Download as pdf or txt
Download as pdf or txt
You are on page 1of 99

University

of Southampton, 2015
FACULTY OF HUMANITIES









AN ARCHAEOLOGICAL REACTION TO THE REMOTE SENSING DATA EXPLOSION.
Reviewing the research on semi-automated pattern recognition and assessing the
potential to integrate artificial intelligence.












By IRIS CAROLINE KRAMER
A dissertation submitted in partial fulfilment
of the degree of Archaeological Computing
(GIS and Survey) by taught course.

I

Acknowledgements
I would like to thank my supervisor David Wheatley for his guidance, interesting suggestions
and critical discussions.

This dissertation was financially supported by the Ordnance Survey and I would like to thank
specifically David Holland for his guidance. I would also like to thank the Geography
department from the University of Southampton for providing me with the TRIMBLE
eCognition License, especially Jadu Dash, Liam Riddy and also Richard Downes from
Issolutions. I’m also very grateful to have been provided with the LiDAR data from the
Environment Agency by Mike, and the archaeological monument information by Lucinda
Walker from Historic England.

Additionally, I’m grateful for the advice from Hans Kamermans, Quentin Bourgeois, Arjan de
Boer, Ralf Hesse, Øivind Due Trier, for great insights into pattern recognition, remote sensing
and archaeological monuments.

Also I would like to thank the great friends I met here for the support and distraction from
too much dissertation work. Special thanks goes to Daniel Joyce for patiently correcting my
writing into English, English. Also my family for the moral support through Skype and
messages.

II


III

Contents
Acknowledgements ........................................................................................................... II
Table of figures ................................................................................................................. VI
Table of tables ................................................................................................................... X
Abstract ........................................................................................................................... XII
Chapter 1: Introduction ..................................................................................................... 1
Chapter 2: Aerial survey in Archaeology ........................................................................... 3
2.1. Remote sensing ...................................................................................................... 3
2.1.1. Aerial photography: Moving up ....................................................................... 3
2.1.2. LiDAR data: Subtle features detected .............................................................. 4
2.2. Image processing .................................................................................................... 5
2.2.1. Aerial photography: More than pretty pictures ............................................... 6
2.2.2. LiDAR: Connecting the dots .............................................................................. 9
2.3. Recognising the patterns of archaeology ............................................................. 12
2.3.1. Aerial photography: extracting features ....................................................... 12
2.3.2. LiDAR: extracting shapes ............................................................................... 13
2.4. Classifying archaeological features ...................................................................... 15
2.4.1. Aerial photography: classifying features ....................................................... 15
2.4.2. LiDAR data classifying .................................................................................... 17
2.5. Reacting to the data explosion ............................................................................. 21
2.5.1. Beyond vision, towards cognition ................................................................. 24
2.5.2. AI, a blast from the past ................................................................................ 25
Chapter 3: Imitating the archaeologist ........................................................................... 27
3.1. Human argumentation: Cognitive computing ...................................................... 27
3.1.1. From single pixels to spatial patterns ............................................................ 28
3.1.2. The integrated approach ............................................................................... 30
3.1.3. Classifying the whole image with easy codes ................................................ 30
3.1.4. What to do with the stranger in our midst? .................................................. 32
3.2. Human experience: machine learning ................................................................. 34
3.2.1. Classified by training ..................................................................................... 34
3.2.2. From supervised classification to supervised learning .................................. 38
3.2.3. Minor training, self-learning algorithms ....................................................... 38
3.3. The struggle of AI in archaeology ......................................................................... 41
Chapter 4: Pilot study on barrow detection .................................................................... 45
4.1. Data sources ......................................................................................................... 46

IV

4.2. Methodology ........................................................................................................ 47
4.2.1. Cognitive argumentation ............................................................................... 48
4.2.2. Towards machine learning ............................................................................ 49
4.3. Preliminary results ............................................................................................... 51
4.3.1. Cognitive argumentation – easy and sufficient ............................................. 51
4.3.2. Towards machine learning – large effort but great potential ....................... 54
4.4. Potential of eCognition for archaeology .............................................................. 56
4.4.1 Comparing eCognition to alternatives ........................................................... 56
Chapter 5: Discussion and future scope .......................................................................... 59
5.1. AI in reaction to the data explosion ..................................................................... 59
5.1.1. Alternating hypes – benefit or burden? ........................................................ 60
5.1.2. Embracing multi-temporal data for heritage management .......................... 60
5.1.3. Harmonising the struggle of AI ...................................................................... 62
5.2. Future scope ......................................................................................................... 62
5.2.1. Consistency in large mapping programmes .................................................. 63
5.2.2. Web-based data repository ........................................................................... 63
5.2.3. The aftermath of the data explosion ............................................................. 66
Chapter 6: Conclusion ..................................................................................................... 67
Bibliography .................................................................................................................... 71
Appendix 1: Snapshots of the workflow for cognitive argumentation
Appendix 2: Snapshots of the workflow for template matching
Appendix 3: Snapshots of the workflow for random forest classifier

V

Table of figures
Figure 1 Interaction of laser pulse with forest canopy resulting multi returns over increasing time
(Challis et al., 2011, 2) ............................................................................................................................ 5

Figure 2 (left) Visual display of 3x3 filter sliding over the digital image to process it (after
blog.teledynedalsa.com accessed 09/09/2015). (right) A few examples of widely applied image filters
(for more examples see Gonzalez and Woods (2008)) ........................................................................... 6

Figure 3 Phenological and spectral characteristics linked to the presence of crop-marks. In the lower
part of figure the QuickBird spectral range are reported (Lasaponara and Masini, 2012a, 26). ........... 9

Figure 4 Illustration of the combined oversampling and REIP calculation available in ARCTIS: (a) true
color aerial photograph; (b) false color composite of REIP indices calculated between bands 50 (676
nm) and 63 (746 nm) using the original data (after Atzberger et al., 2014, 8628). ............................... 9

Figure 5 Different visualizations of grave field Hemlanden at Birka (Sweden). (a): shaded relief model
(light source in NW); (b): positive openness; (c): inverted negative openness; (d): local relief model
(Doneus, 2013, 6433). .......................................................................................................................... 11

Figure 6 A typical palaeochannel feature showing LiDAR topography and intensity data. The profiles
for a transect across the palaeochannel show the intensity values (after Challis et al., 2011, 6). ....... 11

Figure 7 (left) raw satellite image (middle) Hough transform line detections with edges (black) and
candidate points (red) (right) rectangularity heath map (after Zingman et al., 2015, 7) .................... 13

Figure 8 (a) A real image of some archaeological crop marks acquired in Apulian Tavoliere; (b) The
results of the Canny edge detector; (c) Results of the Chan-Vese approach (the green points represent
the manually selected seed points); (d) Results of the isotropic nonlinear filtering ; (e) Results of the
proposed Modified Variance Analysis ; (f) The yellow circles highlight the lines which are detected after
the second iteration with a smaller window size (after D'Orazio et al., 2015, 3582 & 3584). ............. 14

Figure 9 Series of images from Rowlands and Sarris (2007, 799): (a) shaded relief image derived from
the lidar data; (b) (NDVI) which differentiates between vegetation types (light grey) and exposed stone
wall of a basilica near Itanos, eastern Crete, Greece (dark grey); and (c) a classified image of CASI and
lidar data. ............................................................................................................................................. 16

Figure 10 Object-based segmentation and image classification results after de Laet et al. (2007, 839)
at Hisar (southwest Turkey). Classification result of both limestone rock outcrops (i.e. archaeological
features) and shadows of walls and other landscape elements overlaid with the visual interpretation
of the (inferred-) ancient walls. ............................................................................................................ 16

Figure 11 The user interface of the validation step in CultSearcher (Trier et al., 2009b, 23). .............. 17

Figure 12 (left) Template in 2D view from above. The shades indicate the altitude - a lighter shade is
higher. This is a 16m template in a 2 x 2m grid (8x8 gridcells). (right) Template in 3D view. A similar
template as in Fig. 2, but now in projected in 3D. In a 4 X 4 m grid this template has a size of 20 x 20m
(5x5 gridcells) (de Boer, 2005, 249). ..................................................................................................... 18

Figure 13 (left) Results of the detection model; known site boundaries in pink are off from the true site
locations – images are connected with||. (right) Results represent the area in blue under the forest
canopy – difficult field detection (Riley, 2009, 83). .............................................................................. 18

Figure 14 Process of template creation by Trier et al. (2015, 70-71). (top left) Average barrow,
estimated from training data. (top right) Mound template. Bright pixels are positive and dark pixels
are negative. The medium grey pixels outside the black ring edge are exactly zero, thus not part of the

VI

template. (below) Profile of average barrow from training data, with a similarly scaled hemisphere
template as a reference. ...................................................................................................................... 19

Figure 15 (a) The ground plan and cross-section geometry of a charcoal kiln site. (b) The elevation,
slope, topographic position index and hillshade (Schneider et al., 2015, 48). ...................................... 20

Figure 16 Semi-automatic edge tracing. The user identifies the area containing the feature of interest;
the computer then extracts arcs of varying strength, centre, and radius, rejects outliers, and smoothes
the remaining arcs together using weighted moving averages into a coherent shape which is
approximated at weak areas (Redfern, 1997b, 162 20). ...................................................................... 26

Figure 17 The AAS interface (left) Photos tab (right) Monuments tab (Redfern, 1997b, 162 18-19). .. 26

Figure 18 The performance of an automated archaeologist as a three-stage process (Barceló, 2008, 2)


.............................................................................................................................................................. 27

Figure 19 The hierarchy of image objects (TRIMBLE eCognition user guide 2015, 11) ........................ 29

Figure 20 (left) original image, (middle) fusion of intensity and texture gradient images, (right)
segmentation results (Niemeyer et al., 2008). Ordnance Survey Crown Copyright. All rights reserved29

Figure 21 Object-oriented image analysis: the generic procedure (Baatz et al., 2008, 32) .................. 31

Figure 22 Concept of a typical landslide in the soil covered and hilly study area, Belgium (van den
Eeckhaut et al., 2012, 212). .................................................................................................................. 31

Figure 23 Diagram showing some of the objects which comprise a hillfort, as well as attributes to be
used as parameters in feature detection (Sevara and Pregesbauer, 2014, 143). ................................. 32

Figure 24 “This comparative plan of banjo enclosures underlines the morphological similarity between
a range of sites on Cranborne Chase, Wiltshire. All are less than 1 hectare in area and they frequently
occur in pairs, sometimes set within larger compounds. Gussage St Michael 7d is interesting in that
there is a small rectilinear enclosure close to the entrance funnel – a commonly observed trait at a
number of banjo enclosures.” - (English Heritage, 2011, 2) ................................................................. 33

Figure 25 Template matching for palm tree classification from which three templates are derived to
account for the variability in orientation of the leaves. Interface from the eCognition 9.0 (Release Notes,
2014, 3). ............................................................................................................................................... 34

Figure 26 Hillshade maps for (a) possibilities of disturbed kiln sites with diameters of 12, 18 and 24m
and other surface structures on a sloping (left side) and horizontal (right side) surface, and (b) a
synthetic DTM representing kiln sites with diameters of 8–28m without noise and with low, moderate
and high Gaussian noise (Schneider et al., 2015, 49). .......................................................................... 35

Figure 27 (left) A basic, three-layer neural network topology, with a hidden layer (Barceló, 2008, 150)
.............................................................................................................................................................. 36

Figure 28 (right) The hierarchy of MLPs used to classify monuments. Rectangles represent neural
networks, while ovals represent final classification decisions (Redfern et al., 1998, 14) . ................... 36

Figure 29 Overview of the processing steps for the Random Forest algorithm (Stumpf and Kerle, 2011,
2568). ................................................................................................................................................... 37

Figure 30 An illustration of the architecture of the CNN, explicitly showing the delineation of
responsibilities between the two GPUs. One GPU runs the layer-parts at the top of the figure while the
other runs the layer-parts at the bottom. The GPUs communicate only at certain layers (Krizhevsky et
al., 2012, 5). .......................................................................................................................................... 39

VII

Figure 31 Results of test images and labels found most probable by the model (after Krizhevsky et al.,
2012, 8). ............................................................................................................................................... 39

Figure 32 Classification accuracies versus number of training samples (after Hu et al., 2015, 10) ..... 40

Figure 33 Classification of agriculture. From left to right: ground truth, RBF-SVM, and CNN (Hu et al.,
2015, 7). ............................................................................................................................................... 40

Figure 34 (left) Original relief of a Seljuk sultan from the Konya museum. (right) Example of a
decomposed image derived from the relief (Gardin, 1980, 109 & 115)). ............................................. 41

Figure 35 Graphical representation of a Late Bronze Age Warrior Stela, compared to the vectorial
notation for input and output. Many of these formalisations would be used in NN (Barceló 1995).... 42

Figure 36 A neural network to recognize visual textures as use-wear patterns in lithic tools (Barceló,
2008, 241). ........................................................................................................................................... 43

Figure 37 Semantic description of a round barrow. ............................................................................. 44

Figure 38 Various barrow types present in the South of England (after Stonehenge World Heritage Site,
accessed 01/07/2015) .......................................................................................................................... 45

Figure 39 (left) Barlings Abbey barrow cemetery. In this photo the barrows surviving as earthworks in
the field centre-left do not show clearly, but other barrows are strikingly visible as light soil patches in
the centre of the image. (NMR12606/15, 05-SEP-1994. © Crown copyright. NMR (right) Lidar data.
Even with the height exaggerated 20 times many of the barrows are still very unclear. (Lidar courtesy
of Lincolnshire County Council: source, Environment Agency, March 2001. © English Heritage. NMR -
(Crutchley, 2006, 254)). ........................................................................................................................ 47

Figure 40 Avebury case study area depicted with LiDAR-derived slope map and indications of sub areas
which were used to derive parameters, select samples and train classifier from the known barrows. 48

Figure 41 Three barrow types depicted on slope derivative; (left) Bowl (middle) Saucer (right) Bell. .. 49

Figure 42 Cognitive analysis flow diagram. .......................................................................................... 49

Figure 43 Template matching flow diagram. ....................................................................................... 50

Figure 45 Visual depiction of true/false positive detections and false negatives of undetected known
barrow locations. Location: east of Windmill hill and north of Overton hill. ........................................ 52

Figure 46 With red circles the visual interpretation of possible barrows are indicated. The 6 bright
features in the top appear like bowl barrows and show up throughout the forest. The 5 indications
below are of the same size but appear less prominent and might be slightly levelled barrows (reference
image on the topleft from Google Earth). ............................................................................................ 53

Figure 47 In the top left the detection method found a pit structure and number 102 is a barrow which
was indicated by both the OS-first edition and the topographic survey undertaken by the (WANHS) but
missed by HE. ....................................................................................................................................... 53

Figure 48 Snapshot of barrow verification. Image on the right shows the correlation of the pixel to the
templates, in this case white declares higher certainty levels. ............................................................. 54

Figure 49 Snapshots from the application of RF in eCognition showing the known classification below
and the outcome of the machine learning above. Left side shows the general application of all barrows
and right shows the outcome specifically per type of barrow (location: Overton Hill). ....................... 55

Figure 50 Workflow of the change analysis procedure (Niemeyer et al., 2008, 188) ........................... 61

Figure 51 Screenshot of the WebGIS application user interface presented by Duckers (2013). ........... 65

VIII


IX

Table of tables
Table 1 The results from running the automatic confidence assignment on the test set (Trier and Pilo,
2012, 110). ........................................................................................................................................... 19
Table 2 Sum of the deviations between mapped and feature diameters for the kiln templates (Ø =12
m, 18m and 24 m); the table compares the results for single variables and variable combinations. The
numbers in parentheses indicate the number of features that were not detected (Schneider et al., 2015,
55). ....................................................................................................................................................... 20
Table 3 Details of two aerial prospection projects, summarised from Cowley et al. (2013) after Bennett
et al. (2014, 901) .................................................................................................................................. 22
Table 4 Results from the cognitive analysis compared with the known barrows based on the assumption
that the repository from Historic England holds all barrows. ............................................................... 51
Table 5 Evaluation of the rightfully detected barrow types show that the parameters work best for bell
barrows ................................................................................................................................................ 51

XI

Abstract
The increasing amounts and variety within remote sensing data have triggered researchers
to develop semi-automated methods for detecting landscape features. For archaeology it has
been recognised that it is difficult to capture the unique characteristics of archaeological
remains detected in the landscape. This has led to a discussion of the potential from which
was concluded that variable features could not be captured. Nonetheless within geosciences
recent success of pattern recognition allowed for the supervised classification of highly
irregular landslides using cognitive reasoning and machine learning. These concepts form the
basis of artificial intelligence (AI) which has yet to be explored for remote sensing in
archaeology. The success in geoscience is largely provided by the geographical object based
image analysis software, eCognition, which has been explored by archaeologists to some
extent but was not applied in great detail. With the presentation of a pilot study for barrow
recognition around Avebury, United Kingdom, it has been shown that the potential of this
software is far reaching and requires a re-evaluation within the archaeological discipline for
remote sensing. From three basic analysis based on cognitive ruleset, template matching and
machine learning it was concluded that the first is simple to implement but requires a lot of
manual refinement and with the second, simple and consistent features could be detected.
Even though the machine learning algorithm did not yield the best results it is a known fact
that with adequate amounts of training data, this is the best method to detect variable
features as it writes the rules itself based on similarities between training samples. As it was
shown that AI concepts are able to detect unique characteristics of archaeological features
and combine multiple remote sensing layers to do so it was found that the future of semi-
automated pattern recognition in archaeological remote sensing could greatly aid large scale
mapping programmes. Additionally, these methods could be applied for change detection to
monitor archaeological sites for improved heritage management and have the potential to
be widely implemented in an all-encompassing online web application for exchange of
rulesets and training data for generally known features.

XII

XIII

Chapter 1: Introduction
An important task for every heritage management body is the mapping and management of
archaeological features. There are widely applied methods for this, such as desk based
assessment, remote sensing and predictive modelling which have, as of yet, not provided a
satisfactory method for automated large scale recognition of features (Kamermans, 2008).
The rapidly increasing availability and variety of very high resolution (VHR) remote sensing
data has further developed the research of detecting individual archaeological features.
Various techniques from computer vision have been applied to process digital images which
enhance the visibility of archaeological features and even allow the extraction of their
appearance. From these emerging techniques there has advanced a long-term aim to
develop methods for semi-automated pattern recognition within the field of feature
detection and monitoring. Yet the technologies presented thus far have failed to capture the
unique characteristics of archaeological features and have led to the conclusion that
computers will never be able to imitate the abilities of the human interpreter (Parcak, 2009,
110). Although this statement is not shared amongst all archaeologists, for Bennett et al.
(2014) have made efforts to tackle the taboo and prompt archaeologists to further research
the potential. In this scope, the aim of this dissertation is to develop an algorithm to increase
the availability of feature detection, it would be supported by the following research
question;

How has automated pattern recognition been applied for archaeological feature
detection and to what extent is there a future for this technique in specifically multi-scalar
barrow detection on LiDAR imagery?

