Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Contents lists available at ScienceDirect

Remote Sensing Applications: Society and Environment


journal homepage: www.elsevier.com/locate/rsase

Spatial resolution and landscape structure along an urban-rural gradient: Do T


they relate to remote sensing classification accuracy? – A case study in the
megacity of Bengaluru, India
Kwame T. Awuaha,1, Nils Nölkea, , Maximilian Freudenbergb, B.N. Diwakarac, V.P. Tewaric,d,

Christoph Kleinna
a
University of Goettingen, Forest Inventory and Remote Sensing, Faculty of Forest Sciences and Forest Ecology, Büsgenweg 5, D-37077 Göttingen, Germany
b
University of Goettingen, Third Institute of Physics, Friedrich-Hund-Platz 1, 37077 Göttingen, Germany
c
Institute of Wood Science and Technology, Bengaluru, India
d
Himalaya Forest Research Institute HFRI, Shimla, India

ARTICLE INFO ABSTRACT

Keywords: Along the urban-rural gradient in megacities, the extent and material composition of impervious surfaces are dif-
Map accuracy ferent. This leads to variations in the frequently mentioned heat-island property, but possibly also to different spectral
Land-cover signatures and, consequently, different accuracies in remote sensing image classification. This, in turn, creates a
Spatial metrics challenge when it comes to selecting suitable image processing techniques. In this study, we examine how the
Urban-rural gradient
accuracy of land-cover classification changes along an urban-rural gradient as a function of spatial resolution and the
Landscape structure
gradient in landscape structure. RapidEye, Sentinel-2A and Landsat 8 images were used. Land-cover classification
Megacities
was performed using a deep learning model and landscape metrics were used to assess landscape structure. A high
degree of landscape heterogeneity and lowest classification accuracy was observed in the transition zone between
urban and rural domains, within a stretch of 15–20 km from the urban center. As expected, spatial resolution was
found to be influential in classification accuracy. A comparison of classifications indicates that within rural land-
scapes finer resolution images retain more spatial and thematic details in land-cover, e.g., RapidEye and Sentinel-2A
imagery better distinguish built-up areas within the agricultural landscape and discriminate more of the mapped
land-cover/use classes than Landsat 8. Overall accuracy increased with increasing spatial resolution (30 m, 10 m,
5 m) within the urban and rural areas, however, the 10 m resolution image (Sentinel-2A) produced better results in
the transition zone. The findings from this study provide a basis for more focused, consistent and possibly more
accurate time-series analyses of land-use dynamics at the urban-rural interface.

1. Introduction the quality of selected land cover datasets have considerable influence on
simulation and model predictions (Benitez et al., 2004; Sertel et al., 2010).
Assessment of accuracy of land-cover maps produced from remote It is well known that classification accuracy, when looking at the same
sensing imagery is an integral part of image classification processes. In the set of classes, depends on various factors, including landscape hetero-
recent past, information on map accuracy has received heightened atten- geneity and spatial resolution of sensors, which have been reported to be
tion, particularly in the context of the increasing consensus on under- particularly important (Smith et al., 2002, 2003; Wu et al., 2008; Lu et al.,
standing and monitoring global environmental changes and the co- 2010; Fritz et al., 2011). Land cover and its spectral and spatial hetero-
ordination of mitigation and adaptation actions against climate change geneity are considerably different in rural and urban landscapes. For in-
(Mora et al., 2014). Spatially explicit land cover information serves as stance, while rural landscapes are mainly characterized by larger areas of
input into ecosystem, regional and global climate simulation, global cir- agricultural fields, barren land, fallow land, scattered human settlements
culation, land surface (Mora et al., 2014; Fritz et al., 2011) and socio- and forest patches of varying sizes, urban landscapes are predominantly
economic models (Nesbitt and Meitner, 2016; Schüle et al., 2017). Thus, composed of impervious surfaces such as roads, buildings, rooftops and
consistent and quality-checked land cover data are in high demand since parking lots mixed with tree crowns and green areas. The high variation in


Corresponding author.
E-mail address: nnoelke@gwdg.de (N. Nölke).
1
Current address: Department of Geography, Edge Hill University, Ormskirk L39 4QP, UK

https://doi.org/10.1016/j.rsase.2018.10.003
Received 26 July 2018; Received in revised form 15 October 2018; Accepted 17 October 2018
Available online 23 October 2018
2352-9385/ © 2018 Elsevier B.V. All rights reserved.
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

