Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Ocean & Coastal Management 69 (2012) 102e110

Contents lists available at SciVerse ScienceDirect

Ocean & Coastal Management


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

A comparison between three short-term shoreline prediction models


Rodrigo Mikosz Goncalves a, Joseph L. Awange b, *, Claudia Pereira Krueger c, Bernhard Heck d,
Leandro dos Santos Coelho e
a
Department of Cartography Engineering, Federal University of Pernambuco (UFPE), Geodetic Science and Technology of Geoinformation Post Graduation Program,
Recife 50670-901, PE, Brazil
b
Department of Spatial Sciences, Curtin University, GPO Box U1987, Perth, WA 6845, Australia
c
Geodetic Science Post Graduation Program, Federal University of Parana (UFPR), Box 19.001, 81.531-990 Curitiba, PR, Brazil
d
Geodetic Institute, Karlsruhe Institute of Technology, Engler-Strasse 7, D-76131 Karlsruhe, Germany
e
Pontifícal Catholic University of Parana, PUCPR Production and Systems Engineering Graduate Program, LAS/PPGEPS Imaculada Conceicao, 1155, 80215-901 Curitiba, PR, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Monitoring and management of shorelines along populated coastal areas is a very important task, but
Available online 23 August 2012 remains a difficult endeavor. The historical information used for short-term analysis and prediction are
always underpinned by uncertainties associated with old data. Predictions of shoreline positions nor-
mally depend on the accuracy of the input data as well as the validity of the mathematical models used.
With the requirement to study shoreline changes along the Parana (PR) coast in Brazil, it was necessary
to obtain related cartographic information, which included temporal shoreline data obtained from
orthophotos. In this contribution, photogrammetric together with GPS data are used to compare the
capability of three shoreline prediction models; linear regression, robust parameter estimation, and neural
network to predict the 2008 Parana shoreline position, which is then validated using the GPS measured
position of 2008. The results indicate a MAPE (Mean Absolute Percentage Error) of 0.61% for the linear
regression, 0.14% for the robust estimation, and 0.33% for the artificial neural network method. Although
the coefficient of determinant (R2) value for the neural network was the best, i.e., 0.997 compared to
0.994 for the robust model and 0.984 for the linear regression, its maximum deviation from the control
values (i.e., 16.46) was almost twice that of robust model (7.63). On the one hand, the robust estimation
model provides a more suitable approach for managing outliers in shoreline prediction, and also vali-
dating traditional methods such as linear regression. On the other hand, the neural network method
offers an alternative approach to the robust prediction model. The results of the study highlightthe
importance of a model choice for predicting the shoreline position.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction satellite images, aerial photography, profile surveys and bathy-


metric records. Remote sensing is important for creating base maps
Most coastal areas are known to experience erosion. In Brazil, able to interpret landform changes and quantify shoreline move-
for example, hundreds of beaches are known to suffer severe ment. The hydrographic and also large-scale topographic maps are
erosion problems (e.g., Souza, 2009). Beaches are sources of considered the primary sources of information for shoreline posi-
revenue for many countries through tourism, and as such, neces- tions and volumetric change computations. The use of profile
sitate proper management in coastal areas. For the efficient surveys can also contribute to evaluating shoreline changes and
management of beaches, however, one of the important tasks computing beach volume changes along and across the shore.
remains that of accurately monitoring the coastal shorelines. The importance of coastal monitoring information, e.g., on the
Gorman et al. (1998) presented an overview of the methods of data state of the current erosion and the possible prediction of its impact
collection for coastal monitoring and baseline studies using have been shown by Li et al. (1998) to be critical to decision-making
in coastal management. Li et al. (1998) point to the fact that coastal
monitoring information could be useful in land-use zoning,
* Corresponding author.
E-mail addresses: rodrigo.mikosz@ufpe.br (R.M. Goncalves), J.Awange@
construction setbacks, and relocation implementations. They give
curtin.edu.au (J.L. Awange), ckrueger@ufpr.br (C.P. Krueger), heck@kit.edu a practical example of a digital topographical map that is overlaid
(B. Heck), leandro.coelho@pucpr.br (L.dosS. Coelho). with a cadastral map and an erosion condition map, which is used to

0964-5691/$ e see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ocecoaman.2012.07.024
R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110 103