During the research of previously applied methods in both archaeology and geosciences, the
apparent ongoing discussion in archaeology generated a critical opinion which was
henceforth explored in this dissertation. This critique was inspired by the apparent decision
of the archaeological discipline to settle with simple shape detection and template matching
algorithms whereas geoscientists have been able to classify highly irregular landscape
features (e.g. landslides). A brief study of these methods developed an understanding that
concepts of human cognition and artificial intelligence (AI) formed the foundation of their
research. Therefore it was decided to change the scope of the work to an assessment of the
potential to integrate AI into archaeological feature recognition. The aim has remained the
same, to contribute to the discussion and future scope of the increasing availability of remote
sensing data, but the objectives were altered to further explore the methods applied by
geoscientists. This is done by reviewing previous literature and also the potential of the

1

TRIMBLE eCognition software which stands at the foundation of their methods. This
investigation will be presented in the following structure;
In the second chapter the emergence of remote sensing techniques for archaeology will
be discussed together with the specific applications of aerial photography and airborne LiDAR
(Light Detection and Ranging). The interpretation of the resulting images will be exemplified
for visual interpretation, digital image processing and pattern recognition method.
Eventually this chapter will present an extensive account of all the available literature for
semi-automated archaeological feature detection. This will conclude with an evaluation of
the applied methods and the current state of the remote sensing discussion and possible
further potential.
The third chapter will explain how AI has developed a prominent place in geosciences
through cognitive reasoning and hierarchical interpretation of features. Furthermore they
have applied machine learning in cases where it has been difficult to define rules to describe
features, as will be discussed. Finally, advances towards the general integration of AI into the
archaeological discipline will be discussed with examples mainly presented by Barceló (2008).
In the fourth chapter the potential integration of eCognition will be explored to
specifically detect round barrows around Avebury, England. This software is the foundation
for most of the success in feature recognition of geosciences. The pilot study will present 3
types of analysis specifically based on hierarchical cognition, template matching and
supervised classification. An analysis of the results and potential refinement of the applied
methods will allow the evaluation of the value of this software to be further explored by
archaeologists.
In the fifth chapter the discussion of the potential of the data explosion and semi-
automated methods will be assessed. Specifically by returning to the initial aims of pattern
recognition for heritage management will allow for a discussion of the potential research
areas. While also looking forward to the far-reaching integration of this method at a larger
scale, the scope for future development will include a proposed web-based software
solution.
For an accurate contribution to the discussion it was thought to be important to generate
a full understanding of the discipline. Therefore, throughout this dissertation, not only will
scientific papers be discussed but also articles in the annual newsletter of the AARG (Aerial
Archaeological Research Group) in which discussions have been laid out and software is
presented. Additionally, in some cases, personal contact with authors has been sought. This
has generated insightful information on desires, working methods and why some of the
methods were not picked up by a wider community or continued in follow-up publications.

2

Chapter 2: Aerial survey in Archaeology
The purpose of the review of current literature was to define the applied methods for
detecting archaeological features from remote sensing data to identify potential
improvement in this area. The results of the review are presented here with a focus on LiDAR
and aerial photography, starting with an introduction into the data generated by these two
techniques (2.1), then the techniques of processing the image data are discussed (2.2) and
finally the attempts to automate pattern recognition are outlined (2.3). Finally, the current
status of research and critique on these methods are discussed and through an insight of
other disciplines on this matter possible gaps will be identified to assess the future of semi-
automated pattern recognition in archaeology (2.4).

2.1. Remote sensing


In the 1920s O. G. S. Crawford, the archaeological officer of the British Ordnance Survey,
demonstrated that archaeological structures could be identified from aerial photography.
Unlike the previously applied mapping techniques, features could be identified which could
not be seen, or at least appreciated fully, on the ground (Crawford, 1923). The importance of
this technique was quickly picked up by other archaeologists and even led to the appreciation
of new feature types, such as the first classification of Iron Age ‘banjo enclosures’ in the south
of England by Perry (1969). Nowadays the identification of features from a distance is
commonly referred to as Remote Sensing and also includes ground based methods such as
geophysical survey but here only airborne methods will be discussed. From this range of
applications aerial photography and LiDAR (Light Detection and Ranging) will be discussed in
detail as these techniques are commonly known to enable the mapping of archaeological
remains over large areas with great detail and are available at ever decreasing costs.

2.1.1. Aerial photography: Moving up
Aerial photography, in essence, digitizes the earth's upwelling electromagnetic radiation
(reflected solar radiation) and is therefore a passive remote sensing technique. These
photographs have been generally captured from highly oblique angles to detect features
through the shadows of slight earthworks, the different colours and textures of soil or from
low oblique angles to detect the difference in the growth of crops over buried archaeological
features. While initially only the human visual spectrum could be captured, improvements in
technology has also allowed crop marks to be detected within the Near-InfraRed range which
increases the contrast on the photograph. Besides that the shift from film to digital cameras

3

allowed for a great spectrum range to be obtained with, for instance, airborne hyperspectral
scanning (AHS). The latter technique captures the reflected thermal radiation emitted by the
objects in a multitude of small spectral bands (Verhoeven et al., 2013). This method can
detect subtle vegetation characteristics (e.g. stressed versus healthy plants) and soil
properties (e.g. mineral composition) to a much higher extent than any standard
photographic or multispectral method (Traviglia, 2006). The chance of detecting any
archaeological features from photographs depend largely on the time of the year and day,
and might not show up every year as a field might be used for different crops or soil is left
unploughed and thus reveals only topsoil. Therefore it is important to keep recording the
landscape in order to generate a complete repository of features. However, recently there
has been critique on the ‘traditional’ ways of data capturing at high oblique angles. These
oblique angle photographs are difficult to accurately interpret and are generally captured
after a human has already decided that the land below holds something worth recording, yet
it is unlikely to record features that were not noticed from the air which makes it impossible
to record anomalies that are only visible beyond the human spectrum (Palmer, 2005, 97).
Nonetheless, this remote sensing technique is nowadays still recognised to be one of the
most cost-effective methods of site recording (Doneus et al., 2014, 84).

2.1.2. LiDAR data: Subtle features detected
All site information and imagery can be better employed for studying when placed within its
landscape context. Therefore, archaeologists have strived to produce accurate
documentation of the landscape elevation over the past centuries. This was initially, achieved
by the use of triangulation using sighting or theodolites and later aided by digital terrestrial-
based survey technologies including total stations and RTK GPS also called differential GPS
(dGPS). The first large scale elevation mapping was derived from aerial photogrammetry and
the size of the areas captured was later enlarged with the rise of satellite imagery. Although
these methods allow a context for landscape archaeology they are, as of yet, not able to map
the subtle elevated earthwork monument.
Such large scale detailed mapping was introduced to archaeology with airborne LiDAR
(Light Detection and Ranging) also called ALS (Airborne Laser Scanning), first demonstrated
in a collaboration of the UK Environment Agency and English Heritage (Holden et al., 2002).
LiDAR is an active remote sensing technique where the land surface is scanned by a high-
frequency pulsed InfraRed laser beam from which several aspects of the return rate are
measured. While the transmitter sends out laser beams, the point locations are
georeferenced by the use of RTK GPS aboard the airplane and on the ground. Other data

4

recorded is Time and Intensity, of which the time it takes for every pulse to return to the
transmitter is used to derive the elevation, and the intensity of the return reveals the amount
of InfraRed absorbed by the surface (Hyyppä et al., 2009, 336). The latter has been
successfully used to differentiate in surface cover but has also been important in some
archaeological case studies, as will be explained further on in this chapter. Every laser point
might be returned multiple times, depending of the track towards the surface (fig. 1) from
which the surface and ground can be differentiated.
LiDAR became of increasing importance for archaeological landscapes hidden under
dense tree canopy when it was discovered that a full-waveform recording scanner and high
point density could pierce through the dense forest canopy and record very subtle structures
underneath (Doneus and Briese, 2006).


Figure 1 Interaction of laser pulse with forest canopy resulting multi returns over increasing time (Challis et al.,
2011, 2)

2.2. Image processing
In this section basic data processing, digital image processing and image interpretation will
be discussed. The general application of image processing includes the application of filters
to enhance images, to apply image blur and reduce noise (Gonzalez and Woods, 2008). A
filter refers to accepting (passing) or rejecting certain frequency components. For example,
a filter that passes low frequencies is called a lowpass filter – this blurs/smooths the image

5

(Gonzalez and Woods, 2008, 167). A filtered image is created as the centre of the filter passes
over each pixel in the input window (fig. 2, left). These filters (kernels) can be slid over the
image in a similar way as template matching is applied, this will be discussed later on in this
chapter. From the examples of filtered images displayed in figure 2, right, the edge-detection
filter is mostly used with aerial photography while morphological image processing (High-
Pass/Dilate/Erode) is generally applied to binary images, and is thus of most importance to
LiDAR imagery.


Figure 2 (left) Visual display of 3x3 filter sliding over the digital image to process it (after blog.teledynedalsa.com
accessed 09/09/2015). (right) A few examples of widely applied image filters (for more examples see Gonzalez
and Woods (2008))

Compared to traditional methods digital processing may be expensive at first and require
elaborated and dedicated software and hardware, but in the end it has found to be cost-
effective and has many more advantageous than disadvantages (Lasaponara and Masini,
2012c, 8). For instance it allows for consistent results based on objective evaluations,
facilitates the integration of imagery with other data sources (archaeological records,
documentary sources, etc.) and algorithms can be applied to make archaeological
information extraction and interpretation easier. Examples of digital pre-processing of aerial
photographs and LiDAR data will be further discussed and in the next section this will be
elaborated by introducing automated inspection for feature extraction and classification.

2.2.1. Aerial photography: More than pretty pictures
The traditional manual methods of extracting information from aerial photograph prints,
combining it and plotting it onto base maps, were gradually replaced by computer-based
methods which increased speed, accuracy, flexibility of digital editing and the ability to
reproduce images at any scale (Palmer, 2000). However two complications within this
process have been identified (Lock, 2003, 20). First, aerial photographs are usually taken from
an oblique angle rather than from the vertical, and second, the terrain containing the
archaeological features is rarely flat. The procedure generally referred to as rectification
automatically adjusts for these two problems and rectification software has been written
specifically for archaeology, AERIAL (Haigh, 1998) being the most popular. Aside from aerial

6

photography, another digital recording technology used in the capture of landscape data has
risen in popularity with archaeologists, namely satellite imagery. It is increasingly becoming
available for free, or very low cost and at high resolution and with increasing numbers of
spectral bands but most importantly for archaeology it overcomes the bias of the
photographer. Satellite imagery will not be discussed into great detail here but methods of
data processing are largely the same to aerial photography.

Data processing and image enhancement are mostly carried out using GIS (Geographical
Information System) and remote sensing specific software such as ENVI or ERDAS IMAGING.
This has allowed users to manipulate images to make subtle features more visible to the
human eye and, therefore, easier to interpret than raw photographs (Lasaponara and Masini,
2012a). A digital image processing workflow generally consists of various steps, beginning
with pre-processing which consists of an abstraction of the input data and generally
combines pixels based on specified similarity. All these advances have been welcomed by
archaeologists but have also triggered critique on the apparent lack of understanding of the
physical, chemical, biological and environmental processes that determine whether
archaeological features will be identified in one or any sensor (Beck, 2007). Consequently,
digital image processing for enhancing archaeological features became an informed field of
study set out to detect the nature of the archaeology and its relationship with the immediate
matrix (or context). Lasaponara and Masini (2012a, 19) categorized satellite image
enhancement approaches into three main groups which are similar for general aerial
photography images:

(i) Radiometric enhancement based on the values of individual pixels (e.g. contrast
enhancement, use of colour composites);
(ii) Spectral enhancement based on the transformation of the values of each pixel
on a multi-band basis (e.g. vegetation indices);
(iii) Spatial enhancement techniques based on the values of individual given pixels
and their neighbouring areas

The first enhancement is mostly derived from the hypothesis that archaeological features
produce localised contrasts in the landscape matrix which can be detected using an
appropriate sensor under appropriate conditions. This contrast can be measured directly
from an archaeological feature (e.g. ditch, bank) or indirectly (e.g. crop mark), also called
proxy (Beck, 2007). Contrast enhancement is very helpful for sharpening subtle details, but
it is not a good practice to use the created image for classification and/or change detection
(Lasaponara and Masini, 2012a, 20).

7

The second, spectral enhancement, identifies archaeological features with spectral bands
beyond human vision which capture soil constituents and moisture content. Detections are
thus based on the expected archaeology and features and in figure 3 can be seen how red
channel enhances shadow and crop marks while NIR channel makes the detection of crop
marks easier. The variation emitted by vegetation can be enhanced by mathematical
combinations of different bands, also called vegetation indices (Lasaponara and Masini,
2012a, 27). The most widely used index is the well-known Normalized Difference Vegetation
Index (NDVI) obtained by using the following formula:

(𝑁𝐼𝑅 − 𝑅𝐸𝐷)
𝑁𝐷𝑉𝐼 =
(𝑁𝐼𝑅 + 𝑅𝐸𝐷)

NDVI is indicative of plant photosynthetic activity where high (low) values of the vegetation
index identify pixels covered by substantial proportions of healthy (disease or stressed)
vegetation. Lasaponara and Masini (2007) applied this method to detect crop marks, and
found that the near-infrared band was better for crop marks in dry vegetation, and the NDVI
was better for green vegetation. For hyperspectral images the potential use in archaeology
was recognised but the images were too blurry and hence unstable for accurate mapping
(Traviglia, 2006). This was emphasised by Verhoeven and Doneus (2011) who used airborne
hyperspectral scanning for Red Edge Inflection Point (REIP) derived from spectrally smoothed
and oversampled signatures. Therefore Verhoeven et al. (2013) developed a toolbox for
MATLAB called ARCTIS (Archaeological Toolbox for Imaging Spectroscopy) which includes a
successful noise reduction filter and is now ‘open source’ available (Atzberger et al., 2014).
This toolbox provides Different methods of visualizing the data and the ability to highlight
possible crop or soil marks alongside REIP with options to look at the first principle
components, apply various standard and optimized hyperspectral vegetation indices and
even test some edge detection algorithms (fig. 4).

8


Figure 3 Phenological and spectral characteristics linked to the presence of crop-marks. In the lower part of figure
the QuickBird spectral range are reported (Lasaponara and Masini, 2012a, 26).


Figure 4 Illustration of the combined oversampling and REIP calculation available in ARCTIS: (a) true color aerial
photograph; (b) false color composite of REIP indices calculated between bands 50 (676 nm) and 63 (746 nm)
using the original data (after Atzberger et al., 2014, 8628).

The final enhancement method for geospatial data analysis includes a set of geostatistic
tools useful for the characterization of spatial variation, prediction, simulation and
autocorrelation (Lasaponara and Masini, 2012a, 54). For remote sensing the most widely
used tool is variogram analysis which describes the correlation between image pixels in close
proximity to each other. Only limited research in archaeology has been conducted so far, this
method forms the basis for context analysis of related features which is often said to be the
basic characteristic human interpretation of archaeological features. Besides that, it enables
temporal analysis of time series data which can be used in analysis for change detection.

2.2.2. LiDAR: Connecting the dots
The raw data from the LiDAR data collection is roughly processed by, in general, the capture
company and returned to the customer with XYZ elevation as well as a value for intensity of
the InfraRed return for the first and last return (returns in-between can be provided but are
not of general interest for archaeological applications) (Hyyppä et al., 2009, 336). The XYZ
data is used to generate a digital surface model (DSM) using the first return and the last

9

return points, these are further filtered to make sure only bare ground points remain to
generate a digital terrain model (DTM). These can be produced in a GIS by connecting all the
points for a vector based TIN (Triangulated Irregular Network) or by raster based
interpolation between the points based on their relative proximity, averaged neighbourhood
or specific other equations. The latter process reduces the data intensity and generates an
image with specified cell size which the user bases on the amount of points per pixel. This
digital depiction of the landscape can then be further processed to derive images with
meaningful grouped pixels.
Commonly applied methods in geosciences are Slope derivatives which groups pixels
under the same angle and Aspect derivatives which groups pixels under the same wind
direction (ESRI ArcMap user guide 2015). Besides these morphological derivatives for
detailed LiDAR DTM it is important to apply abstractions which emphasis the subtle
earthworks held within the landscape. Generally the most used method is hill-shading which
is used to illuminate the image from one direction from highly oblique angle (Devereux et al.,
2008). This resembles the way that from highly oblique angles aerial photography reveals
shadow marks, but in this case the interpreter has complete control of illumination
conditions regardless of the time and date of the original survey. Despite this however, there
are still significant limitations to the single light source method because some feature may
be located in the shadow of a slope and others may be aligned parallel to the sun’s
illumination angle which, as a result, not always cast shadows and will hence be missed by
the interpreter. To overcome this various other techniques have been developed by mainly
archaeologists and can be summarised in 4 methods;
(i) Devereux et al. (2008) used Principal Component Analysis (PCA) to visualise the
correlation between 16 images with different illumination directions.
(ii) Hesse (2010) has generated Local Relief Model (LRM) which emphasises small-scale
features by extracting local positive and negative relief variations.
(iii) Kokalj et al. (2011) generated the Sky-view factor (SVF) to reflect the illumination
received by an area and emphasis ridges and depressions regardless of its aspect.
(iv) Doneus (2013) created positive and negative Openness displays which show
differences in micro-relief
Doneus (2013, 6439) compared the Openness method and noted that the LRM is the only
method to emphasise very shallow relief features, such as ploughed barrows, but becomes
problematic in areas with steep relief and large differences in slope (fig. 5). In comparison
with the SVF they noted that Openness enhances both convex and concave structures and
creates a visualization which is stripped of the general topography. Clearly the choice of

10

method applied is generally dependant on the expected topography and features so use of
various methods is the best way. These various methods can be applied in the open source
toolbox, called LivT (Lidar Visualisation toolbox) developed by Hesse (2014).
The intensity of the InfraRed return is hardly used for archaeological applications but can
be used in similar ways as infrared is used in digital images – to analyse the variation of
moisture in the ground and hence detect features which are not found the visible spectrum
(Challis et al., 2011, fig. 6).


