RESEARCH ARTICLE
Adv. Sci. Lett. 24, 1382–1387, 2018
Copyright © 2016 American Scientific Publishers All rights reserved Printed in the United States of America
Advanced Science Letters Vol. 24, 1382–1387, 2018
Artificial Neural Networks for Satellite Image Classification of Shoreline Extraction for Land and Water Classes of the North West Coast of Peninsular Malaysia Syaifulnizam Abd Manaf1, Norwati Mustapha1, Md. Nasir Sulaiman1, Nor Azura Husin1, Mohd Radzi Abdul Hamid2 1Intelligent
Computing Research Group, Faculty of Computer Science and Information Technology, Universiti Putra Malaysia, 44300 UPM Serdang, Selangor, Malaysia 2Coastal Management and Oceanography Research Centre, National Hydraulic Research Institute of Malaysia, 43300 Seri Kembangan, Selangor, Malaysia
Monitoring and measuring the shoreline of coastal zones helps establish the boundary of a country. Such an activity entails ground survey, topographic survey, aerial photo, or remote sensing techniques to extract the shoreline. For example, the remote sensing technique to determine shorelines involves the extraction of relevant data from satellite images. Specifically, the satellite image classification enables shorelines to be extracted from land and water classes with a high degree of precision. However, extracting information from satellite images is challenging as it relies on a strong understanding of image processing, machine learning, and data mining techniques. Thus, the researchers discuss the study of the pixel-based classification of machine learning techniques to classify land and water classes in terms of accuracy, training time, and testing time. The research findings showed that the Multilayer Perceptron Artificial Neural Network (MLP ANN) was the most effective technique, compared with other techniques, hence reinforcing its importance in classifying land and water classes. Keywords: Multilayer Perceptron, Artificial Neural Networks, Image Classification, Shoreline Extraction, Machine Learning
1. INTRODUCTION Coastal zones are constantly exposed to natural processes, of which factors such as tidal effect, wind speed, wind direction, sea level, and sediment transport, are dynamically changing the geographic landscape of these regions1,2,3. Socio-economic activities, which triggers many uncontrolled anthropogenic processes, such as tourism, industrialization, sewage coastal pollution and agricultural pollution, further aggravates the problem4. In unison, both these natural and anthropogenic processes are continually reshaping and redefining the coastal areas of countries on a massive, unpredictable * Email Address:
[email protected] 1
Adv. Sci. Lett. Vol. 24, No. 2, 2018
scale. Thus, coastal zones monitoring provides important information about the condition of coastal areas, which are adversely affected by natural or human activities. In fact, coastal zone monitoring is part of a nation’s important endeavors to protect its sovereignty. Essentially, a coastal zone can be monitored by examining changes that are taking place at its shoreline. A shoreline is an interface that physically separates the land and the water, thus creating a boundary between them5,6. In this regard, the extraction of shoreline is useful for several applications that deal with the examination of historical records of climatic changes, detection of coastal changes, and management of coastal zones. Measuring a shoreline using traditional ground survey techniques is difficult, 1936-6612/2011/4/400/008
doi: 10.1166/asl.2018.10754
Adv. Sci. Lett. 24, 1382–1387, 2018 time-consuming, and, in extreme cases involving complex coastal terrains, almost impossible. Irrespective of the nature of the coastal areas, acquiring a shoreline entails a shoreline indicator that represents the true position of the shoreline to be determined. In general, the shoreline indicators can be categorized into three groups: (i) an indicator that is based on visually detectable (or discernible) features, (ii) an indicator that is based on specific tidal datum, and (iii) an indicator that is based on features that are not necessarily visible to the human eye6. In this study, the researchers used the third group as the shoreline indicator. To date, many techniques have been employed to extract shorelines from optical multispectral satellite images using image processing techniques and image classification techniques. The image processing techniques include Band rationing7,8, edge detection9, normalized difference water index (NDWI)10, segmentation11,12, thresholding13, and wavelet14. The satellite image classification techniques based on the machine learning techniques can be divided into two types of classifications: (i) supervised classification, such as Maximum Likelihood15,16, Mahalanobis Distance16, Minimum Distance16, Neural Network15 and Support Vector Machines15,17 and (ii) unsupervised classification, such as ISODATA16. The main aim of this study is to evaluate the most effective and efficient machine learning technique for the extraction of shoreline of the North West coast of Peninsular Malaysia based on the Landsat TM satellite images. Currently, there are 11 machine learning techniques that can be used to classify land-water classes that helps extract a shoreline. For this study, the researchers used the pixel-based approaches to classify land-water classes. Specifically, the Multilayer Perceptron Artificial Neural Network (MLP ANN), and ten other machine learning techniques were chosen to perform the classification process. To aid discussion, this paper is structured as follows: Section 2 discusses the MLP ANN, Section 3 details the methodology used, Section 4 reports the experimental results of the proposed machine learning technique, and Section 5 highlights the discussion and conclusion of the paper.