find lots and parcels that are affected by coastal erosion. Another for observation, a challenge for detecting, monitoring, and pre-
area that requires shoreline monitoring is that of the integrated dicting shoreline movement is to develop an effective mathematical
coastal zone management (ICZM), where such information assist in tool capable of identifying and mapping a feature set as “shoreline”
the process of coastal management and territorial planning/ according to the data sources available (Boak and Turner, 2005). For
decision-making by providing essential knowledge about coastal instance, exploiting the relationship between shoreline and sea
processes and their dynamic evolution (e.g., Veloso-Gomes et al., level changes (i.e., using series of sparsely sampled sea-level values
2008). ICZM has been shown to be important for example in the as surrogate data for shoreline change), Douglas et al. (1998)
detection of landscape and vegetation changes, and to establish the developed an algorithm that evaluates some of the predictive
relations between human impact and vegetation in natural coastal methods such as the end point method, linear regression, and
dune systems (see, e.g., Tzatzanis et al., 2003). minimum description length criterion and established that linear
Monitoring of beach stability, therefore, is vital for environ- regression gave superior results. The need for an effective mathe-
mental as well as resource management, and is essential for matical tool is further emphasized by Li et al. (2001) who stress the
improving databases of information on shoreline evolution in an need to integrate specific coastal engineering and modeling soft-
area. The Metropolitan Borough of Sefton (2002) listed the benefits ware packages in a GIS environment since they are usually not
of shoreline evolution information as; providing input to shoreline provided by commercial GIS software system.
review plans, planned maintenance of coastal defenses, achievement The problem with the linear regression models is that they work
of high government level targets, determination of appropriate design well when the underlying linearity and normal distribution
criteria for coastal works, biodiversity action plans, implementation of assumptions are fulfilled. However, in some cases, where the data
habitats directive, and leisure and amenity management of shoreline maybe of poor quality, the linear regression assumptions may be
areas. While shoreline monitoring, as discussed above is essential, violated. Uncertainties in the extracted shoreline data need to be
predicting its future position is equally vital to support coastal appropriately addressed if they are to be used to predict future
management and impact assessment programs, e.g., predicting the shoreline positions to support sustainable coastal management
necessary future building setbacks from the shoreline to serve as (see, e.g., Addo et al., 2008). Robust estimation models have been
protection for a time comparable to the expected lifetime of new proposed to deal with cases where data contains uncertainties (see,
coastal structures, usually 30 or 60 years (Hecky et al., 1984; e.g., Huber, 1964, 1981; and Hampel et al., 1986).
Crowell et al., 1997). Crowell et al. (1997) point out, however, that Using the positions of shorelines over time, inferred from
determining adequate setbacks require estimating long-term photogrammetric and GPS surveys for the years 2001, 2002, 2005
shoreline change trends from historical data. For historical data- and 2008, this study discusses the possibility of developing a short-
set, metric quality and aerial photographs are very important and term prediction model using different sources of temporal geodetic
are some of the most reliable data sets available for coastal zones. data, with most of them having different accuracies. To achieve this,
Crowell et al. (1997) and Douglas et al. (1998) reckon that often the a comparison and assessment of three different shoreline predic-
data used in such predictions are at times temporally poorly tion models; linear regression, robust parameter estimation, and
sampled historical databases. They point out that short records are artificial neural network is undertaken.
not good predictors of shoreline location. The robust estimation model used in this work has been
Fenster et al. (1993) developed a predictive method that detects successfully applied by Awange and Aduol (1999) to estimate
short-term changes in the long-term trend and identifies linear or geodetic parameters in the case of contaminated observations. The
high-order polynomial models that best fit the data according to artificial neural network method tested in this paper is termed
the minimum description length (MDL) criterion. In this method, Neural Network Multilayer Perception (MLP), and was imple-
only linear models are extrapolated. Predictions shaped or influ- mented using the training of LevenbergeMarquardt algorithm
enced by higher-order polynomial schemes can sometimes be (Hagan and Menhaj, 1994).
superior to those obtained from linear regressions, but they can The study is organized as follows. In Sect. 2, the background,
also be extremely inaccurate (Crowell et al., 1997). Douglas et al. study area, data, and the shoreline prediction models are discussed.
(1998) stressed the need to incorporate long-term erosion trends The results are then presented and discussed in Sect. 3, and the
and historical record of storms, including their impacts on shore- study concluded in Sect. 4.
line positions and beach recovery in predictive models. Thieler et al.
(2000) tested the use of mathematical models to predict beach 2. Data and methods
behavior for U.S. and cited lack of hind-sighting and objective
evaluation of beach behavior predictions for engineering projects, 2.1. Background
and incorrect use of model calibration as some of the motivation
behind their work. Since 1996, the Spatial Geodesy Laboratory and Hydrography
Shorelines are known not to be stable and vary over time. Short- (LAGEH) at the Federal University of Parana (UFPR) have been
term changes occur over decadal time scales, or less, and are related collecting data using GPS surveys over the Parana State coast (see,
to daily, monthly, and seasonal variations in tides, currents, wave e.g. Krueger et al., 2009). The data are composed of digital carto-
climate, episodic events and anthropogenic factors (e.g., Demarest graphic documents with reports showing the post-processing and
and Leatherman, 1985; Galgano et al., 1998; Galgano and Douglas, accuracy reached in a specific survey. With a proposal to study
2000). Fenster et al. (2000) describe shoreline movement as shoreline changes along the Parana coast, some analog images/
a complex phenomenon and outline the difficulties involved in photographs related to the years 1954, 1963, 1980, 1991 and 1997
distinguishing long-term shoreline movement (signal) from short- were retrieved and used for shoreline extraction along a 6 km
term changes (noise). coastal zone of the Matinhos beaches in the state of Parana, Brazil.
Even with the presence of noise in the data, the use of modern
accurate surveying techniques has been shown to improve the 2.2. Study area
quality of forecasts. For example, Douglas and Crowell (2000)
demonstrated that meaningful deduction was achievable even if The study area is located in the Matinhos District along the coast
the inherent variability of shoreline position indicators remained at of Parana State, Brazil (Fig. 1), where the beaches (105 km long)
the level of many meters. Other than the surveying technique used form the second smallest littoral along the Brazilian coast, located
104 R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110

Fig. 1. Matinhos District at the coast of Parana State, Brazil.