Figure 5 Different visualizations of grave field Hemlanden at Birka (Sweden). (a): shaded relief model (light source
in NW); (b): positive openness; (c): inverted negative openness; (d): local relief model (Doneus, 2013, 6433).


Figure 6 A typical palaeochannel feature showing LiDAR topography and intensity data. The profiles for a transect
across the palaeochannel show the intensity values (after Challis et al., 2011, 6).

11

2.3. Recognising the patterns of archaeology
Visual interpretation and image enhancements filters are generally better known than
algorithms for archaeological feature extraction. A number of case studies which have aided
interpreters to get a better perception of archaeological features will be discussed below. In
many cases the final detections are interpreted manually as discussed here but improved use
of these methods could be turned into automated classification of archaeological features as
discussed in the following section.
2.3.1. Aerial photography: extracting features
The first work in archaeology was done by Lemmens et al. (1993) who undertook both noise
reduction by Gaussian smoothing and the Kuwahara filter which also reduces textures to
emphasise edges for their detection. Edges were detected by a circular template and false
positives such as objects with circle-like properties (squares and rectangles) were removed
by including threshold characteristics. A few years later Redfern (1997a) developed a piece
of software called Aerial Archaeology System (AAS) which was based on Windows 95 and
allowed users to apply various digital image processing techniques to digitized aerial
photographs. Besides applying various filters it included a shape extraction algorithm based
on improved edge detection filters for irregular archaeological features, this included the
tracing of line direction. He also improved the method presented by Lemmens et al. (1993)
for circle detection by tracing arc-like shapes to enable detection of earthwork features
which are not perfect circles, but tend to be sub-circular. This method is similar to the later
introduced seed growing algorithms applied in archaeology such as active contours (D'Orazio
et al., 2012, Figorito and Tarantino, 2014), however these are not automated methods (see
fig. 8c).
Other algorithms have been implemented to detect specific shapes such as the
application of Hough (1962) transform, which is a template-based approach that transforms
a binary edge map into multi-dimensional space that represents the parameters of a specific
shape (e.g. straight lines, circles and ellipses). In the most basic Hough transform a line is
defined by the orientation ∅ of the normal and a distance 𝑟 from the origin;

𝑟 = 𝑥 cos ∅ + 𝑦 sin ∅

where 𝑥 , 𝑦 are spatial coordinates of the edge feature. In archaeology, a Hough line
detection was carried out by Zingman et al. (2015) who aimed to detect incomplete
rectangular contours and eventually generated heath maps from the data to point
archaeologists at the locations of possible structures (fig. 7). A modified version, Circular
Hough Transform was applied by Luo et al. (2014) to detect Tops of Qanat Shafts from VHR

12

Imagery in Google Earth. They also applied an edge detection filter (Canny, see also fig. 8b)
and significantly improved the accuracy by using mathematical morphological processing
beforehand to reduce noise effects and variation. Eventually they compared Hough
transform with the Hough circles which improved both accuracy and computation time. The
circle pattern is described as;

𝑟 5 = (𝑥 − 𝑥6 )5 + (𝑦 − 𝑦6 )5

where x6 and y6 are the coordinates of the centre and r is the radius of the circle.
Another kind of approach to extract variable shapes more native to archaeology was
proposed by D'Orazio et al. (2015) using a Modified Variance Analysis. They introduce a new
model for curve fitting to evaluate pixel displacements and estimate the pixel variance within
each region. Knowledge of the expected appearance of features, mainly blurry edges, has
been used to modify the variance analysis and provide a more selective wide line detection
approach which was very successful (fig. 8e). Finally a smaller window size was applied to
detected finer lines, without introducing false positives (fig. 8f). Such digitisation methods
are valuable for generating an unbiased repository and moreover forms an essential basis for
classification of features.


Figure 7 (left) raw satellite image (middle) Hough transform line detections with edges (black) and candidate
points (red) (right) rectangularity heath map (after Zingman et al., 2015, 7)


2.3.2. LiDAR: extracting shapes
For LiDAR derived DTMs there have been no general pattern recognition methods developed.
Yet Openness has been a valuable technique for visualisation but has been determined to be
of importance for (semi-)automatic classification because of the generated sharp contrast
and visible edges in areas of transition (Doneus, 2013, 6439). Further LRM could be useful for
detection of shallow features such as heavily ploughed barrows.

13


Figure 8 (a) A real image of some archaeological crop marks acquired in Apulian Tavoliere; (b) The results of the
Canny edge detector; (c) Results of the Chan-Vese approach (the green points represent the manually selected
seed points); (d) Results of the isotropic nonlinear filtering ; (e) Results of the proposed Modified Variance Analysis
; (f) The yellow circles highlight the lines which are detected after the second iteration with a smaller window size
(after D'Orazio et al., 2015, 3582 & 3584).

14

2.4. Classifying archaeological features
Pattern recognition and, in turn, image classification techniques have the overall objective to
automatically or semi-automatically categorize all pixels of given scenes (Lasaponara and
Masini, 2012b). Classification can be performed directly on pixels based on specified
similarities in pixel values or pre-defined context. Otherwise, an object-based approach finely
groups pixels in an initial stage to then further group these based on a variety of thresholds
such as their spectral and /or shape characteristics, using single or multiple image channels,
until the classification reached is satisfactory. Generally in archaeology, the first method is
applied as will be discussed in this section.
There are a number of other ways to classify images, here the unsupervised, supervised
and decision rules will be discussed. For unsupervised classification the pixels are grouped in
a pre-defined number of classes, which are assigned according to their characteristics, mostly
spectral. On the other hand, supervised classification allows for more control and generally
involves various threshold and iterations. Finally the classification based on decision rules has
only been applied to a limited extent in archaeology. This method is defined by the object-
based approach, and classifies the landscape and feature characteristics in a semantic order
to finally detect the composed feature of interest based on combinations of these.

2.4.1. Aerial photography: classifying features
Integrated into the AAS software created by Redfern (1997a) was a classification algorithm
based on Artificial Neural Networks (ANN) which were meant to generate new classifications
based on training from a repository holding user digitised features. The value of ANNs was
recognised as they are not programmed with rules in the way that most algorithms work,
which is very useful for working with problems where rules are difficult to determine, and
were commonly applied to classification problems (Redfern, 1997a, 37). Yet this method was
not generally taken up in archaeology.

An unsupervised, object based, classification method was implemented by Rowlands and


Sarris (2007) using the Reede Xiaoli (RXD) anomaly detection algorithm within ENVI software.
This method extracts unknown targets that are spectrally distinct from the image
background. The analysis showed an integrated approach using LiDAR DTM and CASI imagery
derived NDVI for which results on the remains of a basilica can be seen in figure 9. Although
the result is presented for a snapshot, at a large scale such an unsupervised method requires
great amounts of interpretation of the classified classes.

15


Figure 9 Series of images from Rowlands and Sarris (2007, 799): (a) shaded relief image derived from the lidar
data; (b) (NDVI) which differentiates between vegetation types (light grey) and exposed stone wall of a basilica
near Itanos, eastern Crete, Greece (dark grey); and (c) a classified image of CASI and lidar data.

Three different ways of feature classification from satellite imagery were presented by
de Laet et al. (2007) to evaluate GIS-, pixel- and object-based techniques. This was carried
out by respectively edge enhancement-, maximum likelihood- and nearest neighbourhood-
classification to detect walls and their shadows. The conclusions from their analysis stated
that object-based analysis worked better than pixel-based but none of the techniques
succeeded in capturing the unique spectral class of features (fig. 10). Even though the
techniques seemed unsuccessful this might have been caused by their applied methods. First
of all, it is questionable if shadows of walls would reveal a unique spectrum compared to
those caused by trees. Secondly, two of these methods were preceded by a ‘training phase’,
yet it is unclear if a significant amount of pixels was used to derive the unique spectrum and
to allow an accurate classification. Finally, it seems that for each application the most basic
analysis was carried out which seems not to represent the full potential of these techniques.


Figure 10 Object-based segmentation and image classification results after de Laet et al. (2007, 839) at Hisar
(southwest Turkey). Classification result of both limestone rock outcrops (i.e. archaeological features) and
shadows of walls and other landscape elements overlaid with the visual interpretation of the (inferred-) ancient
walls.

Another pixel-based approach was presented by Trier et al. (2009a) who created
software called CultSearcher to detect ‘rings’ in the landscape, which represent crop marks

16

(of the indirect remains) of burial mounds in southeast Norway. The first step in the software
is to apply filters to remove unrelated information such as plough furrows. Then local
contrast was enhanced to execute template matching in which an ‘ideal’ image slides across
the image to detect equals. Finally the detections are extracted to further assess indicators
revealing the quality of a match (fig. 11). Experiments with the method demonstrated that
the proposed algorithm was able to detect many circular patterns. But, many are also missed
by the algorithm, and many false detections are made. Finally, they proposed future
improvement including multiple spectral bands and algorithms such as NDVI.


Figure 11 The user interface of the validation step in CultSearcher (Trier et al., 2009b, 23).

The shape detection algorithms discussed in the previous section, could be considered as
a classification method as it seeks to find specifically defined features, although this will
detect all circle like features in the area, it does not consider circles of other nature and is
thus technically merely pattern recognition.

2.4.2. LiDAR data classifying
The first case study of automated classification on LiDAR data was undertaken by de Boer
(2005) for barrow detection in the Netherlands using template matching (fig. 12). The case
study was centred on forested areas as the best preserved barrows are located here because
these are generally uncultivated. To account for varying diameters, the size of the template
was generated for different sizes (diameters: 12, 16, 20, 24 and 28 m). This template was

17

then slid over the image and the Pearson product-moment correlation was measured for
each location, and locations of high correlation were verified afterwards. Although most
registered barrows had high correlation and new detections were made, in the end it was
found not to be particularly successful because of the large amounts of false positives due to
none unique shape and relatively low resolution of LiDAR data, but also because of limited
computational power (de Boer, 2015, personal communication).


Figure 12 (left) Template in 2D view from above. The shades indicate the altitude - a lighter shade is higher. This
is a 16m template in a 2 x 2m grid (8x8 gridcells). (right) Template in 3D view. A similar template as in Fig. 2, but
now in projected in 3D. In a 4 X 4 m grid this template has a size of 20 x 20m (5x5 gridcells) (de Boer, 2005, 249).

A very different approach was presented by Riley (2009) in an MSc dissertation project.
She generated a GIS based approach in the Model Builder function of ArcMap based on the
classification of pixels resembling the morphology of conical barrows in IOWA, USA, based
on context information. Different than most approaches this methods took into account
spatial enhancement, as was discussed in section 2.2.2, to include factors of slope and aspect
and flow direction. For example, a circular kernel with a diameter of average barrow size was
slid over the image to detect areas which have the greatest variety of aspect – a characteristic
of a conical barrow. Although there was no accuracy assessment untaken the snapshots
visible in figure 13 show yellow pixels which look like correctly detected barrows.


Figure 13 (left) Results of the detection model; known site boundaries in pink are off from the true site locations –
images are connected with||. (right) Results represent the area in blue under the forest canopy – difficult field
detection (Riley, 2009, 83).

18

Trier and Pilo (2012) presented an automated detection of pit structures resembling iron
extraction and hunting systems using template matching. Their analysis was divided into a
training phase and a detection phase. The first was undertaken to derive characteristics of
the ideal shape of pits. The resulting template was slid over the image to measure the
correlation of each characteristic and to assign confidence classes to each detection. The
verified results from their test site were satisfactory and were 10 times quicker than manual
reading of the LiDAR data and some of their detections were not picked up visually.
Other work carried out by the same authority (the Norwegian Directorate for Cultural
Heritage) developed a method for barrow detection (Trier et al., 2015, fig. 14). Their
workflow was similar to the previous method using template matching based on a training
phase to identify thresholds for assigning confidence values to the detections (table 2).
Overall, 53% of the true grave mounds were detected with correct geometry, 17% of the true
grave mounds were detected with incorrect geometry, and 30% were missed.


Figure 14 Process of template creation by Trier et al. (2015, 70-71). (top left) Average barrow, estimated from
training data. (top right) Mound template. Bright pixels are positive and dark pixels are negative. The medium
grey pixels outside the black ring edge are exactly zero, thus not part of the template. (below) Profile of average
barrow from training data, with a similarly scaled hemisphere template as a reference.


Table 1 The results from running the automatic confidence assignment on the test set (Trier and Pilo, 2012, 110).

19

The template matching approach presented by Schneider et al. (2015) was rather
different than the other presented case studies. Their aim was to locate charcoal kiln sites in
north of Cottbus, Germany, by generating multiple templates which include different sizes
and imperfections based on training data from the area. They also applied various
morphometric variables on each of these templates to differentiate in relevant spatial
characteristics (fig. 15). To assess their results they evaluated the relevance of each of the
morphometric derivatives and noted that the template matching results differed depending
on the morphometric variable used (table 2). Hill shading provided least accurate results,
because within the case study area slight slopes are evident which face different directions.
Perhaps another one of the LiDAR visualisations discussed in section 2.2.2. would have been
more successful.


Table 2 Sum of the deviations between mapped and feature diameters for the kiln templates (Ø =12 m, 18m and
24 m); the table compares the results for single variables and variable combinations. The numbers in
parentheses indicate the number of features that were not detected (Schneider et al., 2015, 55).


Figure 15 (a) The ground plan and cross-section geometry of a charcoal kiln site. (b) The elevation, slope,
topographic position index and hillshade (Schneider et al., 2015, 48).

20

2.5. Reacting to the data explosion
From the first experimentation with semi-automated pattern recognition in the 1990s,
applications have not generally been picked up with notions that complete manual
interpretation provided better results leading some researchers such as Parcak (2009, 110)
to conclude that there will never be any automated mapping for archaeology and Hanson
(2010) doubted whether future work could provide effective methods due to the numerous
false positive detections. Yet this was not generally agreed upon in a discussion in the
AARGnews of September 2009 which reviewed the work presented at CAA of that year
positively (de Laet and Lambers, 2009). They summarised that archaeological features have
“…near infinite diversity of their shapes, proportions, locations, and contexts is a huge
challenge for semi-automated detection approaches...” and “…may be regarded as
‘uncooperative’...”, yet “In archaeology we are only beginning to understand and describe
typical features that we encounter in given regions in such terms.”. To conclude they mention
that it will mainly be useful to map typical, recurrent features whereas the detection of
atypical, rare, unusual features will always remain the domain of visual image interpretation.
The editor of the newsletter reviewed their input as a major step in the development of
image interpretation and stressed the reader to have faith in even the unsuccessful case
studies but also that researchers should be more critical of their input (Palmer, 2009).
Most of the case studies presented in this literature review have been presented after
this discussion took place. Yet the taboo on semi-automated methods remains in some form
because only limited amounts of further research is undertaken. Despite the work of Trier et
al. (2009-2015) for the Norwegian Directorate for Cultural Heritage, large scale pattern
recognition and classification methods have rarely been adopted in local or national mapping
programmes. An evaluation of large scale mapping programmes was presented by Cowley et
al. (2013), who stressed the importance of consistent mapping techniques which can be
provided by methods of computer vision. As an example they compared the English Heritage
National Mapping Programme (NMP - Horne, 2009) mainly based on the visual work of aerial
photographers and interpreters, and the Baden-Württemberg mapping programme based
on visual interpretation of slight earthworks and ditches which were emphasised on LRM
derived from LiDAR images (Hesse, 2013). In table 2 can be seen that the NMP required a
great investment in time and money on detailed mapping while the LiDAR project took
significantly less time and effort. Despite the fact that not all sites have been verified, this
project has developed some new insights on the nature of archaeology in the landscape. With
this comparison they stress for mapping programmes to be critical when choosing either

21

method as in some cases insights of archaeological patterns in large scale areas can be more
valuable than detailed and accurate mapping of small areas.

English Heritage National Mapping Baden-Württemberg LiDAR


Programme (NMP) Mapping, Germany
Aerial resource used Predominantly aerial photographs, LiDAR project
latterly some LiDAR
Area covered 52 000 km2 35 000 km2
Personnel Multiple operators for each project area Single operator
Method Manual classification and interpretation Semi-automated classification of
based on topographic, soil and digital terrain models based on
vegetation contrast micro topographic variation (LRM),
with manual verification
Timescale 20 years (since 1992/93) 6 years
Number of sites Over 100 000 sites 600 000 sites (estimated)
Table 3 Details of two aerial prospection projects, summarised from Cowley et al. (2013) after Bennett et al.
(2014, 901)

It is commonly agreed that archaeology is at higher threat than ever from for instance
climate change, mechanical ploughing and rapid urbanisation (Parcak, 2009, 211). This
requires action at a greater scale which is enabled by the ‘data explosion’ – with increasing
availability, extent and lower prices of VHR remote sensing. Although this wealth of data
could only be efficiently explored by the application of semi-automated methods. This
statement was developed in an evaluation of semi-automated methods for feature detection
by Bennett et al. (2014). They noted that the rhetoric against automation has softened in
recent years due to the emergence of new visualisation techniques for LiDAR data and the
application of some automated techniques, but also that research in this area remains limited
and hence their aim was to break this taboo. Reflecting on the known case studies, they
conclude that the field should focus on predictable shapes and sizes as these work best within
the presented template matching and shape detection algorithms. Derived from the work of
de Laet et al. (2007), who evaluated GIS-, pixel- and object-based techniques (see 2.4.1), they
note that the over-simplicity and inflexibility of the modes of automated or semi-automated
detection selected for analysis mean they are poorly suited to the diversity in shape, size,
spectral and topographic properties of the features we categorise as archaeological. In the
end, the reasoned archaeological interpretation supported by observation and knowledge is
a valuable specialist skill that computer detection methods, however sophisticated, will
ultimately fail to replace. Therefore the most urgent improvements for future approaches do

22

not encourage researchers to re-evaluate applied techniques but point towards the following
four directions;
i. A credible benchmark, based on Big Research
ii. Improved protocols for recording pre-processing, data categorisation and metadata
iii. Incorporation of multi-sensor data
iv. Innovative manual interpretation methods

The first direction stresses the importance of universally agreed systems for
interpretations of features which could then provide training and validation datasets
essential for the application of many computer vision techniques to image analysis. The
second is intended to define a recording system for the applied pre-processing techniques to
improve the intelligibility of, and confidence in, the results of semi-automated techniques.
This would be accompanied by a categorisation system of different levels of processing and
interpretation which is to be included as metadata in all feature assessments. Besides that,
feature recording protocols should be expanded to include complementary environmental
data such as land use, land management strategies, dates of feature visibility and proportion
of known feature visible to assist definitions and help explain gaps in identification. The third,
concerning the incorporation of multi-sensor data, has yet to be explored in any depth, but
incorporation of tools such as decision tree analysis are believed to form a good start to allow
strategic targeting of attributes within multiple datasets that represent archaeological
features. Finally, their last direction accounts for the mistaken assumption that manual
methods bring about the best results and propose to further develop recent experimentation
of ‘crowd-sourced’ interpretations from LiDAR derivatives as described by Duckers (2013).
Reflecting on their evaluation of working methods and future approaches it seems as
though there is only little room for improvements in the area of semi-automated pattern
recognition. Their conclusions imply that it will be difficult if not impossible to develop a cost-
effective method. As discussed earlier in this literature review, shape matching algorithms
are renowned for their numerous false positives and template matching algorithms are
intrinsically restricted to local environment as the testing phases were intended to alter the
algorithm accordingly. Their main source to dismiss future incorporation of diversity within
archaeological features is de Laet et al. (2007) but this method actually only grasped the most
basic analysis of each of their evaluated techniques. Also, it seems rather contradictory that
they first praise geographical sciences for developing semi-automated methods for capturing
complex structures such as landslides, which have endless variation caused by nature, while
dismissing such developments for archaeology. Perhaps we should actually be looking at

23

their applied methods to revise the possibilities of capturing complex archaeological
structures.