spectral signatures of built-up surfaces, coming from varying sizes, shapes (Fig. 1), which has been found to be an appropriate extent for landscape
and compositions of the used surface materials renders urban landscapes assessment using landscape metrics in other contexts (Messerli et al.,
spectrally altogether more heterogeneous (Lu et al., 2010) which may lead 2009). The extents of urban, transition and rural domains were defined
to more misclassifications and, ultimately, to a reduced classification ac- along the proportion of built-up pixels, using a grid of 14 squares in a
curacy (Tran et al., 2014). Landsat images, which are probably the most north-south direction, each of size 3 km2, and a cross-classification ap-
widely used data source for larger area land cover and land use classifi- proach. Areas with more than 50% built-up pixels were regarded as
cation – in particular when it comes to time series -, often fall short of urban whereas areas with 10–50% and 0–10% built-up pixels were de-
accuracy requirements particularly in complex urban-rural landscapes, fined as transition and rural respectively (Angel et al., 2010).
owing to its relatively coarse spatial resolution (Lu and Weng, 2005). Al-
ternatively, high spatial resolution images like IKONOS, Quickbird and 2.2. Satellite imagery
WorldView have been promoted for land-cover classification (Sugumaran
et al., 2002; Van der Sande et al., 2003; Lu et al., 2010) and mapping of Remote sensing data used in this study included Landsat 8, Sentinel-
impervious surfaces in urban environments (Goetz et al., 2003; Lu and 2, RapidEye and WorldView-3 datasets, all cloud free and covering the
Weng, 2009). Although this has been advantageous in minimizing mixed same extent. The Landsat 8, Sentinel-2A and RapidEye imagery served
pixel problems (Lu and Weng, 2009) with more detailed information ex- as the input for the land cover classifications whereas the high-resolu-
traction potential, such high resolution data are constrained by high tion WorldView-3 image was used as a reference to extract training as
spectral variation within the same land cover class (Lu et al., 2010; Lu and well as validation data. All satellite images were processed to surface
Weng, 2009) and high processing requirements (Lu and Weng, 2007). reflectance; image-to-image co-registration was done by automated
In the surroundings of rapidly emerging megacities, rapid changes in designation of tie-points. Table 1 shows a summary of the sources and
land-use and land-cover is being observed in both the urban and the properties of the image datasets.
surrounding rural domain. There, up-to-date and quality-checked data-
sets are of high importance for analyses of land cover dynamics.
Producing these required high-quality information from free available 2.3. Classification scheme
satellite images like Landsat and Sentinel-2 with 30 m and 10 m resolu-
tion respectively is challenging. Extents, materials and composition of In the study area, a variation of different landuse is found, ranging
impervious surfaces change significantly in such diverse landscapes, from agricultural to industrial uses. Consequently, the area is char-
leading to different spectral responses. This creates challenges in acterized by a broad spectrum of anthropogenic (e.g. concrete, asphalt,
choosing suitable classification algorithms and image processing tech- metals, roofing material, etc.), semi-natural (e.g. grass, fields, re-
niques (Lu et al., 2010). The use of object-oriented approaches (MacLean servoirs) and natural (e.g. lakes, trees, bare soil) land-cover features. A
and Congalton, 2012) and the incorporation of texture and spectral classification scheme for the existing land-cover categories was thus
bands (Lu et al., 2010) have been explored to enhance image classifi- developed. It included seven land use/cover classes and is detailed in
cation accuracy in similar contexts. Those studies, however, are com- Table 2 together with some information about the quantitative basis for
monly restricted to urban areas only, and little is known about the per- classification (number and area of training polygons).
formance of classification processes when simultaneously considering
rural and urban landscapes together with the transition zone in between. 2.4. Training data
This study aims at enhancing the understanding of how classifica-
tion accuracy changes along the urban-rural gradient in one of the A systematic sampling approach was adopted for the collection of
world's most dynamic cities, Bengaluru, India. With this background, training data. Circular response units of 500 m radius were installed at 78
we investigate how variations in accuracy are related to landscape sampling points (corresponding to a grid of 1.89 km by 1.89 km) in the
structure and which role spatial resolution plays in this context. satellite imagery. In this radius observations were made for training of
the classifier. The number of 78 training plots was deemed sufficient to
2. Material and methods cover as much spectral and spatial information as possible along the
gradient, given the apparent variety in land use/cover types. Due to the
2.1. Study area systematic sample, the resulting numbers of training areas per class were
roughly proportional to their respective total area so that a post-stratified
Bengaluru, the capital of the Indian State of Karnataka, is located sample resulted. Training data for the defined land use/cover classes
around 12°58'N, 77°35'E and lies on Southern India's Deccan plateau, at (Table 2) was collected by visually identifying, digitizing and labeling
about 920 m above MSL (Sudhira and Nagendra, 2013). The topography 809 polygons of pixels using the WorldView-3 image as reference such
is relatively flat at Bengaluru North, which is one of four urban districts that their spatial and spectral properties were comprehensively captured.
while Bengaluru South is slightly undulating, with a central ridge run- The same polygon extents and locations were then used to extract
ning in north-east and south-west direction. The city is nowadays very training samples at the different spatial resolutions. To allow for a direct
well-known for Information Technology, Biotechnology, Aerospace and comparability of the results, training samples for the 5 m RapidEye image
other advanced knowledge-based industries and research centers from November 2017 were adjusted to account for temporal differences
(Hiremath et al., 2013) while in earlier times it was known as the ‘garden in land cover, particularly in agricultural landscapes and dried-out water
city’ of India owing to its widespread parks and green spaces. bodies relative to the other image sets which stem from November 2016.
Bengaluru is a rapidly growing megapolis with concomitant in- All efforts were undertaken to ensure spectral separability of the
creases in population (e.g. from 6537,124 in 2001 to 9621,551 in 2011) training classes. Specifically, spectral plots of training classes were re-
(India, 2011), economic activity and urban sprawl. This extensive viewed and repeatedly modified until training sets for all classes had
growth has resulted in landscape fragmentation and drastic changes in attained adequate spectral separability. Further, the Jefferies-Matusita
land use along the urban-rural interface (Sudhira and Nagendra, 2013). (JM) distances for pairs of classes of training sets were computed to
A 50 × 5 km research transect, in the northern part of Bengaluru, was provide a statistical measure of training class separability. The JM
defined in the framework of a larger Indian-German collaborative re- distance is a commonly used measure for defining spectral (dis) simi-
search project. This transect contains different land-use categories and larity of thematic classes (Van Niel et al., 2005; Dabboor et al., 2014;
extends over rural, transition and urban domains. To analyze relation- Momeni et al., 2016). The JM distance asymptotically approaches 2
ships between classification accuracy, land-cover spatial heterogeneity when pairs of class spectral signatures are completely different and falls
and spatial resolution, the transect was divided into 5 × 5 km subsets to 0 when spectral signatures are identical (Richards and Jia, 2006).

90
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Fig. 1. Map of study area centered on Bengaluru District – India with the 250 km2 transect and five-kilometer vertical divisions. Urban center is assumed to be the
lower border of the transect.

Table 1
List of satellite image data used in this study.

Satellite data Acquisition date Resolution Source

Landsat 8 November 2016 30 m Visible, Near Infrared (NIR), Short-Wave Infrared (SWIR), 100 m Thermal Infrared (TIR), 15 m United States Geological Survey (USGS)
panchromatic
Sentinel-2A November 2016 10 m, 20 m and 60 m in Visible and Near Infrared (NIR) to Short-Wave Infrared (SWIR) spectral range European Space Agency (ESA)
RapidEye November 2017 5 m Visible, Red Edge, Near Infrared (NIR) Planet Labs
Worldview-3 November 2016 31 cm panchromatic, 1.24 m multispectral, Digital Globe Inc.

2.5. Multilayer Perceptron classification different classes are not linearly separable are the main reason for the
usage of a neural network whose advantage is, among others, that it can
The pixel-wise image classification was performed for the whole handle such complexity (Bischof et al., 1992; Venkatesh and Raja,
extent of the transect using a so-called Multilayer Perceptron (MLP) 2003). Fig. 2 shows the structure of our MLP network model.
network. It is a simple deep learning model which is trained via back- For network training, the categorical cross-entropy loss function
propagation (Goodfellow et al., 2016). The huge variations in spectral was used in conjunction with the Adam optimizer (Kingma and Ba,
reflectance along the urban-rural interface and the fact that the 2014) and a constant learning rate of 0.001. Each layer is followed by a

Table 2
Land cover classification scheme.