between latitudes 25 S and 26 S and longitudes 48 W and 49 W. study area, indicating the prevalence of the problems of coastal
In geological terms, Soares et al. (1995) describe the area as being erosion.
composed of Pre-Cambrian units of crystalline complexes of
igneous and metamorphic rocks. Pleistocene and Holocene sandy 2.3. Data source
sediments of marine origin from a terrace (beach ridges) are found
between the shoreline and the foot of the Serra do Mar Mountains. Monitoring of shoreline positions nowadays benefit from the
The beaches of Matinhos are subject to both oceanic conditions and state of the art mapping techniques such as global navigation
ebb tidal delta influence from the neighboring Guaratuba Bay satellite systems (GNSS, e.g., Goncalves, 2010; Awange, 2012),
(micro-tidal system). The district of Matinhos, with an area of remote sensing using satellite images, aerial photographs (e.g.
117 km2, comprised 29,172 habitants in 2010. Land use is mostly for Mitishita and Kirchner, 2000), and LIDAR (Light Detection and
recreational purposes and in summer, the city becomes more Ranging). For instance, Mitishita and Kirchner (2000) extracted
densely populated due to the influx of tourists. shoreline position from temporal vertical aerial photographs by the
The settlement in Matinhos is largely near the shore, where monorestitution technique, which required altimetry information
settlements are characterized by constructions at backshore posi- and was dependent on the quality of the photos, the distribution
tions or over the beach leading to the destruction of dunes and and number of control points, and photo-interpretation for shore-
wetlands, thereby forcing rivers to change course. This was a result line extraction.
of lack of urban planning and the fact that the morphology and In the present study, aerial photographs and GPS observations
coastal dynamic environment was not considered during the collected for a 6 km part of the coastal zone were used, where the
settlements. Fig. 2 shows the situation in 2007 and 2008 at the photogrammetric data was converted to digital orthophotos using

Fig. 2. State of the Matinhos beach in 2007 and 2008 (Goncalves, 2010).
R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110 105

control points and digital terrain model (DTM). The shoreline where x and y are the linear regression vectors, while a and c are
positions were then obtained using the monorestitution technique the unknown linear regression coefficient and intercept respec-
based on photo interpretation. The GPS data was collected during tively, and 1 is a column vector consisting of “1” numbers. Equation
a series of geodetic surveys of the shoreline in Matinhos using the (1) can be expressed in terms of the observations and unknowns as
kinematic relative positioning method (see, e.g., Awange, 2012).
The GPS survey was undertaken by LAGEH (Laboratory of Space y ¼ Ab þ ε; (2)
Geodesy and Hydrography) in UFPR (Federal University of Parana)
using dual frequency GPS receivers. The base station was installed with y being an n  1 column vector of observations, A is an n  m
at Pedra located at 25 490 5.779900 S; 48 310 49.136400 W in design matrix (here m ¼ 2), b is the m  1 vector of unknown linear
Matinhos. regression parameters, and ℇ is the n  1 vector of observational
All temporal vectors of shorelines from photogrammetry for the errors. Equation (2) can now be expressed in the form of the
years 1954, 1963, 1980, 1991 and 1997 (Fig. 3a) and from GPS for the GausseMarkov estimation model as:
years 2001, 2002, 2005 and 2008 (see Fig. 3b) were placed in layers  
and processed in the same geodetic reference frame (World y ¼ Ab þ ε; εw 0; s20 W 1 ; (3)
Geodetic System; WGS-84). For distance computations, planar
coordinates (UTM) were used. The shoreline positions obtained whose solution (linear regression) is given by
from photogrammetric data did not always begin and end at the
 1
same location, and also had gaps, and as such did not cover all the b
b ¼ AT WA AT Wy; (4)
information at the study site.

with the dispersion Df b


bg ¼ bs 20 ðAT WAÞ1 and
2.4. Shoreline prediction models sb 20 ¼ ðεT W ε Þ=ðn  mÞ. Here, bs 20 is the variance of unit weight and
W is the n  n positive-definite weight matrix.
In what follows, brief discussions of the linear regression, robust First, we define a reference line from which cross-sections are
estimation, and artificial neural network models are presented. taken at intervals of 100 m. The temporal positions of shorelines are
then obtained from their intersection with the cross-sections.
2.4.1. Linear regression model These positions are then used to calculate the shoreline distances
The linear regression model commonly used in shoreline xn, related to time tn, which forms the elements of the observation
prediction is expressed as vector y in equation (1). The temporal information, i.e. years, are
used in the design matrix A.
The linear regression solution (4) holds when there are no gross
y ¼ ax þ c1; (1) errors and b b becomes an unbiased estimator of b since the

Fig. 3. Temporal resolution of the photogrammetric and GPS data.