2.5.1. Beyond vision, towards cognition
Contradicting to their settlement with simple features Bennett et al. (2014, 902)
acknowledge that within the computer vision process, the role of the expert should be
perceived as an integral part of the feature engineering process. Hence they encourage that
“to do this there has to be willingness to engage with both the practical and theoretical
challenges of explaining and replicating human visual interpretation processes”. This
acknowledgement requires a change in the understanding of how features appear in the
field, on images and in our human mind. In order to reach such a state, different than the
superficial vision which has been grasped to capture, we must dive into the cognitive process
which makes human interpretation successful. The difficulty of detecting features in imagery
is due to the fact that they are often only partially visible due to their bad state of
preservation as well as to spectral similarities between archaeological remains and
surrounding entities (de Laet et al., 2007, 834). Yet the human brain is able to visually
complete the features, differentiate between them and fill in the gaps of familiar shapes and
features. This can be further explained by the following seven image characteristics which
make up archaeological perception of features according to Lasaponara and Masini (2012c,
7);
(i) Tone
(ii) Texture
(iii) Shape
(iv) Size
(v) Pattern
(vi) Shadow
(vii) Association

Yet this process works different in every brain which is caused by the bias of previous
experience which leads to different and sometimes even wrong interpretations (Bennett et
al., 2014, 899). Thus the cognitive interpretation of an experienced interpreter has to be
replicated. The contradiction here, however, is that the applied methods are all set out to
capture one of the image characteristics described by Lasaponara and Masini (2012c, 7)
whereas our brains looks at the complete picture. In the end, no matter the algorithm or
template you apply in any computer program, what you’ll get in return is the image reflection

24

of what you asked for. So, if Bennett et al. (2014, 902) request to replicate the expert then
there has to be a fundamental change in the methods or the way they have been applied.

2.5.2. AI, a blast from the past
Rather than the trend of focussing on developing frameworks for metadata we must look
beyond it towards the methods in which they will be usefully integrated and these must
include;

i. Ability to replicate the human cognitive argumentation


ii. Ability to integrate data from multiple remote sensing sensors
iii. Ability to integrate complementary environmental data
iv. Ability to detect variability within features
v. Ability to detect variability within the appearance of features
vi. Ability to generate data categorisation and record metadata
vii. Ability to use the data categorisation for learning algorithms
viii. Ability to apply the method over large areas

Throughout the rest of this dissertation these points will be further addressed but it is
important to notice that parts of such a method have actually been proposed by Redfern
(1997a), discussed earlier in this chapter (2.3.1.). With the Aerial Archaeology System being
based on artificial intelligence he developed an integrated approach, using both aerial
photography and a DTM generated from the aerial photography (fig. 16). He combined
spectral and spatial analysis and generated an automated morphological-topographical
classification scheme for earthwork monuments with the aim to build up a large database of
known structures. That database would form the training data for the Artificial Neural
Networks which would enable the recognition of future detections with higher accuracy.
It presented an approach to the automatic detection and extraction of low contrast,
damaged objects, in scenes where significant noise and clutter are present. The drawback of
this method was the manual selection of an approximate centre point and size of the feature
of interest (Redfern et al., 1998, fig. 17). To overcome this they proposed future
improvement in automation of the method with Hough transform, yet a follow up publication
is not known. Although it seems to be nowhere mentioned why the AAS was not picked up
by the archaeological community, it might have been subject to the same critique on artificial
intelligence of the time – computational power was not sufficient. This was also the reason
that the barrow detection method developed by de Boer (2005) was not further refined (de
Boer, 2015, personal communication). Another reason might be that during the time of

25

publication most images were still captured under high oblique angles which does not
support the rectified digitization as it was criticised by Palmer (2005, 97). Within the software
this would have been an intrinsic problem for the training phase based on actual sizes and
shapes of the features.
In recent years AI has developed more successful algorithms empowered by an increase
in computational power and computer vision has further developed ANNs. The
implementation of AI and hence modelling human arguments and experience will be further
discussed in the next chapter.


Figure 16 Semi-automatic edge tracing. The user identifies the area containing the feature of interest; the
computer then extracts arcs of varying strength, centre, and radius, rejects outliers, and smoothens the remaining
arcs together using weighted moving averages into a coherent shape which is approximated at weak areas
(Redfern, 1997b, 162 20).


Figure 17 The AAS interface (left) Photos tab (right) Monuments tab (Redfern, 1997b, 162 18-19).

26

Chapter 3: Imitating the archaeologist
Different from many disciplines that apply artificial intelligence to their data organisation
problems, archaeology is based on social studies in which patterns and rules are difficult to
define. Although archaeologists have attempted to formalise arguments and created
computer systems to critically reflect arguments, the creation of successful AI systems are
very rare. The main problems in rule definition within the archaeological record is that it is
almost always incomplete: not all past material things have remained until today or are found
destructed. So, in order to replicate an archaeologist it must be able to reconstruct
incomplete data for which Barceló (2008, 49) identified the following key concepts;

i. Deduction (classification derived from argumentation)


ii. Induction (classification based on learning from training data)
iii. Analogy (information recalled from previous case studies)

This perception is key to simulate the missing parts of the input and further extract features,
recognise their meaning and respond accordingly (fig. 18). Different from archaeology,
geographical sciences have been able to accurately define these concepts and generate
models to replicate human interpretations and use AI where rule definition is difficult. In this
chapter first the current status of AI in archaeology will be discussed and then mainly
geographical examples will be presented to define argued concepts and also the way they
have attempted to replicate experience with machine learning.


Figure 18 The performance of an automated archaeologist as a three-stage process (Barceló, 2008, 2)

3.1. Human argumentation: Cognitive computing


Current ways of replicating what archaeologists interpret is done by means of shape
detection preceded by image filtering to generate a level of smoothing for easy detection by
the algorithm. This filtering also removes other aspects of human perception such as texture,
tone and shadow. In order to replicate the human cognitive argumentation we have to look
at the complete visual process which starts from the very bottom of the hierarchy, with the
smallest entity of the image – pixels. The reason we may want to begin at the pixels is a
perception of patterns we humans don’t explain because the brain does these steps for us.

27

To illustrate this, textures are displayed on images using a specific alternation of pixels,
shadows on the other hand, are recognised as a darker depiction of the true colours. Yet such
a method is new to archaeology; it has been presented in object based analysis, which groups
together pixels of similar values based on defined thresholds, as explained earlier in this
literature review (2.4). Part of the concept has been grasped by de Laet et al. (2007) who
aimed to detect shadows in an image to characterise the presence of walls. Unfortunately,
this was the only parameter used but various characteristics of human perception can be
combined as will be further discussed in this section.

3.1.1. From single pixels to spatial patterns
Remote sensing has generally been manual interpreted or aided by GIS-based image
processing by analysing structural parameters of the image (i.e., colour, tone, texture,
pattern, shape, shadow, context, etc.). Geographic Object Based Image Analysis (GEOBIA)
emerged at the beginning of the 21st century from a general feeling that certain types of
remote sensing image analysis were not possible in general Geographic Information Science.
Blaschke and Strobl (2001) rhetorically asked themselves; ‘‘why are we so focused on the
statistical analysis of single pixels, rather than on the spatial patterns they create?’’. These
patterns could be captured in objects as defined by Baatz et al. (2008, 77);

“Geographic Object-Based Image Analysis (GEOBIA) is a subdiscipline of Geographic


Information Science (GIScience) devoted to developing automated methods to partition
remote sensing imagery into meaningful image-objects, and assessing their characteristics
through spatial, spectral and temporal scales, so as to generate new geographic information
in GIS-ready format.”

This definition is further clarified by the critical bridge between the (often disparate) raster
domain of remote sensing, and the (predominantly) vector domain of GIS. This ‘bridge’ links
both sides of these domains by the generation of polygons representing geographic objects.
These objects are derived from the hierarchical grouping of smaller segmented image
elements based on decision rules (fig. 19).
Since its emergence this field of study has developed a strong community in the field of
remote sensing. This resulted in the development of software called eCognition which
enables this kind of analysis (Baatz and Schäpe, 2000), publication of a handbook (Blaschke
et al., 2008) and every other year there is a conference to streamline new ideas which may
further develop the software and fields of applications (Blaschke et al., 2014, 185).

28


Figure 19 The hierarchy of image objects (TRIMBLE eCognition user guide 2015, 11)

It is generally agreed that GEOBIA builds on older segmentation, edge-detection, feature


extraction and classification concepts that have been used in remote sensing image analysis
for decades, archaeological examples of which were introduced in the literature review. Yet
its emergence has provided an improved link between the spatial concepts applied in multi-
scale landscape analysis in GIS and the synergy between image-objects and their radiometric
characteristics (Blaschke, 2010). This can be further exemplified with a basic land-cover
classification, presented in the handbook, which tackles the problem of scale by using the
texture captured on the image (Niemeyer et al., 2008). The data used in their research
consisted of scanned aerial photography with a 0.25 m ground sample distance of
Southampton, England, obtained from the Ordnance Survey of Great Britain (figure 20, left).
Most boundaries in remotely sensed images consist of both an intensity and texture
boundary. In this case the gradient of each of these images were fused into one image (figure
20, middle) which enabled the segmentation of meaningful objects (figure 20, right). The
achieved results were very accurate and outperformed the generally applied Canny gradient
operator.


Figure 20 (left) original image, (middle) fusion of intensity and texture gradient images, (right) segmentation
results (Niemeyer et al., 2008). Ordnance Survey Crown Copyright. All rights reserved

29

3.1.2. The integrated approach
The eCognition software also enables the analysis of data from multiple remote sensing
sensors in an integrated manner. This can be done by hierarchical analysis; each step of the
analysis defines the grouping of sets of pixels which is based on thresholds which can be
defined on any of the loaded image layers. In this way the user is not just able to use the raw
data but as many derivatives as one may find useful at any point in the analysis. Whereas the
land-cover classification by Niemeyer et al. (2008) was still only based on one image layer, in
later case studies the use of multiple layers becomes the standard. In fact, Verhagen and
Drăguţ (2012) applied an object-based landform delineation and classification from LiDAR
derived DEMs for archaeological predictive mapping. The archaeological heritage
management system in the Netherlands relies on predictions of site locations which are
based on geomorphological maps. The predictive maps are based on 1:50,000 scale maps
which are insufficient for the variable archaeological record and thus their aim was to explore
the possibility to generate geomorphological classifications at a smaller scale. They used
elevation, slope and curvature as input layers which were used at different stages of the
classification. The results show that the GEOBIA classification conforms in broad terms to the
visual interpretation and they hence urge for further development of the technique for
complete integration in predictive modelling.

3.1.3. Classifying the whole image with easy codes
When modelling reality, complex models are the standard which is now possible through the
increasing availability of VHR imagery and LiDAR data. Although GEOBIA methods have
traditionally begun with the identification of objects such as buildings or trees, there have
been recent methodological developments, particularly when addressing complex geospatial
features such as landslides (Blaschke et al., 2014, 189). This level of complexity requires
domain knowledge and semantics, in order to cope with the specific characteristics of the
structure. This advanced model could not be modelled in the two stage workflow where a
segmentation step is followed by a classification step. In this model the objects themselves
should be constantly altered in order to move from object primitives in an early stage
towards objects of interest in a final stage of the analysis (Baatz et al., 2008). As objects form
both the information provider and the building blocks the suggested name for this advanced
workflow is ‘object-oriented image analysis’. This approach can be realised by using Cognition
Network Technology (CNT), in eCognition, which allows for typical programming tasks like

30

branching, looping and variable definition. In figure 21 can be seen how the whole process is
iterative in order to define objects of interest.


Figure 21 Object-oriented image analysis: the generic procedure (Baatz et al., 2008, 32)

In many cases, a previously defined ontology of the image-objects can be used to model
real-world objects. In order to describe natural variability, many models need to be capable
of expressing their vagueness (e.g. by fuzzy rules) and to be adaptable according to
unforeseeable imaging situations (Blaschke et al., 2014, 187). An example of such a
geomorphic fingerprint was presented for densely vegetated landslides by van den Eeckhaut
et al. (2012). They used multiple LiDAR derivatives (such as slope, roughness, SVF, openness
and flow direction) as the vegetation obscured the spectral characteristics on the imagery.
They identified several features with significantly different morphologic signatures and
possible distortions (fig. 22). These were described as rules and used successfully in a
supervised learning approach as will be discussed later on in this chapter (3.3.2).


Figure 22 Concept of a typical landslide in the soil covered and hilly study area, Belgium (van den Eeckhaut et al.,
2012, 212).

31

The eCognition software supporting GEOBIA created a modular programming language CNL
(Cognition Network Language – TRIMBLE). Although it is similar, it should not be confused
with the Object-oriented programming language. Nevertheless the GEOBIA community has
recognised (Hay and Castilla, 2008, 84, Blaschke et al., 2014, 187) the self-similarity across
analysis and that by the integration it could make the software more powerful with semantic
inheritance concepts (is part of, is more specific than, is instance of). This aim has also been
recognised for GIS in which the geo-relational data model, separates spatial and attribute
data and links them by using a common identifier. The potential of object-oriented GIS was
recognised for archaeological applications (Wheatley and Gillings, 2002, 238). As an example
of an OO-GIS they proposed the class of objects termed ‘bronze age burial mound’ which
would have a specific set of attributes: area; height; volume; date; number of burials. Then
they would also specify a series of sub-classes: bell-barrow; round-barrow; dish-barrow;
pond-barrow etc. that inherit all of the attributes of the parent as well as adding their own
unique attribute – shape. This descriptive model could be reproduced in GEOBIA, except for
the attribute data of number of burials. Another archaeological example of a decision tree to
classify an earthwork hillfort, intended to be replicated in GEOBIA, was presented by Sevara
and Pregesbauer (2014) at the GEOBIA conference of 2014 (fig 23).


Figure 23 Diagram showing some of the objects which comprise a hillfort, as well as attributes to be used as
parameters in feature detection (Sevara and Pregesbauer, 2014, 143).


3.1.4. What to do with the stranger in our midst?
This section has focussed on the possibility to formalise features which have been classified
by archaeologists. Although certainly, there are features which can be detected from the air
but are yet to be classified or unique in their shape and form. As it was mentioned in section
2.1, the introduction of remote sensing led to the classification of ‘banjo enclosures’ by Perry
(1969). These features have been defined to consist of small sub circular enclosure with a
narrow approach way consisting of parallel ditches (English Heritage, 2013). To enable such

32

discoveries we may want to generate a hierarchy to differentiate the anomaly from the
natural matrix. In practise it would classify agricultural field at a first level, differentiate
positive from negative crop marks (stone- or ditch structure) in the next, and further apply
fuzzy rules to define the strength of the anomalies and geometrical characteristics (e.g.
rectangularity/parallel lines or circularity and level of enclosing) which could direct
interpreters to certain location of interest (in part similar to the heath map generated with
Hough line detection by Zingman et al. (2015) discussed in section 2.3.1).


Figure 24 “This comparative plan of banjo enclosures underlines the morphological similarity between a range of
sites on Cranborne Chase, Wiltshire. All are less than 1 hectare in area and they frequently occur in pairs,
sometimes set within larger compounds. Gussage St Michael 7d is interesting in that there is a small rectilinear
enclosure close to the entrance funnel – a commonly observed trait at a number of banjo enclosures.” - (English
Heritage, 2011, 2)

33

3.2. Human experience: machine learning
Whereas the expert argumentation has been integrated to some level, experience has been
criticized and avoided for archaeological feature detection. Within machine learning
inductive learning tools are trained to recognize patterns for outcome predictions from the
generalisation of a set of data of which the desired outcome is known (training data) to a
larger set of circumstances (Barceló, 2008, 75). The most recent attempts in archaeology
have been found in template matching but as it will be discussed here, there are some
significant restrictions which would eventually hamper the ability to capture the variability
intrinsic to archaeology. Barceló (2008, 261-264) described how computational intelligence
used in geography could aid the interpretation of remote sensing data but as of yet this has
not been directly implemented. Therefore other supervised classification methods are
proposed which have been successful in remote sensing classification of morphological
distinct features. This section introduces these and eventually discusses improvements of
methods which may enable large scale classification of variable archaeological features.


Figure 25 Template matching for palm tree classification from which three templates are derived to account for
the variability in orientation of the leaves. Interface from the eCognition 9.0 (Release Notes, 2014, 3).


3.2.1. Classified by training
One of the most well-known types of supervised classification is, as discussed in the literature
review, template matching. Templates for classification can be based directly on the training

34

of the identification of the centre locations of training features (fig. 25) or by knowledge
based training which was used in archaeological case studies presented in the literature
review. The drawback of both cases is the feature orientation which is why figure 25 shows
3 tree templates and archaeological examples are focused on symmetrical shapes such as
round barrows and kiln sites. Both will result in as many templates and variations as found
useful by the user, which may result in high computation time or at least some labour (fig.
26). Furthermore, the knowledge based templates will not include unaccounted feature
variation which may be included in the direct training if an adequate amount of training data
is selected and templates are created.