Land cover class Training polygons

ID Name Description Number Area (km2 )

1 Built-up Building structures with walls and roof, including residential buildings, industrial buildings, long surfaced vehicle tracks, mostly 391 2.09
asphalt coated, also including narrow streets and all railroad systems
2 Trees Patches of leaf-on trees within the landscape 175 0.86
3 Grass/Shrub Areas such as urban parks and lawns but also grass-covered areas around agricultural fields 13 0.08
4 Water Bodies Water bodies, including natural lakes, man-made ponds and reservoirs 23 1.67
5 Fallow land Agricultural lands currently without crops 90 3.98
6 Crop field Agricultural lands currently covered with crops 87 1.72
7 Barren Open areas covered by bare soil outside agricultural land use contexts as well as rocky surfaces 30 0.53

91
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

landscape structure by number, size and shape of patches in the different


land cover classes. Taking into account the spatial-contextual information
through landscape metrics (configuration) allows a more thorough inter-
pretation of classification accuracy (Smith et al., 2003). As an example,
high fragmentation can result in an increase in the number of mixed pixels
and may affect the classification error. Including this information may
provide additional insights into potential causes of classification error.
Four landscape metrics were selected and calculated for each of the
5 km × 5 km subsets using Fragstats, a spatial analysis program for raster
maps (McGarigal and Marks, 1995). The selected landscape metrics were
Patch Density (PD), Landscape Shape Index (LSI), Contagion Index
(CONTAG) and Shannon's Diversity Index (SHDI) (Table 3). This combi-
nation of metrics has proven capable of providing robust quantitative
Fig. 2. Illustration of the MLP network. measures of spatial landscape structure (Herold, 2001).

dropout layer (Srivastava et al., 2014), randomly neglecting 20% of the 3. Results
weights in order to minimize overfitting.
Neural networks usually learn fastest, when their input is normal- 3.1. Land cover map
ized and their training data are shuffled (LeCun et al., 1998). Therefore,
each image is normalized by subtracting the mean and dividing by its The satellite images were classified using the Multilayer Perceptron
standard deviation. As pixel-wise classification only takes spectral in- Neural Network classifier to produce land-cover maps at the three
formation into account, the network is fed with pixels in a random different spatial resolutions of 30 m for Landsat 8, 10 m for Sentinel-2
order. To monitor accuracy and thereby to prevent overfitting, each and 5 m for RapidEye. Fig. 3 shows land cover maps for sections of the
dataset was divided into a training and a test set with a ratio of 70–30. five-kilometer vertical divisions of the study strip, representing urban
When the network learns a meaningful representation of the data the (10–15 km), transition (15–20 km) and rural (45–50 km) landscapes,
accuracies on both, training and test set, increase simultaneously. and allowing for an visual examination of differences. Differences in
scale and date of image acquisition (in the case of RapidEye) may result
2.6. Accuracy assessment considerable in disparities in derived land cover at different spatial
resolutions particularly within the transition and rural domains. For
The accuracy assessment was done separately for each land cover map instance, additional locations of water bodies in RapidEye relative to
(whole transect and in 5 km steps) at the different spatial resolutions of Landsat 8 and Sentinel-2A result from temporal differences in image
30 m, 10 m and 5 m. An independent validation dataset was collected from acquisition. Overall, the output maps from finer resolution, as expected,
the very high resolution WorldView-3 image. A systematic sample of gave a better representation of land cover. Eventually, this impression is
26047 points was labeled with their land cover class in conformity with consistent with the pattern of differences in the measures of classifi-
the classification scheme. Each point was visually interpreted. Similar to cation accuracy given in Section 3.2.
the training data, the validation dataset was adjusted to account for The finer resolution images retain more spatial and thematic details
temporal differences in land cover between 2016 and 2017. Error matrices in land-cover, particularly in the rural end of the landscape. For ex-
were constructed that showed the correspondence between reference data ample, RapidEye and Sentinel-2A better distinguish built-up areas
and classification output per land cover class. Users' accuracy, producers' within the agricultural landscape as shown in Fig. 3 and discriminate
accuracy and overall map accuracy was estimated from the error matrix. more land cover/use classes (see appendices A.5, B.6 and C.7) than
Following Olofsson et al. (2014) accuracy-adjusted area estimates for each Landsat 8. However, the level of thematic details is similar for all spatial
class and confidence intervals were estimated both for the whole transect resolutions in the urban and transition zones (appendices A.5, B.6 and
and for each of the 5 km × 5 km subsets. C.7). Furthermore, area estimates for key land cover types were similar
at the different spatial resolutions (see Fig. 5).
2.7. Assessment of landscape structure
3.2. Classification accuracy
Spatial metrics (landscape metrics) provide information about the
composition and configuration of land-cover data sets, describing Confusion matrices for the land-cover maps were computed.

Table 3
Selected landscape metrics. N is the number of patches in the landscape, A is the total landscape area, E is the total edge length and minE is the edge length of the
square patch with equal area, P i is landscape proportion occupied by patch type i, m is number of patches in the landscape and g ik represents the number of
adjacencies (shared sides) between pixels of the patch types i and k.

Landscape Metric Description Value range

PD =
N
(10000)·(100) Mean number of patches per 100 ha. In this context, patch PD > 1, increases without limit. The larger, the more
A refers to land cover/use pixel or pixel groups. Quantifies fragmented
configuration
LSI =
E Expresses the configuration of a landscape. Quantifies LSI 1, increases without limit. The larger, the more
minE
configuration fragmented
gik gik Indicates the overall clumpiness within a landscape. 0 < CONTAG 100, 0 maximally disaggregated and
m (P )
ik i m g · ln (Pi ) m g Quantifies configuration 100 aggregated. The larger, the less fragmented
k = 1 ik k = 1 ik
CONTAG = 1 + ·100
2 ln(m )

SHDI =
m
(Pi·lnPi) Quantifies the number of different patch types (i.e. patch SHDI 0, increases without limit. The larger, the
i=1
richness). Quantifies composition more diverse and likely to be more fragmented

92
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Fig. 3. Example land-cover maps for different spatial resolution images (at 10–15 km, 15–20 km and 45–50 km along the transects represent regions within the
urban, transition and rural areas respectively).

Fig. 4 shows the trend in land cover/use classification accuracy from 3.3. Spatial heterogeneity
the urban towards the rural area which is similar for all three types of
imagery used. The overall classification accuracy decreased sharply Fig. 6 shows the results of spatial composition and configuration of
from the urban region, reaching its lowest in the transition zone land cover types as a function of the distance from urban to rural en-
(15–20 km stretch away from the defined urban center). Thereafter, vironments. The computed landscape metrics point to a trend in spatial
accuracy estimates increased gradually in the rural stretch (see Ap- heterogeneity with distance from urban to rural areas where this trend
pendix A.5, B.6 and C.7 for details). Table 4 shows a summary of overall is similar for the three spatial resolutions, for all four selected frag-
classification accuracies for the entire urban, transition and rural zones. mentation indices.
Overall accuracy generally increased with increasing spatial resolution Regarding spatial heterogeneity, the highest values were observed
(30 m < 10 m < 5 m) within the urban and rural areas; however, the in the transition zone within 15–20 km from the urban center for all
10 m resolution image (Sentinel-2A) gave better results in the transition metrics. The results indicate that PD and LSI are significantly influenced
zone (also see Fig. 4). by spatial resolution, unlike CONTAG and SHDI.
Detailed results for the whole transect are presented in appendix
A.5, B.6 and C.7 for the three spatial resolutions. As with overall ac- 4. Discussion and conclusions
curacy, also the class-wise accuracies increased in principle with finer
spatial resolution (5 m RapidEye and 10 m, Sentinel-2). Nonetheless, 4.1. Classification accuracy changes with landscape structure along the
the coarser resolution (30 m, Landsat 8) proved slightly superior (i.e. urban-rural gradient
8% over 5 m RapidEye and 3% over 10 m Sentinel-2A) with regards to
areas correctly classified as Grassland for user's accuracy (see Appendix Landscapes with higher values of PD, LSI and SHDI, as well as lower
A.5, B.6 and C.7). Overall classification accuracy went down with CONTAG can generally be considered as spatially more heterogeneous
spatial resolution by approximately 2% from 5 m (RapidEye) to 10 m (Smith et al., 2003; Lechner et al., 2009; Tran et al., 2014). Landscape
(Sentinel-2A) and 6% from 10 m to 30 m (Landsat 8) spatial resolution. structure along the studied urban-rural gradient in Bengaluru follows a
‘U’ (for CONTAG) or reversed ‘U’ (for PD, LSI and SHDI) relationship