106 R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110

expectation Ef b
bg ¼ b. With gross errors in the observations, v ¼ y  Ab
b: (13)
however, we have
The values of v obtained from equation (13) are re-introduced
y ¼ Efyg þ 2 þ ε; (5) into equation (11) to obtain a new weight matrix W. The proce-
dure for the estimation of b
b is then repeated until a sufficient level
where2 is the n  1 bias vector on the observational vector y. of convergence is attained. In this iterative approach, the starting
Considering y ¼ y þ dy, the solution of (5) from (4) leads to: initial values of the weight matrix W are those obtained from the
 1  1 linear regression solution in equation (4).
b
b ¼ AT WA AT Wy þ AT WA AT W dy; (6) This method makes use of the fact that any gross error in the
observation will manifest itself in the estimated residuals, which
will tend to be larger compared to the others due to the effect of the
with the expectation
gross errors. In the re-estimation of the elements of the weight
   1  1 matrix W, a relatively large residual will result in a relatively low
E b
b ¼ AT WA AT Wy þ AT WA A T W 2 ¼ b þ g ¼ b0 ; weight for the respective observation. The procedure therefore
down-weights the influence of outlying observations and in so
(7)
doing, reduces the effect of bad observations on the estimated
and y ¼ Efyg; dy ¼ 2 þ ε: It becomes evident then that b b is quantities.
a biased estimator of b with the termðAT WAÞ1 AT W 2. Setting
g ¼ db, equation (7) leads to 2.4.3. Neural network
In beach morphodynamics applications, Álvarez-Ellacuría et al.
b ¼ b0 þ db: (8) (2011) have shown using an example that artificial neural
networks can be used to obtain shoreline positions from daily raw
2.4.2. Robust estimation model video images where the spatial and temporal variability of the
The robust estimation model is defined as follows: Consider beach are split by Empirical Orhthogonal Function (EOF) decom-
x1,x2,.,xn as being randomly independent variables assumed to position. For the setup and development of the artificial neural
have a normal distribution described by the probability density network (NN), the characteristics of neurons, topology, and training
function F0(x). If a fraction k(0 < k < 1) of the observations are rules must be specified. The adaptation of the initial weights and
contaminated by gross errors, then the new distribution will be learning of their behaviour are specified by the rules of training.
represented by Huber (1964): The algorithms of training of a NN have the feature of adjusting
iteratively the weights of connections between neurons until the
FðxÞ ¼ ð1  kÞF0 ðxÞ þ kHðxÞ; (9) pair of inputs and expected outputs are obtained, and consequently
the relations of cause and effect established. When the setting of
where H(x) is the unknown contaminating distribution. The robust a particular problem presented to NNs changes, and the perfor-
estimation procedure estimates parameters from this setup in such mance model does not fit the situation anymore, it is possible to
a way that the influences of the gross errors in the final estimated train the NN with new conditions of input and output to improve
parameters are significantly reduced. the performance (Kröse and Smagt, 1996; Haykin, 1999).
In adopting a robust estimation procedure, we seek a rigorous The schematic model of an artificial neuron considered as the
adjustment method that will solve equation (6) and give the results processor of a neural network was proposed by McCullochePitts,
that are as close as possible to those of equation (4). One advantage where three basic elements can be identified (see, e.g. Kröse and
of using robust estimation methods is that they do not immediately Smagt, 1996; Haykin, 1999; Arbib, 2003). The first element is a set
delete outlying observations from a given set, rather, they isolate of synapses or connecting links, each characterized by a weight wkj.
their effects through, e.g., down-weighting. The iterative weighting The second element relates to a sum of the input signals, weighted
approach proposed by Aduol (1994), for instance, provides a means by the synapses of the neuron (linear combination) as
of down-weighting bad observations.
In this approach, the weights of observations are considered as X
m
functions of both the observational variances and the observational vk wkj $xj : (14)
residuals. In equation (3), the weight matrix is defined as j¼0

W ¼ S1 ; (10) The third element is a transfer function which limits the output
amplitude of a neuron to a finite value, given by
with S being the covariance matrix of the observations. This weight
is modified in the robust iterative weighting approach such that yk ¼ 4ðnk Þ: (15)
equation (10) takes into consideration the residuals. Equation (10)
is thus re-written as A neuron can be described mathematically using equations (14)
and (15), where 4 represents the transfer function of the artificial
 1
W ¼ R2 þ S ; (11) neuron, which processes the set of inputs received, while changes
in the activation state are required to obtain a good fit or model for
the problem at hand. The log-sigmoid transfer function (Equation
where
(16)) can take values between 0 and 1, where a is the slope
2 3 parameter of the sigmoid function and n is the activation of the
n1 0 : 0
60 n2 : 07 neuron.
R ¼ 6
4 :
7; (12)
: : : 5 1
0 0 : nn f ðnÞ ¼ (16)
1 þ eðanÞ
and n1, n2,., nn are the observational residuals from equation (3) Equation (17) is the hyperbolic tangent function, which takes
expressed as values between 1 and 1, where a is the slope parameter of the
R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110 107