Figure 26 Hillshade maps for (a) possibilities of disturbed kiln sites with diameters of 12, 18 and 24m and other
surface structures on a sloping (left side) and horizontal (right side) surface, and (b) a synthetic DTM representing
kiln sites with diameters of 8–28m without noise and with low, moderate and high Gaussian noise (Schneider et
al., 2015, 49).

Besides template matching there are approaches which allow for more feature diversity,
of which the Artificial Neural network (ANN) was applied by Redfern (1997a) for
archaeological classification scheme. ANN simulators are computerised representations of
specific problem domains which function in a way similar to our understanding of how the
human brain operates. ANNs are trained to accept a set of input stimuli (e.g. a set of
measurements) and to respond with one or more numerical outputs (Bishop, 2006). There
are various kinds of ANN to apply but in general they are defined by one or more layers of
nodes (artificial neurons) and through this process some weighted summation of the input
data feeds through them, referred to as a feed-forward architecture. The ANN topology,
multi-layer perceptron (MLP), chosen by Redfern (1997a) was, at the time, found most
successful for pattern recognition tasks. It specifies that the nodes in each layer are
connected to each node in the preceding layer. In common 3 layers of nodes are used, in
which the second layer has no connections to the 'outside world', and is referred to as a

35

hidden layer (fig. 27). It is this layer which aids the internal complexity to enhance recognition
of for complex patterns. For accurate matching of the input and output values, weights at
each node are modified iteratively during the training of the MLP, referred to as
backpropagation of errors from the output layer through the preceding layers. In order to
apply the developed typology as an automatic interpretation task, the MLP was trained with
a large amount of data to the point that it almost perfectly reproduces this training data,
referred to as overfitting. Yet different from this method the monument classification
scheme intended to group the output of similar features together (fig. 28). From the 123
classified monuments, 80 where used for training and 45 (unseen) were used to determine
the point of optimal generalisation. In this experiment all 45 monuments were rightfully
classified.


Figure 27 (left) A basic, three-layer neural network topology, with a hidden layer (Barceló, 2008, 150)

Figure 28 (right) The hierarchy of MLPs used to classify monuments. Rectangles represent neural networks, while
ovals represent final classification decisions (Redfern et al., 1998, 14) .

The supervised method behind the landslide classification created by van den Eeckhaut
et al. (2012) is SVM (Support Vector Machine). This algorithm analyses training data and
recognizes patterns, it is used for classification and regression analysis (eCognition user guide
2015, 132). The standard SVM takes a set of input data and predicts, for each given input,
which of two possible classes the input is a member of. In the segmentation layer of the
landslide classification 12 objects were sampled and also expected false positives for each
class in the 10 km2 test area. These objects were used to train the models, while other objects
were considered testing samples. Object features included in the SVM were mean segment
values (slope, plan curvature, openness and roughness), standard deviation of segment
values (slope and plan curvature), and segment width. The training was executed on a zone,
square-shaped with twice the landslide’s length and width, to include context information.
Resulting from the analysis 71% of the landslides mapped by experts were detected and also
more than 50% of the landslide body was correctly identified. The surface morphology of

36

landslides not detected were those too subdued and affected by anthropogenic
interventions, such as construction of houses and roads in the lower deforested part of the
landslides. This result yielded similar accuracy to the analysis using aerial photography and
was therefore very successful.
Another example of landslide classification was presented by Stumpf and Kerle (2011)
who applied Random Forest (RF) algorithm. RF are an extension of decision trees from which
it was found that small changes in the training data induce a high variance in single
classification trees and often leads to rather low classification accuracies. The underlying idea
of RFs is to grow multiple decision trees on random subsets of the training data and related
variables. For the classification of previously unseen data, RFs take advantage of the high
variance among individual trees, letting each tree vote for the class membership, and
assigning the respective class according to the majority of the votes (Stumpf and Kerle, 2011,
2569). For landslide classification they used landslide inventories to create a sample database
with all objects assigned either as landslide objects (OLS) or non-landslide objects (ONLS, fig.
29b). Of each image object spectral features such as mean brightness were calculated and
used to apply in the RF algorithm at different scales. Eventually by using only 20% of the data
for training, the method resulted in accuracies between 73% and 87% for the affected areas.


Figure 29 Overview of the processing steps for the Random Forest algorithm (Stumpf and Kerle, 2011, 2568).

37

3.2.2. From supervised classification to supervised learning
While the previous mentioned case studies are based on passive learning, advances in
computer vision have allowed for active learning. A problem with passive learning is that the
user is unaware of which samples are valuable for the training data. Besides that, for
landslide mapping, it was recognised that collection of training data is a lot of work so the
aim was set out to focus the sampling efforts on a few but relevant spatial subsets in the
analysed scene (Stumpf et al., 2012). In practice this means that not all of the training data
would be used in the beginning and the model then decides which point to choose next. This
choice then enforces uncertainty of the picked training data to account for the variety within
the subset. This is different from the previous application of RFs because the manual
resampling causes uncertain results with small subsets of data, and also valuable samples of
the majority class may be excluded where variety is advocated by the AL approach. Stumpf
et al. (2012, 274) applied the Query-by-committee algorithm which uses the disagreement
among ensemble members as an uncertainty measure and queries the label of the most
uncertain samples. The uncertainty can be measured as the entropy 𝐻 of the ensemble votes
and for a binary decision is calculated by;
𝑣= 𝑣= 𝑣5 𝑣5
𝐻=− log + log
𝑛 𝑛 𝑛 𝑛
Where 𝑣= is the number of votes for the first class, 𝑣5 the number of votes for the second
class and 𝑛 corresponds to the total number of trees in the ensemble. So, from the first
classification the algorithm knows which locations in the result it is least certain about and
these locations are requested from the user to provide with a classification label. This second
round of classification can be done by the expert who is using the program, or when available,
a reference database and this process goes on iteratively until the desired accuracy is
required. Such a classification can be done per segment/image object or by region which
would take context information into account, and comes closer to human cognition. This was
noted by Stumpf et al. (2014) who further developed the method including a sliding window
which measures the uncertainty within each window. All together it can be stated that the
addition of AL to the supervised approach has evolved to be essential when extensive
reference databases are unavailable or when hazard events need prompt reaction with
localised classification.

3.2.3. Minor training, self-learning algorithms
The last algorithm to be discussed here is the deep convolutional neural network (CNN), first
presented by Krizhevsky et al. (2012), which is also supervised and learns from multiple levels

38

of representation and abstraction that help to make sense of data such as images, sound,
and text. The method better reproduces the cognitive process as the brain has a deep
architecture which also brings machine learning closer to the initial aims of AI. The basis of
this method lies in the convolutional neural network, which has been widely applied to image
data (Bishop, 2006, 267). This method divides an image into planes which share the same
weights of values. The aim is to match similar planes, or patterns, at different locations in the
input image. Due to the weight sharing, the evaluation of the activations of these planes is
equivalent to a convolution of the image pixel intensities with a ‘kernel’ comprising the
weight parameters. The whole network can be trained by error minimization using
backpropagation to evaluate the gradient of the error function (similar to previously
discussed for MLP applied by (Redfern, 1997a)). Different from this CNN the ‘deep’ version
they had a primary concern to prevent overfitting, so the effect they observed is faster
learning. Krizhevsky et al. (2012) used the ImageNet contest dataset to train an algorithm for
object recognition and classified 1.2 million high-resolution images into 100 different classes
with efficient use of two GPU’s (Graphics Processor Unit – fig. 30-31). They allowed the
process to take 5-6 days to finish and drastically reduced the state-of-the-art error rate from
26.1% to 15.3% and had no doubt that increasing computational power would improve that
very quickly.


Figure 30 An illustration of the architecture of the CNN, explicitly showing the delineation of responsibilities
between the two GPUs. One GPU runs the layer-parts at the top of the figure while the other runs the layer-parts
at the bottom. The GPUs communicate only at certain layers (Krizhevsky et al., 2012, 5).


Figure 31 Results of test images and labels found most probable by the model (after Krizhevsky et al., 2012, 8).

39

As this method is especially successful on big datasets and requires only limited training
this could be very promising for remote sensing. It is now starting to be recognised with the
publication of some case studies by for example Hu et al. (2015) who compared the
classification of land cover from hyperspectral aerial photography using SVM and CNN (fig.
32-33). They found an overall increase in accuracy of about 3% and expect that further
refinement of the applied parameters will only increase the accuracy of the method. For
archaeology this algorithm could also become very useful as experts naturally differentiate
archaeological features in different, hierarchical layers. Ideally, such a program could
classify an agricultural field in one layer and identify focussed crop stress in the next which
might hold archaeological structures. The labelled crop stress could then be run through
another database layer which resembles possible archaeological indicators, such as ditch or
wall, from which the variation in representation have been learned as well. Finally, the
classified indicators could iteratively explore the context at different scales and attest
possible combinations of indicators to the database layer of learned classifications of
archaeological structures.


Figure 32 Classification accuracies versus number of training samples (after Hu et al., 2015, 10)

Figure 33 Classification of agriculture. From left to right: ground truth, RBF-SVM, and CNN (Hu et al., 2015, 7).

40

3.3. The struggle of AI in archaeology
The first steps towards AI can be derived from the law based formalization by ‘new
archaeologists’ in the 1970s who argued for the acquisition of archaeological knowledge in
terms of sets of laws, correlates and cultural and natural transformation processes (Barceló,
2009, 100). These laws were, however, not a good representation of archaeological
reasoning and resulted in a reaction from Gardin (1980) who proposed the definition of rules
which are meant to be broken with science intrinsically moving forward. By the
deconstruction of scientific papers he looked for the necessary bridges between facts and
theories and the links between explanations in order to discover logistics and expose possible
flaws in argumentation (fig. 34). When formalising these ideas to be implemented in a
computer program he suggested that decision rules could be written to link visual features
with their causal process (social activity). Based on the degree of match, the candidate causal
event is selected or rejected.


Figure 34 (left) Original relief of a Seljuk sultan from the Konya museum. (right) Example of a decomposed image
derived from the relief (Gardin, 1980, 109 & 115)).

Nevertheless this and other proposed methods were accused of excessive simplification,
of forcing knowledge, or distorting it, and of failing to exploit fully the knowledge of the
expert. Stutt and Shennan (1990) critically reflected that these experts are largely
prescriptive and use highly formalized terms set out to justify a claim or decision rather than
to lay out the argumentation and thus set up unrealistic criteria which archaeological
explanations are supposed to satisfy. Yet they do argue that Gardin and other colleagues
advanced their understanding of how to model archaeological interpretation. In response to
solve these inadequacies they generated the WORkbench Supporting Archaeological
Argument Exploration (WORSAAE) system which used hypertext to interact and clearly

41

articulate the structure of the individual argument components as well as of the overall
argument as it proceeds. Their aim was to include various kinds of knowledge, such as
domain knowledge and common sense, to account for the complete process towards critical
interpretations of archaeologists. This knowledge forms unconscious rules which govern the
creation of archaeological statements and these could be critically reflection within various
stages of the mind and hence their system.
Barceló (2009, 97) suggests that these knowledge-based systems were useful tools,
although they do not simulate how humans actually think. Instead of storing declarative
sentences in the computer’s memory, he advocates to build a machine which is able to learn
from its own assumptions and mistakes. Only by this method could the field of study go
beyond the expert-system approach which focusses on database consultation towards
learning and categorizing which is now possible through a new generation of adaptive
algorithms. Although in landscape pattern recognition the only AI example was proposed by
Redfern (1997a), other fields within archaeology implemented some AI systems.
For example, attempts have been made to create an associative memory for
archaeological explanations by identification-based analysis (Barceló, 2008, 195). These
could be derived from descriptive regularities within a single population of objects as was
implemented by Barceló (1995, fig. 35) for the classification of Late Bronze Age warrior
stelae, using ANN. Although the individual iconographic motifs would not be recognised by
the model, descriptive concepts such as “relevance of shield” would be given by a user to
eventually derive patterns in the model.


Figure 35 Graphical representation of a Late Bronze Age Warrior Stela, compared to the vectorial notation for
input and output. Many of these formalisations would be used in NN (Barceló 1995).

42

For the actual attempt to include human perception of vision, various models have been
created for modelling distinctive textures of archaeological materials. For example, a
classification system for lithic tools was created based on the texture found on the surface
micro-topography (Barceló, 2008, 230). Derived from the assumption that artefacts have
surface properties (e.g. polished areas, scars) because of the way they have been made, or
the way they have been used, archaeologists define a meaning. This process can be recreated
with an ANN in similar ways to those presented by the geoscientists; with parameters for
shapes, sized and texture (fig. 36).


Figure 36 A neural network to recognize visual textures as use-wear patterns in lithic tools (Barceló, 2008, 241).

The current status of AI research for archaeology has not significantly changed from the
level reached in previous examples. Yet unlike the process of modelling individual
components used in AI models, researchers in archaeology have defined the importance of
semantic relationships between properties which define an object or feature. The way that
semantics are currently perceived in archaeology is mainly through the web-based semantic
formalisations. This is implemented in various ways ranging from structuring grey literature
to the definition of universal interpretations to make the same feature interoperable
amongst databases, each with a Uniform Resource Identifier, URI (linked to the online web).
In the session 2D “Linked Data: from interoperable to interoperable” of the Computer
Applications and Quantitative Methods in Archaeology 2015 conference several of these case
studies were presented (Isaksen and May, 2015). In relation to landscape archaeology Cripps

43

and Tudhope (2015) presented a method for generating geosemantics specifically defined
for hierarchical geospatial relationships between archaeological features. Even though their
aim is not specifically related to pattern recognition it could become a useful addition to the
automation. Although not specifically hierarchical, the classification system integrated in the
AERIAL software by Redfern (1997b) also included tags for interpretations (e.g. mound, ditch,
bank) used for training of the ANN model. One of the specified aims for developing
automated applications mentioned by Bennett et al. (2014, 901) was to optimise metadata.
Part of this would be a categorisation system to explain and qualify processing steps to
improve the intelligibility of and confidence in the results of semi-automated techniques.
Additionally, they proposed that feature recording protocols should be expanded to include
complementary environmental data such as land use, land management strategies, dates of
feature visibility and proportion of the known feature visible.
These various side paths of AI could be integrated together with the geographical data
differentiating various hierarchical landscape features. To illustrate, agriculture could be
separated from roads and buildings at one level and the designation of the kind of crops
grown at the second level and the state of the agricultural cycle at the final level.
Furthermore, cross related knowledge on the meaning of crop stress and the formation and
appearance of archaeological structures could be integrated. An example of how nodes in
such a basic semantic model would be interrelated can be seen in figure 37. Once this kind
of data is accurately and consistently documented for ‘training data’ it could allow for deeper
layers within ANN to be trained for the detection of patterns using deep convolutional
networks to expand classification in other parts of the landscape.

has landcover

possibly
Agriculture (partly)
levelled
Round barrow is type of Barrow is type of Earthwork

is defined by is type of
Flora
possibly
possibly possibly
(partly) Mound surrounded by
Ditch surrounded by
Bank
destroyed

Fauna has shape has size


Round
(varied sizes)


Figure 37 Semantic description of a round barrow.

44

Chapter 4: Pilot study on barrow detection
To further explore the opportunity of using AI in archaeology it has been chosen to revise the
use of GEOBIA for feature recognition. Although some scholars have suggested that it is not
possible to capture the unique properties of archaeological features, new features within the
software and advanced case studies (of which some were presented in chapter 3) require a
revision of these statements. For many pattern recognition studies barrows have been the
focus, as will they be here. Barrows, also called burial mounds, and amongst the most
common monuments of prehistory all over the world; and therefore a uniformly working
automated technique would be very valuable. As a case study, an area around Avebury,
United Kingdom, has been chosen for which VHR multispectral aerial photography and 1
meter resolution LiDAR derived DEM is available. The south of England is very rich in Bronze
Age barrows, they are to be differentiated into 5 types which imposes additional challenges
for an automated detection method (fig. 38). Additionally, barrows are also renowned for
their similar appearance on remote sensing data to naturally occurring elements (e.g. fairy
ring) and modern human-made structures (e.g. roundabout), which are likely to cause false
positives in this areas. The aim of developing an AI model is to include the knowledge of an
archaeological remote sensing interpreter of the area and features to develop a more
insightful pattern recognition. As this aim is more wide ranging than a master’s dissertation,
the objective is to lay a foundation for future research. So, rather than presenting an
‘exhausted’ account of one possible method, this chapter will present some basic analysis
and workflows to include AI for archaeological feature detection in eCognition.


Figure 38 Various barrow types present in the South of England (after Stonehenge World Heritage Site, accessed
01/07/2015)

45

4.1. Data sources
In most of the previous barrow detection models LiDAR has been used, however,
documentation from Historic England shows that some barrows are only visible as crop marks.
This is generally the result of extensive ploughing which has levelled the barrow which makes
them undetectable for LiDAR data (fig. 39 Crutchley, 2006, 254). For this case study, data
from both aerial photographs and LiDAR sensors have been acquired. As mentioned in the
chapter 2 choosing the right remote sensing technique for the detection of desired features
is very important as well as knowledge about the acquisition time of the appearance of
features. Besides the remote sensing data, locations of known barrows were acquired to use
for training locations and comparison of the results.
The LiDAR data used for the analysis were provided by the Environment Agency, UK. The
data was captured on the 3th of May 2006, filtered from surface features and resampled to
0.5m resolution. It was provided in 90 tiles which were mosaicked in ESRI ArcMap 10.2.2 to
generate a seamless DTM. The data was found to be of high resolution and shows slightly
elevated earthworks; although as said before data capture in the winter would be a better
time as there are fewer leaves on the trees and other vegetation to blur images. Nevertheless,
the area is not generally characterised by forest so overall slight earthworks should be
recognised.
The aerial photography was provided by the Ordinance Survey, the National Mapping
Agency of England. Unfortunately this data was provided at a final stage when most of the
analysis was already concluded. Therefor only a small area has been used to briefly show
some abilities of aerial photography in eCognition.
Point locations and digitised features in the area have been provided by Historic England.
These have been acquired over a long period and were brought together during the National
Mapping Programme and are mainly based on aerial photography and some LiDAR (NMP
discussed in section 2.5). The data was provided together with a complete account of the
archaeological features with each having a description of their manner of acquisition and
certainty of detection. This information was queried on the various kinds of round barrows
to generate feature specific shape files.
Other data sources for visual verification of detections, such as the OS first edition, were
downloaded from EDINA Digimap and reports of local surveys by the Wiltshire Archaeology
and Natural History Society’s (WANHS) were downloaded from the Wiltshire Museum
website (accessed 04/09/2015).