93
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

between spatial landscape heterogeneity and urbanization gradient


(Luck and Wu, 2002; Yu et al., 2016). It should be noted that, while
results of the selected spatial metrics are undeniably correlated. The
depiction of an increasing trend in spatial complexity from the urban
center through the transition zone and subsequently decreasing and
leveling-off within the rural zone, their magnitude may be affected by
the scale of analysis (Wu et al., 2000; Luck and Wu, 2002) particularly
for PD and LSI as shown in Fig. 6.
The image classification accuracy results exhibited a near sinusoidal
pattern of changes from urban to rural landscapes. This trend is similar for
the three spatial resolutions (5 m RapidEye, 10 m Sentinel-2A and 30 m
Landsat 8). Generally, the highest and lowest accuracies were recorded in
the dense urban region and the transition zone respectively. Results of the
assessment of spatial heterogeneity showed that the transition zone along
the gradient has a relatively high number of isolated land cover/use patches
and longer edge lengths, and is less clumpy and more diverse with regards
to the land cover/use types present. This peculiar feature of the transition
landscape relates with the findings of Solecka et al. (2017) that transition
regions exhibit rapid urban sprawl leading to conversion of farmlands and
other vegetated areas into discontinuous urban fabric. Therefore, the decline
in classification accuracy from the urban center through the transition zone
correlates with the concomitant increase in landscape spatial heterogeneity
within the same stretch of the transect. Similarly, Chen et al. (2016) studied
Fig. 4. Changes in overall accuracy along the urban-rural gradient for different the impact of spatial heterogeneity on crop area classification accuracy and
resolutions. Dashed horizontal lines give the whole transect overall accuracy reported a decreasing trend in overall accuracy with increase in composi-
values at each resolution. tional and configurational heterogeneity. On the other hand, the results of
PD, LSI and CONTAG metrics showed the rural and urban ends as having
Table 4 more contiguous and less fragmented patches with many linked cell ad-
Overall accuracies for different image resolutions within urban, transition and jacencies (McGarigal et al., 2002; Lechner et al., 2009) and lower compo-
rural areas along the gradient. Confidence interval estimates were derived sitional heterogeneity based on SHDI, which can be attributed to the pre-
following Olofsson et al. (2014). dominance of agricultural (crop fields and fallow land) and built-up pixels
respectively. Nevertheless overall accuracy was better within the urban
Image Region
region although slightly more spatially heterogeneous with expected higher
Urban Transition Rural spectral heterogeneity (Lu et al., 2010). In part, this may be explained by
spectral confusion between barren, fallow land, cropfields and grassland
RapidEye 0.77 ± 0.007 0.62 ± 0.015 0.78 ± 0.007 classes given that the rural end is primarily composed of agricultural land
Sentinel-2A 0.75 ± 0.007 0.67 ± 0.013 0.76 ± 0.006
uses which are further distinguished as crop fields and fallow land classes.
Landsat 8 0.64 ± 0.006 0.61 ± 0.014 0.70 ± 0.008
The observed trend demonstrates that the changes in classification accuracy
are sensitive to spatial composition and configuration, and point to the fact
that an spatial heterogeneity that is associated with different landscape
conditions exerts enormous influence on image classification accuracy
(Congalton, 1991; Smith et al., 2002, 2003; Congalton and Green, 2008).

4.2. Classification accuracy changes with spatial resolution along the urban-
rural gradient

Differences in accuracy along the urban-rural gradient were similar


for the three spatial resolutions. Yet the results show decreasing dif-
ferences in overall accuracy from the urban end towards the rural end
of the transect, with better and comparable results recorded for the 5 m
RapidEye and 10 m Sentinel-2A image. This may be related to the fact
that there are less mixed pixels for higher spatial resolution, which is
particularly relevant with increasing compositional and configurational
heterogeneity (Lim et al., 2009; Chen et al., 2016). This finding cor-
responds to earlier observations that the use of fine spatial resolution
images in complex regions further enhances classification accuracy (Lu
et al., 2010). Additionally, fine spatial resolution images enabled better
discrimination of land cover/use classes. Accuracy estimates of built-up
surfaces decreased with decreasing spatial resolution (5 m > 10 m >
30 m) along the urban-rural gradient. This phenomenon may pose
considerable challenges when using Landsat images as data for time
series analyses and change detection of urban sprawl (however, the
long historical records of Landsat archive provide great options in
Fig. 5. Area estimates of the classes Built-up and Tree cover along the urban- general for time series analysis and detection of changes in land cover).
rural gradient. Additionally, the observed differences in mapping built-up areas also
point to the critical role of spatial resolution in the calculation of

94
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Fig. 6. Metrics of spatial composition and configuration as a function of distance from urban center, at 5 m, 10 m and 30 m spatial resolutions (Shaded area represents
95% confidence interval (CI) from class-level metrics).

indexes that quantify urbanization (Wei et al., 2017). resolutions even within the same landscape (Hsieh and Lee, 2000;
Other than the observed changes in classification accuracy, similar Huiping et al., 2003; Chen et al., 2004; Alsheri, 2010). Chen et al.
area estimates particularly for built-up and tree cover were recorded (2004) recommends finer resolution images for more fragmented
from the different spatial resolution images along the urban-rural gra- landscapes, while Huiping et al. (2003) iterates that different classes
dient. Area coverage for built-up surfaces decreased rapidly in tandem require different image resolutions for mapping and that the choice of
with tree cover along the urban-rural gradient for all spatial resolutions. the resolution should be guided by the most important class.
Sudhira and Nagendra (2013) reported the dominance of large canopy Ultimately, the results of this study demonstrate a considerable
shade trees along streets and in parks within the urban district of variation in image classification accuracy from urban to rural areas
Bengaluru which we observed in our ground surveys. The higher cov- where accuracies are generally higher in the urban and rural zones than
erage of trees in the urban zone is characteristic for Bengaluru and in the transition zone, and this reflects the degree of landscape spatial
deviates from popular assumptions about the urban-rural gradient heterogeneity along the urban-rural gradient. Thus remote sensing
(URG) hypothesis of ecosystem processes, that the proportion of parks analyses of cities and surrounding areas should take into consideration
and green spaces increases towards the outskirts of cities as evidence of such possible differences and the role of compositional and configura-
a decreasing degree of urbanization (Malkinson et al., 2018). tional landscape structure as well as the scale of analyses in outcomes of
investigated processes particularly along such complex urban-rural
4.3. Spatial resolution and classification accuracy for the whole transect gradient.