curve; b is the lower and higher limiting values and n is the value of Table 1 show the setting used to perform the neural network. It is
activation. important to report that we tried many more neural network
configurations with different input and hidden layers (or neurons)
eðbnÞ  eðbnÞ to check the performance of each model, with the best one selected
f ð nÞ ¼ a (17)
eðbnÞ þ eðbnÞ and compared to the other methods.
The linear activation function is defined by Equation (18) where
2.5. Evaluation criteria
a is a real number that defines the linear output to the input values,
y is the output and x is the input.
Three experiments were performed; the first with the linear
y ¼ ax (18) regression model (equation (4)), the second with the robust esti-
mation model (equations (10)e(13)), and the third with the neural
An important concept of a NN is the definition of the architec- network (Table 1). To assess the efficiency of the three methods, we
ture, or the way in which neurons in a network may be arranged, computed the following:
which is an important parameter that restricts the type of problem
that can be treated in a specific network. Networks with a single (1) The mean of the deviation from the GPS 2008 derived shore-
layer of nodes, for example, can only solve linearly separable line, which we used as a control. This is computed using
problems. Recurrent networks, in turn, are more appropriate to
solve problems that involve temporal processing (Haykin, 1999).
1X n
The Multilayer Feed-Forward Network was used with the x ¼ ðGPS 2008i  Prediction 2008i Þ; (19)
LevenbergeMarquardt (LM) algorithm for training in this work.
n i¼1
This type of architecture is composed of an input layer of source,
one or more layers of hidden neurons and another layer of output where (GPS 2008i  Prediction 2008i) denotes the distance
neurons. There are several types of methods for training networks between the predicted shoreline and the control shoreline at the
that are grouped into two categories: supervised learning and cross-section, while i and n are the number of sections.
unsupervised learning. The case of supervised learning or learning
with a teacher has a set of inputeoutput examples. However in the (2) The standard deviation is then obtained from the square root of
case of learning without a teacher (unsupervised learning), there is the average variance of the estimated shoreline position, i.e.,
neither teacher to oversee the learning process nor are there
labeled examples of the function to be learned by the network (see, vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
e.g., Haykin, 1999). u 1 X n
The most popular algorithm of error management is the back s¼ t ðGPS 2008i  Prediction 2008i  xÞ2 (20)
n  1 i¼1
propagation, were the idea is to correct the errors according to
a learning rule. The learning consists of two steps through the
different layers of the net according to their direction; one step (3) The root mean square (RMS) of the shoreline deviation is
forward, the propagation, and one step backward, the back prop- computed to gain a measure of the corresponding effectiveness
agation (e.g. Kröse and Smagt, 1996; Haykin, 1999; Arbib, 2003). of the method, i.e.,
In the forward step, one vector of input is applied to the
sensorial nodes and it in effect propagates through the net, layer by
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
layer. This results in one set of output to be produced with a real u
u 1 X n
answer or estimate by the net. During the propagation through the RMS ¼ t ðGPS 2008i  Prediction 2008i Þ2 : (21)
net, all weights are fixed. During the backward step, all weights are n  1 i¼1
adjusted according to a rule of correction of errors (see, e.g. Kröse
and Smagt, 1996; Haykin, 1999; Arbib, 2003). Specifically, the
answer of the net is subtracted from that desired to produce the (4) The coefficient of determinant R2 was computed to assess the
error signal. This signal is propagated back through the net, hence fit of the predicted value to the true shoreline position using
the name back propagation. The weights are adjusted to make the (see, e.g., Schaible and Lee, 1996).
answer of the real net closer to the desired output (see, e.g. Kröse
and Smagt, 1996; Haykin, 1999; Arbib, 2003). Pn
In practice, the use of the back propagation algorithm tends to ðGPS 2008i  Predictioni Þ2
R2 ¼ 1  Pn i ¼ 1 2
(22)
converge very slowly, requiring a great computational effort. To
i ¼ 1 ðGPS 2008i  Mean GPS 2008Þ
solve this problem, several techniques have been incorporated to
improve the performance in order to reduce the convergence time
(5) Finally, we determine the mean absolute percentage error
and the computational effort required. Such techniques include,
(MAPE), which measures the accuracy of a fitted time series
e.g., the LevenbergeMarquardt algorithm (see, e.g. Hagan and
and usually expresses accuracy as a percentage (see, e.g., Hayati
Menhaj, 1994; Lera and Pinzolas, 2002).
and Shirvany, 2007).
A test using neural networks is done with the characteristic of
non-supervised training. The inputs are the vectors of observations Table 1
y described in equation (2). The network learns the relationships Artificial neural network settings.
provided by the input information and finds out what the output Architecture e Multilayer Feed-Forward Networks e Structure
for the forward step representing the 2008 predictions are. The
Artificial neural network MLP
data for the year 2008 does not participate in the training. Similarly Training method LevenbergeMarquardt(LM)
to the linear regression and robust estimation methods, the 2008 Number of hidden neuron 1
data serves only as a control to verify the answers found by the Number of hidden layer 4
Activation function used in hidden layer Hyperbolic tangent
respective predictive model. It is emphasized that any kind of
Activation function used in output layer Linear
weight is selected as the input data for the neural network case.
108 R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110

n  
Table 2 100 X 
GPS 2008i  Prediction 2008i 
MAPE ¼ (23)
Cross-sections characteristics.
n i ¼ 1 GPS 2008i 
Cross-sections Temporal data belonging to the set Degrees of
freedom
10 1963 2001 2002 2005 2
15 1963 1980 2001 2002 2005 3
10 1954 1963 1980 2001 2002 2005 4 2.6. The 2008 predictions
7 1954 1963 1980 1991 2001 2002 2005 5
9 1954 1963 1980 1991 1997 2001 2002 2005 6
All historical data are placed in layers and once this is done, one
1 1963 1980 1997 2001 2002 2005 4
5 1963 1980 1997 2001 2002 3 is chosen as a shoreline reference upon which the cross-sections
Total ¼ 57 are drawn. In this case, the first line obtained by GPS (2001) was
selected as a reference, since it comprised the limits of extension of

Table 3
Results of the prediction models for the 2008 shoreline position compared to that of the GPS survey in 2008 (GPS 2008-model results).