46


Figure 39 (left) Barlings Abbey barrow cemetery. In this photo the barrows surviving as earthworks in the field
centre-left do not show clearly, but other barrows are strikingly visible as light soil patches in the centre of the
image. (NMR12606/15, 05-SEP-1994. © Crown copyright. NMR (right) Lidar data. Even with the height
exaggerated 20 times many of the barrows are still very unclear. (Lidar courtesy of Lincolnshire County Council:
source, Environment Agency, March 2001. © English Heritage. NMR - (Crutchley, 2006, 254)).


4.2. Methodology
The three different types of analysis that have been undertaken conform to some of the
previously discussed analysis. The first is based on cognitive analysis for which GEOBIA is
most renowned while the other two are based on machine learning. The analyses have been
undertaken in eCognition Developer 9.1 from Trimble and snapshots of the analysis have
been included for clarification. Furthermore, the path of trial and error are portrayed with
screenshots added to the appendices 1-3 and meta data of the project has been included
with the submission, mainly including generated eCognition rulesets and exported results in
shape files, compliant with geospatial standards. For illustration, maps are presented in this
chapter using ArcMap 10.2.2 referenced to the British National Grid with OSGB 1936 datum.
In order to provide a clear overview it has been chosen to maintain consistency of
parameters between examples without adding too much detail. The analysis of the LiDAR
data has been undertaken using slope derivative (for explanation of slope see subsection
2.2.2). As discussed before (3.1.1), the first process step of each GEOBIA analysis the image
is segmented into objects of similar pixel values. Within the GEOBIA community it is advised
to create small segments which could be merged at a later stage according to set thresholds
and this convention has been followed in this pilot study. As each method required some sort
of training area, whether to train cognitive analysis or to select training samples, two subset
areas have been selected for this purpose (fig. 40). Another important part of the potential
assessment is the computation time and amount of false positives (as discussed in section
2.5), so a close account of this has been kept.

47


Figure 40 Avebury case study area depicted with LiDAR-derived slope map and indications of sub areas which
were used to derive parameters, select samples and train classifier from the known barrows.


4.2.1. Cognitive argumentation
In order to generate a workflow based on cognitive rules it is important to formalise the
feature characteristics at the start of the analysis. In the Avebury three barrow types have
been found of which the slope derivative can be seen in figure 41. Within this it can be seen
that the barrows stand apart from their surroundings by their brightness and are of roughly
the same size and shape. For a first test phase the analysis was undertaken for Overton Hill
where a threshold for the brightness was defined to characterise the outline of the barrow,
these objects were eventually merged. This process resulted in various false positives which
could be excluded by additional threshold variables for the area and the length/width ratio.
Once the results were determined to be satisfactory the same parameters were undertaken
on the complete case study area. The complete workflow is visualised in figure 42. The final
classification was exported to ArcMap to verify the results with the database from Historic
England. The verification is based on the assumption that all features in the area have been
identified and no new features are available.

48


Figure 41 Three barrow types depicted on slope derivative; (left) Bowl (middle) Saucer (right) Bell.

Iterate process

Define
Open test Image Define
segmentation
image segmentation features
threshold

Iterate process

Generate Review
Classify features Add threshold
threshold classification

Open
Evaluate Export
verification Apply ruleset
result classification
image

Iterate process

Figure 42 Cognitive analysis flow diagram.


4.2.2. Towards machine learning
Within this dissertation it has been noted that template matching generally results in
numerous false positives. While, in areas with constant landscape modification templates
should account for some derivation, this could build up quickly resulting in a lot of additional
manual labour. Within eCognition another kind of template matching is available. This
method derives the mean of samples and hence actively updates templates with the addition
of new samples. Additionally it allows the user to select negative templates of commonly
occurring false positives. During the process the generated templates could be tested on
their correlation to amongst the samples. The false positive rates can finally be checked in an
interactive way of lowering the threshold and revising the error rates. As one proceeds
adding more templates typically error rates go down or samples which increase the error
could be excluded. To account for variable features it is possible to generate multiple
templates which combine similar features. The complete workflow is visualised in figure 43.

49

Open test Sample Generate
Test template
image selection template

Iterate process
or add negative template of false positives

Open
Define Review Update
verification
threshold targets template
image

Iterate process

Create Evaluate Execute Export


correlation map correlation classification classification

Iterate process

Figure 43 Template matching flow diagram.

Although machine learning requires more prior knowledge than the other two methods the
most basic workflow will be presented here. All of the locations of known barrows at Overton
and Windmill Hill are used to generate the machine learning. For this case study Random
Forests was applied as this worked well in the case study presented by Stumpf and Kerle
(2011, discussed in 3.2.1.). Although it is generally advised to use at least 100 samples only
22 were available. Additionally, active learning could provide more knowledge on the error
rates of samples to provide the best data for the learning phase. An important part of this
process is the identification of the most stable parameters, and as brightness was chosen for
the cognitive modelling this was used. The complete workflow is visualised in figure 44.

Iterate process

Define
Open test Image
segmentation
image segmentation
threshold

Open test Assign class to Train RF Apply RF


features test features classifier classifier

Open
Review Export
verification Apply ruleset
classification classification
image

Iterate process
Figure 44 Machine learning flow diagram.

50

4.3. Preliminary results
As only the basics of the applied methods have been explored the results are not fully
representative of them and will only be briefly discussed. Although each of the analysis was
successful on the small areas the software seemed to crash or display error messages once
it was performed on the full case study area. It was discovered that this was not caused by
the size of the area but the presence of NoData values. Most of the analysis time was devoted
to the segmentation into objects, which took 20 minutes, yet this is only done once and every
classification or threshold variables afterwards took under 10 minutes.

4.3.1. Cognitive argumentation – easy and sufficient
The cognitive analysis was simple to undertake and yielded the best preliminary result. The
results were exported to ArcMap for verification with the known round barrows from Historic
England (table 4). These known locations still include levelled barrows which were already
impossible to detect by visual inspection, and they probably take up a large part of the 163
false negatives. In total 50 barrows out of 213 were detected. In figure 45 can be seen how
these barrows were compared based on their detected or digitised area. This process has
caused for some unnecessary false positives as the digitisation of historic England did not
fully cover the brightness classified by the algorithm. In the future this could be improved in
the future by including a buffer when assessing the number of false positives.
When comparing the type of barrows detected it is clear that the applied parameters are
especially successful for the detection of bell barrows (table 5, figure 41 could be a reminder
of their brightness). Further research may also be able to define more specific parameters
for the detection of the remaining barrow types.

FALSE POSITIVE FALSE NEGATIVE TRUE POSITIVE
DETECTIONS 196 163 50
Table 4 Results from the cognitive analysis compared with the known barrows based on the assumption that the
repository from Historic England holds all barrows.


OTHER (ROUND) SAUCER BELL BOWL
TRUE POSITIVE 10 3 14 23
FALSE NEGATIVE 76 7 6 74
PERECENTAGE P/N 12% 30% 70% 24%
Table 5 Evaluation of the rightfully detected barrow types show that the parameters work best for bell barrows

51


Figure 45 Visual depiction of true/false positive detections and false negatives of undetected known barrow
locations. Location: east of Windmill hill and north of Overton hill.

Given the fact that the results might include new detections, some locations were briefly
scanned and compared to the slope derivate, the OS first edition and local field surveys from
which it was noted that a concentration of ‘false positives’ was located in the south east of
the case study area, under the West Wood tree canopy (fig. 46-47). As it is a known fact that
aerial photography and field surveys are difficult to undertake in the forest, LiDAR data forms
a key addition to the detection of subtle earthworks. Even though the NMP used some of the
LiDAR data, in this area they only noted the extensive field system next to the forest. Further
ground truthing in the area would allow for the verification of the possible barrow locations.

52


Figure 46 With red circles the visual interpretation of possible barrows are indicated. The 6 bright features in the
top appear like bowl barrows and show up throughout the forest. The 5 indications below are of the same size but
appear less prominent and might be slightly levelled barrows (reference image on the topleft from Google Earth).


Figure 47 In the top left the detection method found a pit structure and number 102 is a barrow which was
indicated by both the OS-first edition and the topographic survey undertaken by the (WANHS) but missed by HE.

53

4.3.2. Towards machine learning – large effort but great potential
The results of machine learning have been explored in less detail as it was discovered that
the methods need more refinement before they can be actually verified.
Template matching, for instance, resulted in most detections being concentrated along
the ridges of hills. Supposedly, this reveals that the samples were selected at different angles
to the sloping terrain which resulted in averaged templates of the surrounding matrix rather
than the mound itself. Enhancement could be further researched by using other types of
derivatives which are not affected by the sloping terrain. Also, the samples were selected
with the surrounding matrix to account for their different sizes. For the future an approach
using different template sizes, advised in Trier et al. (2015), could generate a better
representation of the known variety. Furthermore, additional different templates could be
created for the barrow types and more samples could be integrated to actually account for
imperfections, like Schneider et al. (2015) advised. In the verification process a 0.55 threshold
was selected, but only trial and error could define the best threshold for the analysis (fig. 48).


Figure 48 Snapshot of barrow verification. Image on the right shows the correlation of the pixel to the templates,
in this case white declares higher certainty levels.

The results of the RF algorithm can be seen in figure 49 which shows the numerous false
positives caused by many insufficiencies in the applied methodology. First of all, the method
was applied to the individually created objects which lack the unique characteristics of the
complete barrows. Context information could improve this for the GEOBIA approach or an
object based analysis could also be applied in eCognition. Additionally, the detection method
does not have various stages like the cognitive analysis which means that the area and

54

length/width are executed on individual objects and are hence not able to provide the same
analysis. Extensive trial and error to derive the best variables and LiDAR derivatives, like
advised by Stumpf and Kerle (2011), and the addition of active learning would be a next step.


Figure 49 Snapshots from the application of RF in eCognition showing the known classification below and the
outcome of the machine learning above. Left side shows the general application of all barrows and right shows
the outcome specifically per type of barrow (location: Overton Hill).

55

4.4. Potential of eCognition for archaeology
In order to create a good working method it is important that explanations should be based
on purposeful, goal-directed mechanisms emerging from a dynamical system that has been
calibrated by learning (trial and error, experimentation, analogy) to make the right choices in
the proper circumstances (Barceló, 2009, 106). Within this pilot study this aim has not been
reached but it could be further explored. Yet even the very basic analysis of three methods
in eCognition has shown that barrows can be detected. With the knowledge of fully
developed methods discussed in the sections 3.2 and 3.3 it can be stated that adjustments
will significantly improve the methods. Especially by using other data layers of LiDAR
derivatives and aerial photography. For automated detection using LiDAR it has already been
recognised (Hesse, 2015, personal communication) that it would be very promising to
undertake such analysis on visualisation techniques such as those discussed in 2.2.2. (Hesse,
2010, Kokalj et al., 2011, Doneus, 2013). Also, closer assessment of the known barrows, their
sizes and assigned certainty, and specific digitisation of the mound, ditch and bank could
account better for the variability within known types. This would require more experience in
interpreting the appearance of certain features on the data layers. This conclusion is a lot
different from the previously applied methods in archaeology and shows not only a different
method than the conventional template matching but also shows success in implementing
GEOBIA concepts. Even though it was previously stated that the latter could not capture the
unique characteristics of archaeological features (de Laet 2007, 840).
In a future when more researchers apply methods in eCognition for feature detection,
created rulesets could be easily exchanged so that similar features can be detected and small
adjustments could be made to fit the analysis to other areas.

4.4.1 Comparing eCognition to alternatives
Although the use of the GEOBIA eCognition software is advocated here, there are some
disadvantages to its use. Firstly, a permanent license is quite expensive, maximum €2800,
although educational discounts apply (Lorenzo, 2015, personal communication). Secondly, it
can take for quite a while to learn and comprehend the concepts of GEOBIA. These factors
have also been recognised by the software provider and therefore a free trial version is
available for one month which allows the user to get accustomed to most of the available
analysis steps, although it does not allow one to export results. Also, there are different kinds
of licenses available at different prices ranging from Architect, for the mere application of
previously developed rulesets, to Developer, for the development of rulesets, to Server, for

56

the addition of batch processing large projects. Software users also have access to an online
Community which provides additional resources, such as wiki pages, status messages,
tutorials and discussions. Within the discussions users exchange rulesets for general tasks
and also interact on the shortcomings of the software which are generally integrated or
improved in software updates.
Open source alternatives are yet to provide a successful alternative to the software’s
overall abilities but for some tasks users have presented alternatives. Stumpf and Kerle
(2011) for instance, extensively used the randomForest package and its extension for variable
selection varSelRF implemented in the R statistical programming environment (R-
Development-Core-Team, 2009). In their follow up research, Stumpf et al. (2012) only used
eCognition for the first segmentation stages and then turned to R for the RF algorithm and
the active learning. As archaeologists are getting more familiar with the use of R (Baxter,
2015) an integration of pixel based machine learning could be a great improvement to
current applications. Also, the automated detection routine for template matching by
Schneider et al. (2015) was implemented in ArcGIS using functions from ArcPy and the image
processing module from the OpenCV (Open Source Computer Vision, Version 2.4, OpenCV
Development Team, 2014) library. Additionally, the generation of software specifically for
archaeology has its benefits; such as Redfern (1997b) developing AERIAL while Trier et al.
(2015) developed software specified for their application of template matching (Trier, 2015,
personal communication). Other image enhancement software created specifically for
archaeological purpose such as ARCTIS (Atzberger et al., 2014) for aerial photography, and
LivT (Hesse 2014), for LiDAR visualisations are very helpful for development of further
methods.

57

58

Chapter 5: Discussion and future scope
Semi-automated pattern recognition in archaeology has encountered a lot of criticism since
its introduction and is yet to become an accepted source for archaeological landscape
research. Despite the valid argument that archaeology is intrinsically variable and hence
difficult to extract from remote sensing data, similar difficulties have hampered geoscientists.
Instead of rejecting complex features they recognised the challenge which they have
resolved using AI concepts. Using both the hierarchy in visual recognition of features and the
process of learning from previous examples they have been successful in accurately mapping
complex features. Furthermore, their concepts are ever developing to further increase the
accuracy and speed of mapping, and to determine the most cost-effective methods. This is
possible due to developments within machine learning algorithms created by computer
vision and therefore they have kept a close eye on developments and evaluated how these
could be effectively integrated.
Throughout this dissertation examples of pattern recognition together with a pilot study
have been introduced and in this chapter the potential of these for archaeology will be
critically reflected upon. Additionally, the future scope will be discussed to evaluate in which
field of landscape archaeology AI could be of most value and how the methods could be
integrated as a tool to all of remote sensing research.

5.1. AI in reaction to the data explosion


Increasing availability of remote sensing is an advantage which should be embraced, yet it
requires an approach beyond visual interpretation. Concluding from the second chapter the
following areas of improvement were listed to enable pattern recognition with AI;
i. Ability to replicate the human cognitive argumentation
ii. Ability to integrate data from multiple remote sensing sensors
iii. Ability to integrate complementary environmental data
iv. Ability to detect variability within features
v. Ability to detect variability within the appearance of features
vi. Ability to generate data categorisation and record metadata
vii. Ability to use the data categorisation for learning algorithms
viii. Ability to apply the method over large areas

In this section, the potential of AI to form an adequate reaction to the data explosion will be
discussed. Arguments from archaeologist about how to handle the increasing amount of data

59

will be evaluated against the solutions proposed in geosciences and computer vision to
efficiently analyse big datasets of very high resolution.

5.1.1. Alternating hypes – benefit or burden?
Unlike the many decades only using aerial photography as a source for remote sensing, the
last decade has been characterised by the rise of many different techniques. In this
dissertation only the surface of abilities of VHR satellite images, LiDAR and airborne
hyperspectral scanning has been scratched. Despite that, other immerging hypes have not
even been introduced, such as the rise of UAVs (Unmanned Aerial Vehicles) for low altitude
scanning, which can be instrumented with both photo cameras and LiDAR resulting in more
detail than ever before. Yet the growing availability of different sensor types, resolution and
spectrum also provide some obstacles. Parcak (2009, 239) rhetorically asked; “Is satellite
technology advancing faster than archaeologists’ ability to learn, apply, and analyse the data
and programs, and all the inherent implications?”. She emphasises that this obstacle is
triggered by the delay of book publishing and further proposes another way of publishing, an
online journal, to present case studies with useful advances to rapidly update the community.
Although this seems like a very important step forward, the focus to explore the full extent
of a method should not be underestimated. Only when significant time is invested will the
reasons behind the appearance of features be deduced, as it was critically reflected for aerial
photography (Palmer, 2005). This was successfully done by Verhoeven et al. (2013), who
studied the plants physiology visible in NIR InfraRed anomalies which led towards a better
understanding of hyperspectral imaging for crop mark detection. This knowledge enabled
the creation of a toolbox with meaningful filters to emphasise relevant features (Atzberger
et al., 2014). This is the essential interpreter’s knowledge which is necessary to actively refine
the machine learning algorithms. Even though it both takes time to develop another AI
technique for a different sensor and also to learn, apply and analyse the data derived for
improving the technique, eventually the scale of detection would enable this method to
become cost-effective.

5.1.2. Embracing multi-temporal data for heritage management
Multi-temporal data has been successfully used in archaeology to detect changes in known
features. As it was recognised that it is difficult for humans to visually assess changes
between two or more periods of data, digital image processing tools are largely applied for
this kind of research. Traditionally, temporal data can be used for two types of analysis, those

60

related to either short-term (e.g. looting) or long-term (e.g. levelling by agriculture)
phenomena. Long term analysis for example, has recently been used to assessed the
destructive impact of late twentieth-century development with time-change analysis of the
condition of known archaeological monuments but also to discover archaeological sites now
destroyed (Hanson and Oltean, 2013). An ‘automated’ change detection between historic
images and two DEM was presented by Risbøl et al. (2015). The changes were derived from
elevation models generated from both historic aerial photographs from the years 1968, 1979
and 1999, and from LiDAR datasets from 2008 and 2010. Their resulting difference models
were of decimetre accuracy which allowed for suitable visual interpretation, yet they also
acknowledged that knowledge of the areas studied is a prerequisite in terms of
understanding the identified changes. With the rapid increase of data capturing at higher
resolution in the last decades the previous analysis can be carried out with more detail to
define even slighter changes, but also to detect looting. The possibility of such methods was
researched by Lasaponara et al. (2014) who used satellite images to detect holes from
looting. By applying image processing to clearly emphasise the holes, unsupervised cluster
analysis was undertaken to detect the significant changes between images. The method
proved successful and could also in other areas could prevent the monitoring of sites
becoming financially prohibitive, a limiting factor mentioned by Parcak (2009, 224).
This kind of analysis have also been undertaken with cognitive analysis using GEOBIA as
illustrated in figure 50, although the analysis could be taken even further. Once
characteristics of certain types of destructions are known cognitive analysis could be
generated and if rules are difficult to define they may be able to be found using machine
learning algorithms. If such methods are further developed, in the future, monitoring could
be a computed task and there would be more time available for impact assessment, ground
truthing and action to hazards.