Classification accuracy for the whole transect was generally en-


Declarations of interest
hanced at higher spatial resolutions, which highlights the role of fine
spatial image resolution in enhancing classification accuracy (Ponzoni
None.
et al., 2002; Ming et al., 2011; Suwanprasit and Srichai, 2012; Huiping
et al., 2003). However, for areas identified as grassland, user's accuracy
showed an increasing trend as spatial resolution decreased (i.e. 30 m > Acknowledgments
10 m > 5 m). Similarly, Huiping et al. (2003) found that unlike small
area classes like arbor which had a high accuracy when mapped from The authors gratefully acknowledge the financial support provided
fine resolution image, large area classes like grassland had higher ac- by the German Research Foundation, Germany, DFG, through grant
curacy when extracted from coarser resolution images. number KL894/23-1 as part of the Research Unit FOR2432/1. We are
Other findings caution that finer resolution images may not ne- also thankful for the cooperation and infrastructural support provided
cessarily result in higher classification accuracy and that land cover by our Indian partners at the Institute of Wood Science and Technology
classes exhibit different classification accuracies from various image (IWST), Bengaluru.

95
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Appendix A. Landsat 8 accuracy measures

Table A.5
User(UA) and producer accuracy(PA) for each land cover class(ID) and each 5 × 5 km subset and for the whole transect.

ID Landsat 8

Urban Transition

UA PA UA PA UA PA UA PA UA PA

1 0.77 ± 0.02 0.89 ± 0.02 0.80 ± 0.02 0.86 ± 0.02 0.72 ± 0.02 0.81 ± 0.02 0.71 ± 0.03 0.72 ± 0.02 0.61 ± 0.07 0.51 ± 0.06
2 0.70 ± 0.03 0.52 ± 0.04 0.77 ± 0.03 0.65 ± 0.04 0.50 ± 0.05 0.29 ± 0.05 0.54 ± 0.05 0.33 ± 0.05 0.30 ± 0.11 0.11 ± 0.06
3 0.62 ± 0.27 0.06 ± 0.24 0.25 ± 0.50 0.01 ± 0.14 0.20 ± 0.39 0.01 ± 0.05 1.00 ± 0.00 0.01 ± 0.00 0.00 ± 0.00 0.00 ± 0.00
4 0.42 ± 0.20 0.83 ± 0.16 0.17 ± 0.11 0.42 ± 0.12 0.62 ± 0.09 0.80 ± 0.08 0.74 ± 0.07 0.90 ± 0.06 0.60 ± 0.32 0.46 ± 0.28
5 0.04 ± 0.07 0.25 ± 0.24 0.17 ± 0.07 0.43 ± 0.09 0.64 ± 0.04 0.53 ± 0.04 0.67 ± 0.03 0.68 ± 0.03 0.82 ± 0.02 0.73 ± 0.02
6 0.02 ± 0.04 0.50 ± 0.15 0.05 ± 0.04 0.43 ± 0.09 0.16 ± 0.03 0.44 ± 0.04 0.22 ± 0.03 0.69 ± 0.03 0.61 ± 0.03 0.84 ± 0.02
7 0.40 ± 0.13 0.28 ± 0.13 0.45 ± 0.12 0.41 ± 0.10 0.15 ± 0.05 0.23 ± 0.05 0.29 ± 0.07 0.30 ± 0.07 0.15 ± 0.07 0.26 ± 0.07
OA 0.72 ± 0.01 0.71 ± 0.01 0.56 ± 0.02 0.54 ± 0.02 0.68 ± 0.01

ID Landsat 8

Rural Transect

UA PA UA PA UA PA UA PA UA PA

1 0.48 ± 0.12 0.34 ± 0.09 0.58 ± 0.09 0.43 ± 0.09 0.84 ± 0.111 0.41 ± 0.12 0.74 ± 0.10 0.57 ± 0.11 0.75 ± 0.01 0.81 ± 0.01
2 0.13 ± 0.24 0.01 ± 0.10 0.00 ± 0.00 0.00 ± 0.00 0.42 ± 0.291 0.19 ± 0.20 0.75 ± 0.22 0.07 ± 0.22 0.67 ± 0.02 0.42 ± 0.02
3 0.00 ± 0.00 0.00 ± 0.00 – – – – – – 0.50 ± 0.188 0.02 ± 0.089
4 0.62 ± 0.21 0.57 ± 0.21 0.57 ± 0.27 0.62 ± 0.22 0.81 ± 0.154 0.55 ± 0.13 0.83 ± 0.33 0.50 ± 0.26 0.64 ± 0.05 0.74 ± 0.04
5 0.89 ± 0.02 0.68 ± 0.02 0.88 ± 0.02 0.70 ± 0.02 0.93 ± 0.012 0.72 ± 0.02 0.90 ± 0.01 0.73 ± 0.02 0.85 ± 0.01 0.70 ± 0.00
6 0.51 ± 0.03 0.85 ± 0.02 0.51 ± 0.03 0.84 ± 0.02 0.48 ± 0.028 0.89 ± 0.02 0.48 ± 0.03 0.88 ± 0.02 0.46 ± 0.01 0.84 ± 0.00
7 0.12 ± 0.08 0.25 ± 0.12 0.04 ± 0.05 0.11 ± 0.09 – – – – 0.21 ± 0.03 0.29 ± 0.03
OA 0.68 ± 0.01 0.69 ± 0.02 0.73 ± 0.01 0.71 ± 0.013 0.68 ± 0.005

Appendix B. Sentinel-2 accuracy measures

Table B.6
User(UA) and producer accuracy(PA) for each land cover class(ID) and each 5 × 5 km subset and for the whole transect.