Sample number Number of iteration for Robust (GPS2008  Prediction) Linear regression (GPS2008  Prediction) Artificial neural network
robust best solution (GPS2008  Prediction)

Absolute value (m) Absolute value (m) Absolute value (m)


1 12 0.85 14.35 3.54
2 13 1.36 14.54 3.89
3 28 1.01 8.94 2.06
4 1 1.31 1.31 2.33
5 14 0.9 5.29 1.48
6 3 0.93 6.48 1.35
7 1 0.29 0.29 0.19
8 20 0.47 7.58 0.03
9 2 2.13 2.20 4.62
10 2 0.45 4.75 3.30
11 27 2.24 12.84 4.94
12 2 0.68 6.54 0.34
13 2 0.81 6.62 0.31
14 2 1.58 3.97 0.88
15 2 0.23 5.23 1.05
16 2 1.76 4.04 0.68
17 2 0.23 5.35 1.41
18 1 1.61 1.61 0.73
19 1 0.79 0.79 1.29
20 1 1.60 1.60 0.12
21 1 0.78 0.78 0.79
22 1 1.24 1.24 1.04
23 1 0.32 0.32 0.91
24 1 1.29 1.29 1.96
25 1 1.24 1.24 1.78
26 2 0.06 3.12 7.86
27 21 0.50 1.85 3.98
28 4 0.45 3.08 3.84
29 4 0.70 3.28 2.66
30 3 0.12 0.36 4.28
31 3 0.69 1.44 5.38
32 2 1.18 3.70 0.54
33 2 3.01 8.57 2.61
34 2 2.15 9.22 0.25
35 5 0.30 28.74 2.62
36 7 6.00 57.29 12.09
37 19 0.36 47.36 14.15
38 3 0.36 13.68 16.46
39 1 1.57 1.57 2.32
40 3 0.07 5.37 1.88
41 1 4.21 4.21 2.83
42 3 1.46 4.51 2.30
43 1 4.77 4.77 4.96
44 5 0.01 1.99 1.26
45 1 6.73 6.73 11.81
46 1 6.21 6.21 11.96
47 1 0.73 0.73 3.43
48 3 0.28 3.23 1.18
49 2 1.97 2.10 3.14
50 2 0.28 1.36 2.35
51 2 0.68 3.28 3.05
52 1 7.63 7.63 13.57
53 2 2.36 2.62 12.51
54 1 0.27 0.27 4.35
55 3 0.04 1.77 2.45
56 1 1.20 1.20 1.52
57 7 6.20 7.00 3.33
R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110 109

the coastline studied. From it, cross-sections with a distance of Table 4


100 m were drawn. Statistical data.

In practice there are 57 transverse lines and Table 2 presents the Statistics results Robust Linear Neural
respective information; number of cross-sections, temporal infor- model (m) regression (m) network (m)
mation of the data, and the degrees of freedom relative to the linear Maximum 7.63 57.29 16.46
mathematical model, i.e., the difference between variables and the Minimum 0.01 0.27 0.03
Mean 1.56 6.45 3.69
number of independent equations.
Standard 1.83 10.1 4.03
For the weight matrix in equation (10), the photogrammetry deviation
data are assigned a standard error of 16 m, i.e. 2s following Crowell RMS 2.39 11.9 5.36
et al. (1991) who suggests an error of about s ¼ 8 m. GPS measuring R2
0.994 0.984 0.997
accuracy of 2 cm is adopted for the kinematic relative GPS data (see, MAPE 0.14 0.61 0.33

e.g., Awange, 2012). With this information, the vector of the


unknowns comprising the linear regression parameters a and c is observation from the rest of the sample used in the estimation. This
estimated. The established linear regression models are then used way, it avoids the trap of falls rejection and falls retention (see
to predict the shoreline position for the year 2008. This procedure is Hampel et al. 1986).
repeated with the robust estimation and the neural network For the artificial neural network, the choice of the architecture of
methods. The results of the predicted shoreline for 2008 from all the training method and the neurons in the hidden layers are
the three methods under study are compared with that of the 2008 important factors influencing the response of the predictive model.
GPS survey. In this regard, the results of artificial neural network in Table 3 and
Fig. 4 show significant variations when the neurons in the hidden
3. Results and discussion layer are modified. For example, performing a test with 2 and 10
neurons, maximum deviations of 126.04 m and 242.08 m were
Table 3 shows the results for the three prediction methods: obtained, respectively, and with 4 neurons in the hidden layers, the
linear regression, robust estimation and neural networks, result was 16.4 m.
compared with the true values based on the 2008 GPS survey used The values listed in Table 4 were calculated using equations
as the control for the set of 57 samples used in the study. The (19)e(23), and include the minimum and maximum values for the
second column of Table 3 shows the number of iterations per- data. The results are (arithmetic mean, standard deviation, RMS, R2
formed to obtain the best solution of the robust estimation. When and MAPE) 6.45 m, 10.10 m, 11.90 m, 98.44% and 0.61% for linear
the number of iterations is 1, the solution found by the robust regression; 1.56 m, 1.83 m, 2.39 m, 99.94% and 0.14% for the robust
estimation method is essentially the same as that of the linear estimation; and 3.69 m, 4.03 m, 5.36 m, 99.68% and 0.33% for the
regression. As can be seen from the results of Table 3, Robust neural networks.
estimation method has the smallest values of the absolute errors, The linear regression results were more than 8 m in comparison
i.e., GPS position of 2008 e the predicted model positions. This is to the control shoreline position in 10 test cases, where the
visible in Fig. 4, which presents the deviation of the predicted 2008 maximum deviation was 57.29 m. The best statistic result was
shoreline position from the GPS control data of 2008 constructed presented by the robust estimation model with the largest devia-
from the data in Table 3. The absolute errors are indicated in y-axis tion from the control shoreline being 7.63 m. The artificial neural
(m) and the samples appear on the x-axis. network results had a mean value of 3.69 m with the best coeffi-
In Fig. 4, the results of the 37th sample are seen to have the cient of determinant R2 value but a maximum deviation of 16.46 m.
largest absolute error as clearly depicted by the results of the
Robust estimation method. This sample could possibly be an outlier 4. Conclusion
due to the uncertainty in its data. The advantage of the Robust
estimation method is that despite this uncertainty in the data (i.e., In this work, historical photogrammetric data describing the
outlier), it simply isolates the influence of this particular positional variation of the shoreline in the municipality of Matinhos
in Brazil was combined with GPS-based observations to compare
the predictive capability of three models; linear regression, robust
Absolute Errors (GPS2008 - Prediction2008)
estimation, and artificial neural network. This was achieved by
60 comparing the short-term model predicted shoreline position for
Robust Model
the year 2008 to the GPS measured position for the same period. In
50 Linear Regression so doing, temporal and prediction analysis of shoreline was used to
Artificial Neural generate digital maps of shoreline and critical erosion sites, thus
40 Network providing scientific data, thereby integrating coastal spatial infor-
mation that have been used to support decision making processes.
Errors (m)