Figure 50 Workflow of the change analysis procedure (Niemeyer et al., 2008, 188)

61

5.1.3. Harmonising the struggle of AI
The introduction of AI to archaeology has not found a lot of support and is yet to be fully
accepted. Barceló (2009, 103) stated that “robots are not the guilty ones, humans don’t know
how to program them!”. Based on the research presented in this dissertation it could be
added that the failure of semi-automated pattern recognition in landscape archaeology is
not merely caused by limitations in software or algorithms, it is caused by the researchers
who settle with simple features after exploring the surface of possibilities, or reflecting on
researcher who have. Nonetheless it has been shown here that concepts of AI and
geosciences can be successfully applied to archaeology. Further exploration into the process
of archaeological interpretations could enhance the AI models significantly and possibly point
archaeologists into conclusions which are not self-evident, such as Gardin (1980) aspired with
his reflection on the decomposed Hayek theory (discussed in section 3.3.). In current practise
of remote sensing in archaeology, interpreters and aerial photographers will store a
compendium of types in their brain and these help to identify the more obvious features
(Palmer, 2005, 98). Yet experienced interpreters are biased towards their knowledge and
expectation which may be limiting, also it is a known fact that computers can store more
known features and combine additional information where humans need additional
resources. So, why not provide remote sensing archaeologists with an AI programme to
detect at least the locations of visible ditches and banks depicted in crop marks or earthworks
with perhaps some additional information on possible structures related to the visible range
of geometry, textures, spectrum and context.
Contrasting to the views of current researchers of AI in archaeology and geosciences it
has been mentioned here that the full-range recording and integration of different levels of
interpretation, context, temporal changes and semantic relationships could be valuable
training data for AI applications. Even though it is yet to be applicable, this may become
integrated in the coming years when deep convolutional networks reach the remote sensing
domain.

5.2. Future scope


In case we are able to reach advanced levels of semi-automated pattern recognition this
would require that heritage management bodies adopt the large scale use of methods for a
broad range of applications. In this section will be discussed how that future can be envisaged
together with the possible applications which would enable this.

62

5.2.1. Consistency in large mapping programmes
To this day most national mapping programmes are undertaken with an interpreter’s bias
from the decision of capturing an aerial photograph from the sky to their digitisation.
Although some advances are made towards more consistency. For example, the English
Heritage National Mapping Programme was compared to the Baden-Württemberg LiDAR
Mapping, Germany, in section 2.5. Using the proposed AI methods rulesets for general
pattern recognition tasks (e.g. detection of ditch and bank) could be exchanged to streamline
parts of the tasks. The compendium of features known to interpreters could be extracted in
a uniform matter, as discussed in section 2.3, and used for training data. Further exchange of
this data amongst every subdivision of heritage projects and beyond may possibly allow for
even the unexpected features to be detected.
Optimisation in batch processing of data could enable efficient use of the ever increasing
amount of remote sensing data. For example, data for geographic mapping projects may
have been acquired in those months that archaeological features are general not visible.
Manually browsing the data for features, at this point, would probably not be worthwhile as
they are mostly invisible. This kind of data is available through national mapping agencies
such as the Ordinance Survey of Great Britain which has captured a wealth of multi-
spectral/VHR images. In the scope of consistency this data ticks all the boxes with multi-
temporal, orthorectified images captured under highly oblique angles from 2000-today with
the same sensor covering all of the isles, which is yet to be used for temporal analysis in
archaeology. For the future they would be very interested to provide archaeologists with
their data to enable research for this optimisation (Holland, 2015, personal communication).
When locations are already known and have been successfully detected in one generation of
aerial photography or LiDAR data it can be quickly adapted to new types of data. This would
allow for accurate testing of usability of newly created additions to available data derivatives.
As discussed before, the choice of data derivative is important for successful pattern
recognition.

5.2.2. Web-based data repository
A solution to the large scale adoption of semi-automated methods is by optimising the
availability of the methods and to enable a uniform platform for discussion and exchange.
This could be created in a streamlined web and desktop application which would hold widely
relevant methods in a similar way as eCognition evokes discussions amongst likeminded
users. Different from merely being a discussion forum, this application could hold a
repository of data features to allow rapid growth of sometimes limited training data.

63

Especially in areas without extensive databases of feature locations, exchange is vital as
knowledge of features could still be sparse. A platform such as that proposed here would
improve the availability and consistency within training data used to feed algorithms. This
software should be based on general Human Computer Interaction (HCI) conventions but
also be specifically designed for the archaeological purpose. So, the following three concepts
should be integrated;

- Simple (e.g. LiVT for LiDAR image processing)


- Effective (e.g. relevant tools for archaeology)
- Useful (e.g. serve multiple functions)
- Accessible (e.g. not every archaeologist is comfortable using MATLAB)

Additionally in terms of archaeology, specifically the following concept should be a priority;

- Safety (e.g. protecting site locations to prevent looting at all security cost)

This could only become profitable if it is intended for large scale research and should hence
be generated together with future users. As large scale data is already available from national
mapping agencies and heritage bodies such informative layers could form the base of large
scale research extended with more detailed information from local authorities. For the
application in England, it has been shown in the pilot study that Historic England has digitised
polygons of site locations for the availability of which covers the whole country.
The accumulation of data would aid in making the program smarter but the return of new
possible detections also requires verification. An extension that would account for large scale
classification could by Citizen Science (CS). It was already mentioned by Bennett et al. (2014,
903) that crowd-sourced interpretation of LiDAR data could be integrated as an innovative
interpretation method. Yet unlike their justification for manual detection being the best
method, it would be implemented to verify detections. This could be done by the use of web-
based tools and further evaluation in the field as presented by Duckers (2013). At the root of
this project an open source WebGIS was integrated for the support of their geospatial
project. Various layers such as Google Earth and LiDAR derivatives could be consulted to
improve interpretation and tools were available to draw polygons of detections (fig. 51). A
similar system could be integrated for the web-based data repository, although the location
of new feature detections should be secured with a specific user login.
An online platform as such could be expanded for various kinds of archaeological projects
concerning image processing and AI tasks. As it was discussed in section 3.3. archaeologists
have various relevant fields in which AI could contribute to the availability of resources. Yet
none of these projects are brought into wide use beyond the research groups.

64


Figure 51 Screenshot of the WebGIS application user interface presented by Duckers (2013).

65

5.2.3. The aftermath of the data explosion
It has previously been stated, in the literature and this dissertation, that the increasing
availability of data requires another approach for efficient browsing of archaeological
features. Yet in the further future data will be continuously updated so it is important to
mention how this could all be consistently integrated into archaeological research.
For example, in the ideal future where all the remote sensing data is available online on
one website, a system should be integrated which automatically georeferences updates of
new remote sensing data. With every update, local heritage managers would be notified on
change detection and the extent of it so that they could adequately act upon it. Uncertain
locations could be checked with a field visit or otherwise drones could be sent out for
updating where remote sensing data is not updated soon enough. However, there are also
adverse effects, such as the availability of storage with the increasing availability of data.
Therefore it would be important to look into processing of LiDAR data directly from the cloud
and identifying semantic descriptions straight through deep learning.
Even though mostly archaeological applications are discussed, geosciences are also
concerned with the increase of remote sensing data. Even though most of the information
found to support this dissertation was based on the generation of rulesets and algorithms, it
is easy to imagine that they are also concerned with the development of similar AI systems.
The projects discussed mainly focussed on the classification of (landslide-) hazards from
which the results could be streamlined with the known archaeological sites to make
interdisciplinary responses. In cases where they would also derive risk assessments,
archaeological projects could derive locations where archaeology is at risk from hazards.

66

Chapter 6: Conclusion
Remote sensing has for a long time been a domain for aerial photographers and interpreters,
but with the recent data explosion archaeologists are able to request various kinds of remote
sensing data, at low cost, which is now a standard for many research or large scale mapping
projects. Yet this wealth of data has also triggered archaeologists to improve image
understanding and to further develop a desire for semi-automated methods to enable large
scale feature recognition.
The application of digital image processing has had a great impact on this. For instance,
archaeological structures are better recognised if the contrast between features and their
surroundings is enhanced. Besides enhancing images it has been used in attempts to extract
features by pattern recognition. This is especially true for the work of D'Orazio et al. (2015)
which has enabled extraction of irregular features at multiple levels, this is of importance for
digitising features in a consistent and unbiased manner. Looking to the future, the resulting
digitisations could be extracted and attested to an arrange of known site shapes, although
for now, the process of semi-automation remains a desired next step.
Various attempts have been made to classify ‘easy’ known shapes, such as circles and
lines in an image but the conclusion of the results is that most archaeological features will
remain too unique for automated recognition. This settlement with simple well-known
features has resulted in the emergence of state of the art template matching which now
includes templates of various sizes and accounted distortions, and an outcome provided with
certainty levels (Schneider et al., 2015, Trier et al., 2015). Although this method has
generated good results and can trace monuments which occur throughout the globe, the
method is optimised for local conditions based on the expected sizes and distortions. Also,
the known algorithms are created within a commercial context or in a specialised
programming software and both require a lot of manual input which might prevent it being
used by heritage managers elsewhere.
Repeatability of other presented shape algorithms is also limited because it requires
highly qualified training in mathematics, which is not common amongst archaeologists. Of
the known case studies it is rare to find follow-up research, supposedly money runs out or
results are unsatisfactory. This has resulted in a lack of experts and only a few researchers
leading the field (specifically the projects for of the Norwegian Directorate for Cultural
Heritage lead by Trier 2009-2015). These researchers are increasingly able to decrease
amounts of false positives but remain focussed on easy shapes and extensively occurring

67

features. The limited amount of researchers is this field also implies the absence of actual
discussion amongst users which could trigger the exploration of other methods.
The evaluation of other methods has been of central importance to this dissertation.
Through the exploration of developments in geosciences and computer vision in semi-
automated pattern recognition, new insights were acquired. The field of artificial intelligence
seemed especially valuable in bridging the existing gaps in the research of capturing the
unique characteristics of archaeological features. Although it was found that some work in
archaeology has been undertaken this mainly concerned studies of material culture (Barceló,
2008). Again this field is occupied by a single consistent researcher and only a few have made
an isolated attempt. Although some aspects of visual cognition have been included in this
sort of research, such as the texture on lithic tools, the field has yet to move beyond
individual components of an image and include hierarchical levels of composed knowledge.
The latter underlies semi-automated pattern recognition in landscape archaeology which
implies that other concepts ought to be investigated.
For the connection of cognitive concepts with computer vision algorithms, a
reconsideration of geographic object based image analysis was proposed. This concept was
rejected for archaeological feature detection after the failure of the method to explore
shadows revealing ancient walls (de Laet et al., 2007), although this method was based on
the most basic abilities. Since this publication the available software has further advanced
with more progressive applications, branched of as object oriented analysis, were able to
capture the unique characteristics of landscape features. This was first made possible by the
application of fuzzy rulesets and where the definition of rulesets is found difficult or time
consuming, learning algorithms have been applied. These analyses have allowed for the
detection of irregular features such as landslides and over the years, the machine learning
components have been developed even further. Specifically advantageous is the creation of
interactive computer programmes for iterative selection of training data, based on active
learning, where knowledge on features is sparse. These concepts have recently been further
developed in computer vision, allowing larger datasets to be analysed in shorter time by
reaching deeper levels of understanding. It has been suggested here that these algorithms
could be applied for archaeological use to recreate cognitive interpretation process derived
from given hierarchical metadata on the location, time, context, experience and multiple
data layers. If this multi-dimensional data AI-model could be created semi-automated
pattern recognition could be much closer to replicating an archaeological interpreter.
In a search for new applications to enable archaeological feature detection the potential
of eCognition, the software underlying most of the GEOBIA projects, was explored. By means

68

of a pilot study three different methods were evaluated on their basic workability and
possibility to expand into a more advanced analysis. With only a little effort a simple ruleset
was generated to extract 50 out of 213 round barrows based on visual description of the
appearance of these features on a LiDAR derived Slope image. The possible refinements of
this model are numerous with the available algorithms and integration of other data sources
would probably enable the detection of levelled barrows and exclusion of commonly
detected false positives. Beside these basic GEOBIA rulesets, the software offers other
abilities for template matching and machine learning. The template matching algorithm is
different than previously discussed in that the software allowed users to generate a template
based on the mean of training targets. This way the templates include variation in feature
appearance, of for instance destruction, without manually interpreting and modelling the
kind of destruction that cause it (Schneider et al., 2015). Yet it still requires the input of
somewhat homogenous features. The only method which could actually allow for the
generation of rules without extensive programming is the supervised classification section.
This allows the user to choose a select training data based on previously digitised features or
could be drawn on the image within the software. Derived from the pilot study it was clear
that this method requires more considerate choice of input layers to derive accurate
classification. Despite the fact that it provided least impressive results, it is thought to be
most promising for future applications. These algorithms can be undertaken with both object
and pixel based approaches and can be applied in various software solutions, even open
source. As this is also the foundation of active learning techniques it could be especially useful
in areas where archaeological information is sparse.
Build upon logical assumptions and insight information from remote sensing developers
various aspects where derived which should be integrated in future archaeological
applications for semi-automated pattern recognition. Some of these are already possible
with the reconsideration of eCognition and others could be initiated in an all-encompassing
web-based system. With the enforcement to compile datasets and interpretation of
archaeological features based on pre-defined conventions, solid training data could be
generated which could be exchanged amongst heritage managers and research projects to
derive more accuracy from classification projects. As well as the exchange, efficient
management of data layers and known site location could allow for change detection of
monuments as newly available remote sensing data is acquired. With the anticipation of
increasing return of new detections, locations of sites would have to be verified at a larger
scale, this could be done with the emerging field of Citizen Science. As archaeology is a field
in which many volunteers are already active in various aspects of the discipline, this may yield

69

great success. Using this improved efficient method of heritage management, archaeologists
will be able to react to hazards to the archaeology more quickly and have more time to spend
on the actual management of sites.
The techniques discussed in this dissertation have the potential to greatly improve the
way archaeological reconnaissance is conducted together with reducing costs and saving
time in other areas, and may change archaeology as we know it. Yet the presentation of an
AI model is not expected to be accepted without argument. As Barceló (2009, 95) noted that
the “intelligent” machines incite instinctive fear and anger by resembling ancestral threats –
a rival for our social position as more or less respected specialists, it is emphasised that the
eventual aim is not to replace archaeologists but to model a smart companion. The
companion would in the near future be integrated in every desk based assessment and make
it worthwhile for large scale landscape archaeology projects to verify their expected
archaeological features.

70

Bibliography

Atzberger, C., Wess, M., Doneus, M., and Verhoeven, G. 2014. ARCTIS - A MATLAB® Toolbox
for Archaeological Imaging Spectroscopy. Remote Sensing, 6, 8617-8638.
Baatz, M., Hoffmann, C., and Willhauck, G. 2008. Progressing from object-based to object-
oriented image analysis. In: Blaschke, T., Lang, S., and Hay, G. J. (eds.) Object-Based
Image Analysis: Spatial Concepts for Knowledge-Driven Remote Sensing Applications.
Verlag: Springer.
Baatz, M., and Schäpe, A. 2000. Multiresolution segmentation: An optimization approach for
high quality multi-scale image segmentation. In: Strobl, J., Blaschke, T., and
Griesebner, G. (eds.) Angewandte Geographische Informationsverarbeitung
Heidelberg: Wichmann-Verlag.
Barceló, J. A. 1995. Back-propagation algorithms to compute similarity relationships among
archaeological artefacts. In: Wilcock, J., and Lockyear, K. (eds.) CAA 1993. Computer
Applications and Quantitative Methods in Archaeology Oxford: ArcheoPress.
Barceló, J. A. 2008. Computational Intelligence in Archaeology, Hershey, New York, IGI.
Barceló, J. A. 2009. The birth and historical development of computational intelligence
applications in archaeology. Archeologia e calcolatori, 95-109.
Baxter, M. 2015. Notes on Quantitative Archaeology and R. Unpublished work available
through
https://www.academia.edu/12545743/Notes_on_Quantitative_Archaeology_and_
R.
Beck, A. R. 2007. Archaeological site detection: the importance of contrast. Annual
Conference of the Remote Sensing and Photogrammetry Society. Newcastle upon
Tyne.
Bennett, R., Cowley, D., and De Laet, V. 2014. The data explosion: tackling the taboo of
automatic feature recognition in airborne survey data. Antiquity, 88, 896-905.
Bishop, C. M. 2006. Pattern Recognition and Machine Learning, New York, Springer.
Blaschke, T. 2010. Object based image analysis for remote sensing. Isprs Journal of
Photogrammetry and Remote Sensing, 65, 2-16.
Blaschke, T., Hay, G. J., Kelly, M., Lang, S., Hofmann, P., Addink, E., Queiroz Feitosa, R., van
der Meer, F., van der Werff, H., van Coillie, F., and Tiede, D. 2014. Geographic Object-
Based Image Analysis – Towards a new paradigm. ISPRS Journal of Photogrammetry
and Remote Sensing, 87, 180-191.
Blaschke, T., Lang, S., and Hay, G. J. (eds.) 2008. Object-Based Image Analysis: Spatial
Concepts for Knowledge-Driven Remote Sensing Applications, Verlag: Springer.
Blaschke, T., and Strobl, J. 2001. What's wrong with pixels? Some recent developments
interfacing remote sensing and GIS. Geo-Informations-Systeme, 14, 12-17.
Challis, K., Carey, C., Kincey, M., and Howard, A. J. 2011. Airborne lidar intensity and
geoarchaeological prospection in river valley floors. Archaeological Prospection, 18,
1-13.