ID Sentinel-2

Urban Suburban

UA PA UA PA UA PA UA PA UA PA

1 0.83 ± 0.016 0.93 ± 0.033 0.86 ± 0.016 0.92 ± 0.017 0.81 ± 0.021 0.87 ± 0.015 0.80 ± 0.026 0.84 ± 0.020 0.66 ± 0.056 0.81 ± 0.045
2 0.82 ± 0.028 0.66 ± 0.027 0.87 ± 0.022 0.70 ± 0.046 0.73 ± 0.044 0.48 ± 0.048 0.68 ± 0.043 0.50 ± 0.049 0.36 ± 0.075 0.30 ± 0.059
3 0.67 ± 0.326 0.04 ± 0.287 0.67 ± 0.326 0.04 ± 0.190 0.44 ± 0.344 0.02 ± 0.164 0.76 ± 0.207 0.05 ± 0.150 0.00 ± 0.000 0.00 ± 0.000
4 0.69 ± 0.234 0.92 ± 0.135 0.81 ± 0.172 0.89 ± 0.159 0.93 ± 0.056 0.96 ± 0.052 0.88 ± 0.055 0.93 ± 0.035 0.75 ± 0.255 0.69 ± 0.219
5 0.02 ± 0.046 0.25 ± 0.206 0.20 ± 0.071 0.53 ± 0.156 0.66 ± 0.040 0.58 ± 0.043 0.75 ± 0.032 0.71 ± 0.030 0.89 ± 0.017 0.78 ± 0.017
6 0.01 ± 0.024 0.50 ± 0.126 0.04 ± 0.028 0.43 ± 0.069 0.14 ± 0.032 0.44 ± 0.039 0.26 ± 0.034 0.73 ± 0.029 0.65 ± 0.029 0.78 ± 0.020
7 0.73 ± 0.121 0.48 ± 0.114 0.55 ± 0.107 0.62 ± 0.118 0.28 ± 0.077 0.33 ± 0.074 0.45 ± 0.083 0.41 ± 0.075 0.19 ± 0.080 0.28 ± 0.080
OA 0.79 ± 0.014 0.78 ± 0.012 0.64 ± 0.015 0.64 ± 0.015 0.72 ± 0.014

ID Sentinel-2

Rural Transect

UA PA UA PA UA PA UA PA UA PA

1 0.50 ± 0.095 0.57 ± 0.075 0.70 ± 0.067 0.77 ± 0.059 0.73 ± 0.090 0.77 ± 0.075 0.64 ± 0.090 0.80 ± 0.069 0.81 ± 0.009 0.90 ± 0.006
2 0.29 ± 0.118 0.23 ± 0.089 0.25 ± 0.103 0.26 ± 0.084 0.15 ± 0.080 0.44 ± 0.091 0.34 ± 0.097 0.18 ± 0.056 0.73 ± 0.016 0.57 ± 0.017
3 0.00 ± 0.000 0.00 ± 0.000 0.33 ± 0.653 0.08 ± 0.854 – – 0.00 ± 0.000 0.00 ± 0.000 0.47 ± 0.114 0.04 ± 0.055
4 0.88 ± 0.157 0.65 ± 0.202 0.86 ± 0.190 0.86 ± 0.174 0.95 ± 0.066 0.89 ± 0.088 0.86 ± 0.280 0.60 ± 0.202 0.86 ± 0.036 0.86 ± 0.030
5 0.91 ± 0.013 0.78 ± 0.016 0.94 ± 0.011 0.76 ± 0.014 0.96 ± 0.009 0.72 ± 0.014 0.94 ± 0.012 0.73 ± 0.015 0.89 ± 0.006 0.74 ± 0.007
6 0.61 ± 0.029 0.82 ± 0.020 0.59 ± 0.029 0.87 ± 0.020 0.48 ± 0.029 0.85 ± 0.020 0.48 ± 0.029 0.85 ± 0.018 0.49 ± 0.012 0.80 ± 0.008
7 0.04 ± 0.060 0.07 ± 0.095 0.15 ± 0.112 0.33 ± 0.128 – – – – 0.35 ± 0.038 0.39 ± 0.035
OA 0.75 ± 0.013 0.77 ± 0.012 0.74 ± 0.011 0.72 ± 0.012 0.74 ± 0.005

96
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Appendix C. RapidEye accuracy measures

Table C.7
User(UA) and producer accuracy(PA) for each land cover class(ID) and each 5 × 5 km subset and for the whole transect.

ID RapidEye

Urban Suburban

UA PA UA PA UA PA UA PA UA PA

1 0.90 ± 0.013 0.97 ± 0.000 0.90 ± 0.014 0.94 ± 0.013 0.82 ± 0.020 0.87 ± 0.016 0.86 ± 0.024 0.80 ± 0.020 0.75 ± 0.061 0.66 ± 0.059
2 0.95 ± 0.015 0.81 ± 0.000 0.94 ± 0.015 0.81 ± 0.024 0.76 ± 0.041 0.51 ± 0.039 0.74 ± 0.042 0.52 ± 0.044 0.57 ± 0.105 0.26 ± 0.093
3 0.91 ± 0.117 0.17 ± 0.185 0.58 ± 0.201 0.10 ± 0.108 0.25 ± 0.084 0.16 ± 0.063 0.57 ± 0.166 0.08 ± 0.114 – –
4 0.50 ± 0.204 0.92 ± 0.156 0.59 ± 0.188 0.89 ± 0.140 0.65 ± 0.087 0.93 ± 0.085 0.78 ± 0.067 0.91 ± 0.062 0.24 ± 0.117 0.92 ± 0.147
5 0.00 ± 0.000 0.00 ± 0.000 0.12 ± 0.048 0.81 ± 0.062 0.54 ± 0.037 0.64 ± 0.028 0.53 ± 0.030 0.74 ± 0.023 0.75 ± 0.020 0.85 ± 0.013
6 0.00 ± 0.000 0.00 ± 0.000 0.07 ± 0.073 0.23 ± 0.127 0.27 ± 0.073 0.25 ± 0.079 0.29 ± 0.045 0.51 ± 0.042 0.70 ± 0.032 0.61 ± 0.027
7 0.52 ± 0.129 0.41 ± 0.000 0.63 ± 0.099 0.65 ± 0.185 0.24 ± 0.065 0.37 ± 0.061 0.33 ± 0.081 0.28 ± 0.072 0.25 ± 0.106 0.27 ± 0.108
OA 0.87 ± 0.016 0.83 ± 0.015 0.66 ± 0.016 0.62 ± 0.016 0.69 ± 0.016