The results indicate the importance of continuous monitoring, thus


30
the Federal University of Paraná must keep monitoring the area.
One of the difficulties that deserve a special mention, however, is
20 the recovery of historical data series. The preparation phase of such
data in practice takes a great deal of effort.
10 On the one hand, the result of this comparison demonstrated
the efficiency of the robust estimation model, with residuals of less
0 than 8 m. The results of the linear regression model in some cases
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 were unexpected and completely misrepresent reality, showing
residuals of more than 10 m in seven cases. The artificial neural
Samples
network provides an attractive alternative model for prediction to
Fig. 4. Deviations of the predicted 2008 shoreline positions from the actual measured the robust estimation model with residuals less than 16.5 m. On the
GPS position in 2008. other hand, this study demonstrated that, despite the uncertainties
110 R.M. Goncalves et al. / Ocean & Coastal Management 69 (2012) 102e110

in the data, as often the case for shoreline predictions; the robust Fenster, M.S., Dolan, R., Morton, R.A., 2000. Coastal storms and shoreline change:
signal or noise? Journal of Coastal Research 17 (3), 714e720.
estimation method offers the possibility of appropriately dealing
Galgano, F.A., Douglas, B.C., Leatherman, S.P., 1998. Trends and variability of
with such uncertainties. Its attractive feature is that users do not shoreline position. Journal of Coastal Research 26, 282e291.
need to delete any portion of the data that appears to be outliers. Galgano, F.A., Douglas, B.C., 2000. Shoreline position prediction: methods and
Robust procedures are, therefore, capable of the following; errors. Environmental Geosciences 7 (1), 1e10.
Goncalves, R.M. Short-term trend modeling of the shoreline through geodetic data
using linear regression, robust estimation and artificial neural networks, 2010.
1 providing a method of managing outliers in shoreline predic- Ph.D. thesis. Geodetic Sciences Post-graduate Program, Federal University of
tion data, and Parana (UFPR), Curitiba, Brazil, p. 152.
Gorman, L., Morang, A., Larson, R., 1998. Monitoring the coastal environment; part
2 providing a diagnosis and control tool that may be used to IV: mapping, shoreline changes, and bathymetric analysis. Journal of Coastal
check the efficiency of traditional shoreline prediction Research 14 (1), 61e92.
methods, such as the linear regression. Hagan, M.T., Menhaj, M.B., 1994. Training feedforward networks with the Mar-
quardt algorithm. IEE Transactions on Neural Networks 5 (6), 989e993.
Hampel, F.R., Ronchetti, E.M., Rousseeuw, P.J., Stahel, W.A., 1986. Robust Statistics:
Acknowledgments the Approach Based on Influence Functions. Wiley, New York.
Hayati, M., Shirvany, Y., 2007. Artificial neural network approach for short term load
forecasting for Illam region. World Academy of Science. Engineering and
The authors are grateful to the three anonymous reviewers and Technology 28, 280e284.
the Editor-in-Chief Prof. Victor N de Jonge for their comments, Haykin, S., 1999. Neural Networks e A Comprehensive Foundation. McMaster
which helped to improve the quality of this manuscript. RMG and University Hamilton, Ontario, Canada, Pearson Education.
Hecky, R.E., Newbury, R.W., Bodaly, R.A., Patalas, K., Rosenberg, D.M., 1984.
CPK acknowledge the support of PROBRAL (CAPES/DAAD) and Environmental impact prediction and assessment: the southern Indian lake
CNPq; while JLA acknowledges the financial support of the Alex- experience. Canadian Journal of Fisheries and Aquatic Sciences 41 (4),
ander von Humboldt (Ludwig Leichhardt Memorial Fellowship) 720e732.
Huber, P.J., 1964. Robust estimation of a location parameter. Annals of Mahtematical
and Curtin Research Fellowship. RMG, JLA, and CPK would also like
Statistics 35, 73e101.
to thank their host Prof. Bernhard Heck and his team at the Huber, P.J., 1981. Robust Statistics. John Wiley & Sons, New York.
Geodetic Institute, Karlsruhe Institute of Technology, for providing Kröse, B., Smagt, P.V.D., 1996. An Introduction to Neural Networks, eighth ed. The
University of Amsterddam.
a conducive working atmosphere. The authors thank K. Fleming of