71

Cowley, D. C., De Laet, V., and Bennett, R. Auto-extraction techniques and cultural heritage
databases. In: Neubauer, W., Trinks, I., Salisbury, R., and Einwogerer, C., eds.
Proceedings of the 10th International Conference on Archaeological Prospection,
May 29–June 2 2013, 2013 Vienna. Vienna: Ludwig Boltzmann Institute.
Crawford, O. G. S. 1923. Air survey and archaeology. Geographical Journal, 342-360.
Cripps, P. J., and Tudhope, D. 2015. From interoperable to interoperating Geosemantic
resources; practical examples of producing and using Linked Geospatial Data. Keep
the Revolution going. CAA2015. Computer Applications and Quantitative Methods in
Archaeology. Siena, Italy.
Crutchley, S. 2006. Light detection and ranging (lidar) in the Witham Valley, Lincolnshire: an
assessment of new remote sensing techniques. Archaeological Prospection, 13, 251-
257.
D'Orazio, T., Da Pelo, P., Marani, R., and Guaragnella, C. 2015. Automated Extraction of
Archaeological Traces by a Modified Variance Analysis. Remote Sensing, 7, 3565-
3587.
D'Orazio, T., Palumbo, F., and Guaragnella, C. 2012. Archaeological trace extraction by a local
directional active contour approach. Pattern Recognition, 45, 3427-3438.
de Boer, A. Using pattern recognition to search LIDAR data for archeological sites. In:
Figueiredo, A., and Leite Velho, G., eds. The world is in your eyes. CAA2005.
Computer Applications and Quantitative Methods in Archaeology, 2005 Tomar,
Portugal. 95-96.
de Boer, A. G. 02/04/2015 2015. RE: Advies over master scriptie onderzoek naar
patroonherkenning in LiDAR
de Laet, V., and Lambers, K. 2009. Archaeological Prospecting Using High-Resolution Digital
Satellite Imagery: Recent Advances and Future Prospects - A Session Held at the
Computer Applications and Quantitative Methods in Archaeology (CAA) Conference,
Williamsburg, USA, March 2009. AARGnews, 39, 9-17.
de Laet, V., Paulissen, E., and Waelkens, M. 2007. Methods for the extraction of
archaeological features from very high-resolution Ikonos-2 remote sensing imagery,
Hisar (southwest Turkey). Journal of Archaeological Science, 34, 830-841.
Devereux, B. J., Amable, G. S., and Crow, P. 2008. Visualisation of LiDAR terrain models for
archaeological feature detection. Antiquity, 82, 470-479.
Doneus, M. 2013. Openness as Visualization Technique for Interpretative Mapping of
Airborne Lidar Derived Digital Terrain Models. Remote Sensing, 5, 6427-6442.
Doneus, M., and Briese, C. Full-waveform airborne laser scanning as a tool for archaeological
reconnaissance. In: Campana, S., and Forte, M., eds. From Space to Place.
Proceedings of the 2nd International Conference on Remote Sensing in Archeology.
CNR, Rome, Italy, December 2-4, 2006, 2006 Oxford. Oxford: Archaeopress, 99-106.
Doneus, M., Verhoeven, G., Atzberger, C., Wess, M., and Ruš, M. 2014. New ways to extract
archaeological information from hyperspectral pixels. Journal of Archaeological
Science, 52, 84-96.

72

Duckers, G. 2013. Bridging the ‘geospatial divide’ in archaeology: community based
interpretation of LIDAR data. Internet Archaeology, 35.
English Heritage 2011. Banjo Enclosures. Introductions to Heritage Assets. English Heritage.
English Heritage 2013. Banjo enclosure. 2013-07-17T08:43:50 ed. SENESCHAL:
http://purl.org/heritagedata/schemes/eh_tmt2/concepts/68979.
Figorito, B., and Tarantino, E. 2014. Semi-automatic detection of linear archaeological traces
from orthorectified aerial images. International Journal of Applied Earth Observation
and Geoinformation, 26, 458-463.
Gardin, J. C. 1980. Archaeological constructs: an aspect of theoretical archaeology,
Cambridge, Cambridge University Press.
Gonzalez, R. C., and Woods, R. E. 2008. Digital Image Processing, Upper Saddle River, N.J.,
Pearson Prentice Hall.
Haigh, J. G. B. 1998. Rectification of aerial images under Microsoft Windows. Archaeological
Computing Newsletter, 51, 12-20.
Hanson, W. S. 2010. The future of aerial archaeology in Europe. Photo Interprétation:
European Journal of Applied Remote Sensing, 46, 3-11.
Hanson, W. S., and Oltean, I. A. (eds.) 2013. Archaeology from Historical Aerial and Satellite
Archives, New York: Springer
Hay, G. J., and Castilla, G. 2008. Geographic Object-Based Image Analysis (GEOBIA): A new
name for a new discipline. In: Blaschke, T., Lang, S., and Hay, G. J. (eds.) Object-Based
Image Analysis: Spatial Concepts for Knowledge-Driven Remote Sensing Applications.
Verlag: Springer.
Hesse, R. 2010. LiDAR-derived Local Relief Models - a new tool for archaeological
prospection. Archaeological Prospection, 17, 67-72.
Hesse, R. 2013. The changing picture of archaeological landscapes: lidar prospection over
very large areas as part of a cultural heritage strategy. In: Opitz, R. S., and Crowley,
D. C. (eds.) Interpreting Archaeological Topography. 3D Data, Visualisation and
Observation. Oxford: Oxbow Books.
Hesse, R. 14/04/2015 2015. RE: Advise on the potential of semi-automated pattern
recognition on LiDAR derived images.
Holden, N., Horne, P., and Bewley, R. 2002. High-resolution digital airborne mapping and
archaeology. In: Bewley, R. H., and Raczkowski, W. (eds.) Aerial Archaeology –
Developing Future Practice, NATO Life Sciences, Series. In: Proceedings of the NATO
Advanced Research Workshop, Leszno, Poland, 2000. Amsterdam: IOS Press.
Holland, D. 2015. Opportunity for enlarging dissertation on semi-automated pattern
recognition.
Horne, P. D. 2009. A strategy for the National Mapping Programme, Swindon, English
Heritage.
Hough, P. V. C. 1962. Method and Means for Recognizing Complex Patterns. U.S. patent
application 3,069,654.
Hu, W., Huang, Y., Wei, L., Zhang, F., and Li, H. 2015. Deep Convolutional Neural Networks
for Hyperspectral Image Classification. Journal of Sensors, 2015, 1-12.

73

Hyyppä, J., Hyyppä, H., Yu, X., Kaartinen, H., Kukko, A., and Holopainen, M. 2009. Forest
inventory using small-footprint airborne LiDAR. Topographic laser ranging and
scanning: Principles and processing.
Isaksen, L., and May, K. 2015. 2D-Linked Data: From Interoperable to Interoperating, session
held at 'Keep the Revolution going', CAA2015, Computer Applications and
Quantitative Methods in Archaeology
http://2015.caaconference.org/program/sessions/2d/. Siena, Italy.
Kamermans, H. 2008. Smashing the Crystal Ball. A Critical Evaluation of the Dutch National
Archaeological Predictive Model (IKAW). . International Journal of Humanities and
Arts Computing, 1, 71-84.
Kokalj, Z., Zaksek, K., and Ostir, K. 2011. Application of sky-view factor for the visualisation of
historic landscape features in lidar-derived relief models. Antiquity, 85, 263-273.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional
neural networks. Advances in neural information processing systems, 2012. 1097-
1105.
Lasaponara, R., Leucci, G., Masini, N., and Persico, R. 2014. Investigating archaeological
looting using satellite images and GEORADAR: the experience in Lambayeque in
North Peru. Journal of Archaeological Science, 42, 216-230.
Lasaponara, R., and Masini, N. 2007. Detection of archaeological crop marks by using satellite
QuickBird multispectral imagery. Journal of Archaeological Science, 34, 214-221.
Lasaponara, R., and Masini, N. 2012a. Image Enhancement, Feature Extraction and
Geospatial Analysis in an Archaeological Perspective. In: Lasaponara, R., and Masini,
N. (eds.) Satellite Remote Sensing: a New Tool for Archaeology. New York: Springer.
Lasaponara, R., and Masini, N. 2012b. Pattern Recognition and Classification Using VHR Data
for Archaeological Research. In: Lasaponara, R., and Masini, N. (eds.) Satellite Remote
Sensing: a New Tool for Archaeology. New York: Springer.
Lasaponara, R., and Masini, N. 2012c. Remote Sensing in Archaeology: From Visual Data
Interpretation to Digital Data Manipulation. In: Lasaponara, R., and Masini, N. (eds.)
Satellite Remote Sensing: a New Tool for Archaeology. New York: Springer.
Lemmens, J. P. M. M., Stančič, Z., and Verwaal, R. G. Automated archaeological feature
extraction from digital aerial photographs. In: Andresen, J., Madsen, T., and Scollar,
I., eds. Computing the Past. CAA 1992. Computer Applications and Quantitative
Methods in Archaeology, 1993. Aarhus, Denmark: Aarhus University Press, 45-52.
Lock, G. R. 2003. Using computers in archaeology: towards virtual pasts, Psychology Press.
Lorenzo, A. 10/06/2015 2015. RE: Enquiry about eCognition licenses.
Luo, L., Wang, X. Y., Guo, H. D., Liu, C. S., Liu, J., Li, L., Du, X. C., and Qian, G. Q. 2014.
Automated Extraction of the Archaeological Tops of Qanat Shafts from VHR Imagery
in Google Earth. Remote Sensing, 6, 11956-11976.
Niemeyer, I., Marpu, P. R., and Nussbaum, S. 2008. Change detection using object features.
In: Blaschke, T., Lang, S., and Hay, G. J. (eds.) Object-Based Image Analysis: Spatial
Concepts for Knowledge-Driven Remote Sensing Applications. Verlag: Springer.

74

Palmer, R. 2000. View From Above: Can Computers Help Aerial Survey? In: Lock, G. R., and
Brown, K. (eds.) On the theory and practice of archaeological computing. Oxford:
Oxford University Committee for Archaeology.
Palmer, R. 2005. If they used their own photographs they wouldn’t take them like that. In:
Brophy, K., and Cowley, D. C. (eds.) From the air: Understanding aerial archaeology.
Stroud: Tempus.
Palmer, R. 2009. Editorial. AARGnews, 39, 2-4.
Parcak, S. 2009. Satellite Remote Sensing for Archaeology, New York, Taylor & Francis.
Perry, B. T. 1969. Iron Age enclosures and settlements on the Hampshire chalklands.
Archaeological Journal, 126, 29-43.
Redfern, S. 1997a. Computer assisted classification from aerial photographs. AARGnews, 14,
33-38.
Redfern, S. 1999. A PC-based system for computer assisted archaeological interpretation of
aerial photographs. In: Dingwall, L., Exon, S., Gaffney, V., Laflin, S., and van Leusen,
M., eds. Archaeology in the age of the internet. CAA 1997. Computer Applications
and Quantative Methods in Archaeology Conference, 1997b Birmingham, United
Kingdom. Oxford: Archaeopress, 47.
Redfern, S., Lyons, G., and Redfern, R. M. The Automatic Morphological Description and
Classification of Archaeological Monuments from Vertical Aerial Photographs.
OESI/IMVIP Joint Conference, 1998 Maynooth, Ireland.
Riley, M. A. 2009. Automated Detection of Prehistoric Conical Burial Mounds: From LIDAR
Bare-Earth Digital Elevation Models. Northwest Missouri State University.
Risbøl, O., Briese, C., Doneus, M., and Nesbakken, A. 2015. Monitoring cultural heritage by
comparing DEMs derived from historical aerial photographs and airborne laser
scanning. Journal of Cultural Heritage, 16, 202-209.
Rowlands, A., and Sarris, A. 2007. Detection of exposed and subsurface archaeological
remains using multi-sensor remote sensing. Journal of Archaeological Science, 34,
795-803.
Schneider, A., Takla, M., Nicolay, A., Raab, A., and Raab, T. 2015. A Template-matching
Approach Combining Morphometric Variables for Automated Mapping of Charcoal
Kiln Sites. Archaeological Prospection, 22, 45-62.
Sevara, C., and Pregesbauer, M. 2014. Archaeological feature classification: An object
oriented approach. South-Eastern European Journal of Earth Observation and
Geomatics. Special Issue: GEOBIA, 3, 139-143.
Stonehenge World Heritage Site. 2015. The Different Types of Round Barrows [Online].
Available: http://89.206.186.140/stonehengemap/sites/barrows/all.html [Accessed
01/07/2015.
Stumpf, A., and Kerle, N. 2011. Object-oriented mapping of landslides using Random Forests.
Remote Sensing of Environment, 115, 2564-2577.
Stumpf, A., Lachiche, N., Kerle, N., Malet, J.-P., and Puissant, A. Adaptive spatial sampling
with active random forest for object-oriented landslide mapping. Geoscience and
Remote Sensing Symposium (IGARSS), 2012 IEEE International, 2012. IEEE, 87-90.

75

Stumpf, A., Lachiche, N., Malet, J. P., Kerle, N., and Puissant, A. 2014. Active Learning in the
Spatial Domain for Remote Sensing Image Classification. IEEE Transactions on
Geoscience and Remote Sensing, 52, 2492-2507.
Stutt, A., and Shennan, S. 1990. The nature of archaeological arguments. Antiquity, 64, 766-
777.
Traviglia, A. 2007. MIVIS hyperspectral sensors for the detection and GIS supported
interpretation of subsoil archaeological sites. In: Clark, J. T., and Hagemeister, E. M.,
eds. Digital Discovery: Exploring New Frontiers in Human Heritage. CAA2006.
Computer Applications and Quantative Methods in Archaeology Conference, 2006
Fargo, United States. Budapest: Archaeolingua.
Trier, Ø. D. 20/05/2015 2015. RE: Further information of barrow detection on LiDAR data.
Trier, Ø. D., Larsen, S. Ø., and Solberg, R. 2009a. Automatic detection of circular structures in
high-resolution satellite images of agricultural land. Archaeological Prospection, 16,
1-15.
Trier, Ø. D., Larsen, S. Ø., and Solberg, R. 2009b. Finding burial mounds from space: automatic
detection of circular soil marks and crop marks in QuickBird imagery of agricultural
land in south-east Norway. AARGnews, 39, 18-24.
Trier, Ø. D., and Pilo, L. H. 2012. Automatic Detection of Pit Structures in Airborne Laser
Scanning Data. Archaeological Prospection, 19, 103-121.
Trier, Ø. D., Zortea, M., and Tonning, C. 2015. Automatic detection of mound structures in
airborne laser scanning data. Journal of Archaeological Science: Reports, 2, 69-79.
van den Eeckhaut, M., Kerle, N., Poesen, J., and Hervás, J. 2012. Identification of vegetated
landslides using only a Lidar-based terrain model and derivatives in an object-
oriented environment. Proceedings of the 4th GEOBIA, 211.
Verhagen, P., and Drăguţ, L. 2012. Object-based landform delineation and classification from
DEMs for archaeological predictive mapping. Journal of Archaeological Science, 39,
698-703.
Verhoeven, G. J., and Doneus, M. 2011. Balancing on the Borderline - a Low-cost Approach
to Visualize the Red-edge Shift for the Benefit of Aerial Archaeology. Archaeological
Prospection, 18, 267-278.
Verhoeven, G. J., Doneus, M., Atzberger, C., Wess, M., Rus, M., Pregesbauer, M., and Briese,
C. 2013. New approaches for archaeological feature extraction of airborne imaging
spectroscopy data. Archaeological Prospection. Proceedings of the 10th International
Conference on Archaeological Prospection. Vienna, Austria. Vienna: Austrian
Academy of Sciences.
Wheatley, D., and Gillings, M. 2002. Spatial Technology and Archaeology: the Archaeological
Applications of GIS, London, New York Taylor & Francis.
Wiltshire Museum. 2015. Document Library [Online]. Available:
http://www.wiltshiremuseum.org.uk/documents/?LibraryID=39 [Accessed
04/09/2015.
Zingman, I., Saupe, D., and Lambers, K. 2015. Detection of incomplete rectangular contours
with application in archaeology. Technical Report, University of Konstanz.

76

Appendix 1: Snapshots of the workflow for cognitive argumentation


Figure 1 Result of the image segmenaion into objects. Colours oulining the objects display the range of
brightness with highest values in green (location: Overton Hill).


Figure 2 Images showing the effect of different brightness thresholds (left) 10+ (right) 20+ (location: Overton Hill).


Figure 3 (left) After the first classification based on mean brightness, classified segments were merged resulting
in very large and small classifications of false positives. These were later unclassified based on thresholds for ‘Area’
and ‘Length/Width’ (right)The selected image object including two barrows (merged during brightness
classification) does not fit the area threshold in which objects with total of pixels higher than 3000 were
unclassified. Finer object segmentation at the start of the process might overcome this (location: Overton Hill).



Figure 4 left below is the end of the process with most barrows classified and only one false positive


Appendix 2: Snapshots of the workflow for template matching


Figure 1 Barrow sample selection on aspect layer (image location: Overton hill).


Figure 2 Iterative process of defining the right threshold for the chosen samples54 on aspect layer (image
location: Overton hill).

Figure 3 Generated templates for aspect layer with sample correlation of 0.653 based on 22 samples (image
location: Windmill hill).



Figure 4 Barrow sample selection at Overton hill on slope layer (image location: Overton hill).

Figure 5 Iterative process of defining the right threshold for the chosen samples on the slope layer (image
location: Windmill hill).


Figure 6 template correlation of 0.755 based on 22 samples (image location: Overton hill).



Figure 7 long barrow detected at 0.4 threshold value. Eventually classified as false positive.


Figure 8 (left) Side by side view of the slope and correlation map with indications of known barrows and
detected barrows (right) hoovering over the image indicates the correlation which detected a barrow with 0.68
correlation.


Appendix 3: Snapshots of the workflow for random forest classifier


Figure 1 Importing barrows and classifying them accordingly.


Figure 2 First trial of applying the RF classifier on all round barrows.



Figure 4 Classification of barrows according to their types.


Figure 5 Expansion of the first analysis with training based of types of barrows. Still a lot of false positives.
Classification was based on mean Slope and Aspect.



Figure 5 These error messages cryptically showed that the analysis could not be undertaken if there are areas
with noData values (for template matching the whole program crashed).


Figure 6 As the error was resolved the analysis was undertaken on the whole dataset using the mean-brightness,
area and length/width which caused almost all segments to be classified as barrows (bottom right). As the
mean-brightness (upper right) was not merged like it was in the cognitive workflow, the length/width (upper
left) and area (bottom left) didn’t mean anything on their self. Therefore the context of the segments should be
included.

You might also like