ID RapidEye

Rural Transect

UA PA UA PA UA PA UA PA UA PA

1 0.49 ± 0.110 0.41 ± 0.095 0.74 ± 0.068 0.67 ± 0.077 0.74 ± 0.117 0.41 ± 0.112 0.82 ± 0.092 0.56 ± 0.092 0.86 ± 0.008 0.88 ± 0.006
2 0.46 ± 0.282 0.08 ± 0.278 0.32 ± 0.144 0.17 ± 0.110 0.53 ± 0.156 0.28 ± 0.145 0.55 ± 0.160 0.11 ± 0.118 0.85 ± 0.013 0.62 ± 0.013
3 0.00 ± 0.000 0.00 ± 0.000 0.00 ± 0.000 0.00 ± 0.000 0.00 ± 0.000 0.00 ± 0.000 – – 0.42 ± 0.071 0.10 ± 0.050
4 0.23 ± 0.090 0.83 ± 0.099 0.37 ± 0.111 0.84 ± 0.101 0.88 ± 0.065 0.85 ± 0.064 0.95 ± 0.045 0.85 ± 0.061 0.63 ± 0.035 0.89 ± 0.033
5 0.83 ± 0.017 0.84 ± 0.012 0.86 ± 0.015 0.85 ± 0.014 0.88 ± 0.014 0.87 ± 0.013 0.91 ± 0.013 0.87 ± 0.014 0.78 ± 0.007 0.84 ± 0.006
6 0.69 ± 0.032 0.67 ± 0.027 0.69 ± 0.032 0.72 ± 0.023 0.64 ± 0.034 0.73 ± 0.025 0.54 ± 0.035 0.80 ± 0.023 0.60 ± 0.014 0.67 ± 0.011
7 0.06 ± 0.062 0.11 ± 0.078 0.12 ± 0.130 0.17 ± 0.229 – – 0.07 ± 0.097 1.00 0.174 0.32 ± 0.036 0.37 ± 0.034
OA 0.74 ± 0.013 0.78 ± 0.013 0.80 ± 0.012 0.79 ± 0.013 0.76 ± 0.005

References accuracy, segmentation scale, image resolution. In: Proceedings of the Geoscience
and Remote Sensing Symposium, IGARSS'03. IEEE International., vol. 6. pp.
3671–3673.
Alsheri, M., 2010. The effect of spatial resolution on classification accuracy of marine and India, 2011. Census of India, 2011. Directorate of Census Operations, Karnataka and
coastal areas. URL 〈https://www.geospatialworld.net/article/the-effect-of-spatial- [Office of the Registrar General & Census Commissioner, India, Ministry of Home
resolution-on-classification-accuracy-of-marine-and-coastal-areas/〉. Affairs, Govt. of India], Bangalore and [New Delhi].
Angel, S., Parent, J., Civco, D., Blei, A., 2010. Atlas of urban expansion, Cambridge, MA: Kingma, D. P., Ba, J., Dec. 2014. Adam: A Method for Stochastic Optimization.ArXiv e-
Lincoln institute of land policy. prints.
Benitez, P., McCallum, I., Obersteiner, M., Yamagata, Y., May 2004. Global supply for Lechner, A.M., Stein, A., Jones, S.D., Ferwerda, J.G., 2009. Remote sensing of small and
carbon sequestration: Identifying least-cost afforestation sites under country risk linear features: quantifying the effects of patch size and length, grid position and
consideration. IIASA interim report, IIASA, Laxenburg, Austria. URL 〈http://pure. detectability on land cover mapping. Remote Sens. Environ. 113 (10), 2194–2204.
iiasa.ac.at/id/eprint/7424/〉. LeCun, Y., Bottou, L., Orr, G.B., Müller, K.-R., 1998. Efficient backprop. In: Neural net-
Bischof, H., Schneider, W., Pinz, A., 1992. Multispectral classification of landsat-images works: Tricks of the trade. Springer, pp. 9–50.
using neural networks. IEEE Trans. Geosci. Remote Sens. 30, 3. Lim, A., Hedley, J.D., LeDrew, E., Mumby, P.J., Roelfsema, C., 2009. The effects of
Chen, D., Stow, D., Gong, P., 2004. Examining the effect of spatial resolution and texture ecologically determined spatial complexity on the classification accuracy of simu-
window size on classification accuracy: an urban environment case. Int. J. Remote lated coral reef images. Remote Sens. Environ. 113 (5), 965–978.
Sens. 25 (11), 2177–2192. Lu, D., Hetrick, S., Moran, E., 2010. Land cover classification in a complex urban-rural
Chen, Y., Song, X., Wang, S., Huang, J., Mansaray, L.R., 2016. Impacts of spatial het- landscape with quickbird imagery. Photogramm. Eng. Remote Sens. 76 (10),
erogeneity on crop area mapping in canada using modis data. ISPRS J. Photogramm. 1159–1168.
Remote Sens. 119, 451–461. Lu, D., Weng, Q., 2005. Urban classification using full spectral information of landsat etm.
Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely imagery in marion county, indiana. Photogramm. Eng. Remote Sens. 71 (11),
sensed data. Remote Sens. Environ. 37 (1), 35–46. 1275–1284.
Congalton, R.G., Green, K., 2008. Assessing the Accuracy of Remotely Sensed Data: Lu, D., Weng, Q., 2007. A survey of image classification methods and techniques for
Principles and Practices. CRC Press. improving classification performance. Int. J. Remote Sens. 28 (5), 823–870.
Dabboor, M., Howell, S., Shokr, M., Yackel, J., 2014. The jeffries-matusita distance for the Lu, D., Weng, Q., 2009. Extraction of urban impervious surfaces from an ikonos image.
case of complex wishart distribution as a separability criterion for fully polarimetric Int. J. Remote Sens. 30 (5), 1297–1311.
sar data. Int. J. Remote Sens. 35 (19), 6859–6873. Luck, M., Wu, J., 2002. A gradient analysis of urban landscape pattern: a case study from
Fritz, S., See, L., McCallum, I., Schill, C., Obersteiner, M., Van der Velde, M., Boettcher, the phoenix metropolitan region, Arizona, USA. Landsc. Ecol. 17 (4), 327–339.
H., Havlík, P., Achard, F., 2011. Highlighting continued uncertainty in global land MacLean, M.G., Congalton, R.G., 2012. Map accuracy assessment issues when using an
cover maps for the user community. Environ. Res. Lett. 6 (4), 044005. object-oriented approach. In: Proceedings of the American Society for
Goetz, S.J., Wright, R.K., Smith, A.J., Zinecker, E., Schaub, E., 2003. Ikonos imagery for Photogrammetry and Remote Sensing Annual Conference. pp. 1–5.
resource management: tree cover, impervious surfaces, and riparian buffer analyses Malkinson, D., Kopel, D., Wittenberg, L., 2018. From rural-urban gradients to patch-
in the mid-atlantic region. Remote Sens. Environ. 88 (1–2), 195–208. matrix frameworks: plant diversity patterns in urban landscapes. Landsc. Urban Plan.
Goodfellow, I., Bengio, Y., Courville, A., 2016. Deep Learning. MIT Press(〈http://www. 169, 260–268.
deeplearningbook.org〉). McGarigal, K., Cushman, S., Neel, M., Ene, E., 01 2002. Fragstats: Spatial pattern analysis
Herold, A., 2001. Remote sensing and spatial metrics-a new approach for the description program for categorical maps.
of structures and changes in urban areas. In: Proceedings of Geoscience and Remote McGarigal, K., Marks, B.J., 1995. Spatial pattern analysis program for quantifying land-
Sensing International Symposium, IGARSS'01. vol. 1. IEEE, pp. 366–368. scape structure. Gen. Tech. Rep. PNW-GTR-351. US Department of Agriculture,
Hiremath, S., Prabhuraj, D.K., Lakshmikantha, B.P., Chakraborty, S.D., 2013. Land use/ Forest Service, Pacific Northwest Research Station.
land cover change analysis of bangalore urban district and its impact on land surface Messerli, P., Heinimann, A., Epprecht, M., 2009. Finding homogeneity in hetero-
temperature. In: Proceedings of ISRS Conference Paper, pp. 4–6. geneity—a new approach to quantifying landscape mosaics developed for the lao pdr.
Hsieh, P.-F., Lee, L.-C., 2000. Effect of spatial resolution on classification error in remote Hum. Ecol. 37 (3), 291–304.
sensing. In: Proceedings of the Geoscience and Remote Sensing International Ming, D., Yang, J., Li, L., Song, Z., 2011. Modified alv for selecting the optimal spatial
Symposium, IGARSS 2000, vol. 1. IEEE, pp. 171–173. resolution and its scale effect on image classification accuracy. Math. Comput. Model.
Huiping, H., Bingfang, W., Jinlong, F., 2003. Analysis to the relationship of classification 54 (3), 1061–1068.