Krueger, C.P., Goncalves, R.M., Heck, B., 2009. Surveys at the coast of Parana, Brazil,
GfZ for assistance with the English but take full responsibility. This to determinate the temporal coastal changes. Journal of Coastal Research 1,
is a TiGER Publication No. 422. 632e635.
Lera, G., Pinzolas, M., 2002. Neighborhood based Levenberg-Marquardt algorithm
References for neural network training. IEE Transactions on Neural Networks 13 (5), 1200e
1203.
Addo, K.A., Walkden, M., Mills, J.P., 2008. Detection, measurement and prediction of Li, R., Di, K., Ma, R., 2001. A comparative study of shoreline mapping techniques. In:
shoreline recession in Accra, Ghana. ISPRS Journal of Photogrammetry & The 4th International Symposium on Computer Mapping and GIS for Coastal
Remote Sensing 63, 543e558. Zone Management, Nova Scotia.
Aduol, F.W.O., 1994. Robust geodetic parameter estimation through iterative Li, R., Keong, C.W., Ramcharan, E., Kjerfve, B., Willis, D., 1998. A costal GIS for
weighting. Survey Review 32 (252), 359e367. shoreline monitoring and management e case study in Malaysia. Surveying and
Álvarez-Ellacuría, A., Orfila, L., Gómez-Pujol, G., Simarro, N. Obregon, 2011. Decou- Land Information Systems 58 (3), 157e166.
pling spatial and temporal patterns in short-term beach shoreline response to Metropolitan Borough of Sefton, 2002. Shoreline monitoring annual report 2001/
wave climate. Geomorphology 128 (3e4), 199e208. 2002. http://www.sefton.gov.uk/pdf/TS_cdef_monitor_20012.pdf (accessed
Arbib, M.A., 2003. The Handbook of Brain Theory and Neural Networks. Massa- 14.11.08.).
chusetts Institute of Technology. Mitishita, E.A., Kirchner F.F. Digital mono-diferential restitution of airphotos applied
Awange, J.L., Aduol, F.W.O., 1999. An evaluation of some robust estimation techniques to planimetric mapping. In: International Archieves of Photogrammetry and
in the estimation of geodetic parameters. Survey Review 35 (273), 146e162. Remote Sensing, vol. XXXIII, Part B4. Amsterdam 2000, pp. 655e662.
Awange, J.L., 2012. Environmental Monitoring Using GNSS. Springer, Schaible, B., Lee, Y.C., 1996. Fuzzy logic models with improved accuracy and
Berlin-New York. continuous differentiability. IEEE Transactions on Components, Packaging, and
Boak, E.H., Turner, I.L., 2005. Shoreline definition and detection: a Review. Journal of Manufacturing Technology 19 (1), 37e47.
Coastal Research 21 (4), 688e703. Soares, C.R., Vobel, I., Paranhos Filho, A.C. 1995. The marine erosion problem in
Crowell, M., Letherman, S.P., Buckley, M.K., 1991. Historical shoreline change: error Matinhos municipality. In: Land Ocean Interactions on the Coastal Zone, 1995,
analysis and mapping accuracy. Journal of Coastal Research 7 (3), 839e852. São Paulo, pp. 48e50.
Crowell, M., Douglas, B.C., Leatherman, S.P., 1997. On forecasting future U.S. shore- Souza, C.R.G., 2009. Coastal erosion and the coastal zone management challenges in
line positions: a test of algorithms. Journal of Coastal Research 13 (4), 1245e Brazil. Journal of Integrated Coastal Zone Management 9 (1), 17e37.
1255. Thieler, E.R., Pilkey Jr., O.H., Young, R.S., Bush, D.M., Chai, F., 2000. Behavior for
Demarest, J.M., Leatherman, S.P., 1985. Mainland influence on coastal transgression: U.S. coastal engineering: a critical review. Journal of Coastal Research 16 (1),
Delmarva Peninsula. Marine Geology 63, 19e33. 48e70.
Douglas, B.C., Crowell, M., Leatherman, S.P., 1998. Considerations for shoreline Tzatzanis, M., Wrbka, T., Sauberer, N., 2003. Landscape and vegetation responses to
position prediction. Journal of Coastal Research 14 (3), 1025e1033. human impact in sandy coasts of Western Crete, Greece. Journal for Nature
Douglas, B.C., Crowell, M., 2000. Long-term shoreline position prediction and error Conservation 11 (3), 187e195.
propagation. Journal of Coastal Research 16 (1), 145e152. Veloso-Gomes, A., Barroco, A.R., Pereira, C.S., Reis, H., Calado, J.G., Ferreira, M.D.C.,
Fenster, M.S., Dolan, R., Elder, J.F., 1993. New method for predicting shoreline Freitas, M. Biscoito, 2008. Basis for a national strategy for integrated coastal
positions from historical data. Journal of Coastal Research 9 (1), 147e171. zone management e in Portugal. Jornal of Coast Conservatin 12, 3e9.

You might also like