97
K.T. Awuah et al. Remote Sensing Applications: Society and Environment 12 (2018) 89–98

Momeni, R., Aplin, P., Boyd, D.S., 2016. Mapping complex urban land cover from 1929–1958.
spaceborne imagery: the influence of spatial resolution, spectral band set and clas- Sudhira, H., Nagendra, H., 2013. Local assessment of bangalore: graying and greening in
sification approach. Remote Sens. 8 (2), 88. bangalore–impacts of urbanization on ecosystems, ecosystem services and biodi-
Mora, B., Tsendbazar, N.-E., Herold, M., Arino, O., 2014. Global land cover mapping: versity. In: Urbanization, Biodiversity and Ecosystem Services: Challenges and
Current status and future trends. In: Land Use and Land Cover Mapping in Europe. Opportunities. Springer, pp. 75–91.
Springer, pp. 11–30. Sugumaran, R., Zerr, D., Prato, T., 2002. Improved urban land cover mapping using multi-
Nesbitt, L., Meitner, M.J., 2016. Exploring relationships between socioeconomic back- temporal ikonos images for local government planning. Can. J. Remote Sens. 28 (1),
ground and urban greenery in Portland, OR. Forests 7 (8), 162. 90–95.
Olofsson, P., Foody, G.M., Herold, M., Stehman, S.V., Woodcock, C.E., Wulder, M.A., Suwanprasit, C., Srichai, N., 2012. Impacts of spatial resolution on land cover classifi-
2014. Good practices for estimating area and assessing accuracy of land change. cation. Proc. Asia-Pac. Adv. Netw. 33, 39–47.
Remote Sens. Environ. 148, 42–57. Tran, T.V., Julian, J.P., de Beurs, K.M., 2014. Land cover heterogeneity effects on sub-
Ponzoni, F.J., Galvão, L.S., Epiphanio, J.C., 2002. Spatial resolution influence on the pixel and per-pixel classifications. ISPRS Int. J. Geo-Inform. 3 (2), 540–553.
identification of land cover classes in the amazon environment. An. da Acad. Bras. Van der Sande, C., De Jong, S., De Roo, A., 2003. A segmentation and classification ap-
De. Ciências 74 (4), 717–725. proach of ikonos-2 imagery for land cover mapping to assist flood risk and flood
Richards, J.A., Jia, X., 2006. Interpretation of hyperspectral image data. In: Remote damage assessment. Int. J. Appl. Earth Obs. Geoinform. 4 (3), 217–229.
Sensing Digital Image Analysis: An Introduction, pp. 359–388. Van Niel, T.G., McVicar, T.R., Datt, B., 2005. On the relationship between training sample
Schüle, S.A., Gabriel, K.M., Bolte, G., 2017. Relationship between neighbourhood socio- size and data dimensionality: monte carlo analysis of broadband multi-temporal
economic position and neighbourhood public green space availability: An environ- classification. Remote Sens. Environ. 98 (4), 468–480.
mental inequality analysis in a large german city applying generalized linear models. Venkatesh, Y., Raja, S.K., 2003. On the classification of multispectral satellite images
Int. J. Hyg. Environ. Health. using the multilayer perceptron. Pattern Recognit. 36 (9), 2161–2175(Kernel and
Sertel, E., Robock, A., Ormeci, C., 2010. Impacts of land cover data quality on regional Subspace Methods for Computer Vision. URL 〈http://www.sciencedirect.com/
climate simulations. Int. J. Climatol. 30 (13), 1942–1953. science/article/pii/S003132030300013X〉).
Smith, J.H., Stehman, S.V., Wickham, J.D., Yang, L., 2003. Effects of landscape char- Wei, C., Blaschke, T., Kazakopoulos, P., Taubenböck, H., Tiede, D., 2017. Is spatial re-
acteristics on land-cover class accuracy. Remote Sens. Environ. 84 (3), 342–349. solution critical in urbanization velocity analysis? Investigations in the pearl river
Smith, J.H., Wickham, J.D., Stehman, S.V., Yang, L., 2002. Impacts of patch size and land- delta. Remote Sens. 9 (1), 80.
cover heterogeneity on thematic image classification accuracy. Photogramm. Eng. Wu, J., Jelinski, D.E., Luck, M., Tueller, P.T., 2000. Multiscale analysis of landscape
Remote Sens. 68 (1), 65–70. heterogeneity: scale variance and pattern metrics. Geogr. Inform. Sci. 6 (1), 6–19.
Solecka, I., Sylla, M., Świader, M., 2017. Urban sprawl impact on farmland conversion in Wu, W., Shibasaki, R., Yang, P., Ongaro, L., Zhou, Q., Tang, H., 2008. Validation and
suburban area of wroclaw, poland. IOP Conf. Ser.: Mater. Sci. Eng. 245 (7), comparison of 1 km global land cover products in china. Int. J. Remote Sens. 29 (13),
072002(URL 〈http://stacks.iop.org/1757-899X/245/i=7/a=072002〉). 3769–3785.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout: Yu, B.-j., Tang, Y.-h., Li, H., Bi, L.-l., 2016. Study on spatial heterogeneity of landscape
a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15 (1), fragmentation and its urbanization response. J. Residuals Sci. Technol. 13, 7.

98

You might also like