RESEARCH ARTICLE among nodes carries an associated weight, and each node computes the weighted sum of the inputs and passes the sum through an activation function that provides the output value of this node19. For image classification, the input layer represents the image used, with each node corresponds to one band of the image. The hidden layers are used for computation, and the number of layers and nodes are set by the researchers. The output layer represents the classification results, in which each node corresponds to one class18. In this study, the researchers adopted a four-layer network structure consisting of one input layer, two hidden layers, and one output layer, as shown in Figure 1. For the input layer, three nodes were used as inputs from the 5-4-3 band combination of Landsat 5 TM image. For the hidden layers, five and three nodes were used for the first and second hidden layers, respectively. For the output layer, two output nodes were used to represent the final classes.
Fig.1. The four-layer network structure of ANN used in this study 3. MATERIALS AND METHODS In this study, the method used to extract information was based on the satellite image classification approach, which consists of four phases. The four phases are preprocessing, classification, evaluation and post-processing, as shown in Figure 2.
2. ARTIFICIAL NEURAL NETWORKS Artificial Neural Network (ANN) is the most common approach of nonparametric classification that does not rely on the statistical frequency distribution of data. There are many types of ANNs, with each having specific characteristics that are based on its architecture, training or learning algorithm, and activation function. Among numerous types of neural networks, the most widely used algorithm in satellite image classification is the architecture of MLP, which consists of one input layer, at least one hidden layer, and one output layer18. Each layer consists of several neuron nodes, and all the nodes of a layer are connected to the nodes of the adjacent layers. However, there are no interconnections between nodes of the same layer. Each interconnection
Fig.2. The four phases of the extraction method
2
RESEARCH ARTICLE
Adv. Sci. Lett. 24, 1382–1387, 2018
Study Area The chosen study area was in the Langkawi Island, which located at the North West coast of Peninsular Malaysia, as shown in Figure 3. Specifically, this island is located between 6o 15’N and 6o 29’N latitude and 99o 37’E and 99o 57’E longitude. The total area of Langkawi Island is about 47,848 ha. This island is a major tourist attraction since 1987, the year in which it acquired the duty-free status20. The island comprises many small islands, however only the main landmass was considered in this study (see Fig. 3).
which represents the background signature of the band. Mosaicking process was then applied to combine two scenes of an image of the study area into a single large image using the same coordinate system. Then, image registration process was performed using image-to-image geometric correction process with RMS value of 0.457. Later, Rectified Skew Orthomorphic (RSO) Kertau was chosen to project the image data onto a local projection system for the West Coast of Peninsular Malaysia. Finally, the image sub-setting was performed to fit the image to the area of study. Classification In this phase, supervised classification approaches were used to classify land and water classes of satellite images. Training set was created to build the model, while testing set was produced to measure its performance. For this study, only one set of training and testing was created in the form of polygons as depicted in Figure 4. There were 260 and 65 polygons generated for land and water classes, respectivley. To ensure separability of the training and testing set, Jeffries-Matusita distance and Transformed Divergence were used22, yielding a separability index of 1.97, which was close to 2.0 that represents perfect separability.
Fig.3. The study area of the research Data The data used in this research consisted of two scenes of multispectral Landsat-5 Thematic Mapper (TM) data. Two scenes, instead of one, were used to ensure the whole study area would be covered sufficiently. The scenes were acquired on two different days as shown in Table 1. Table.1. The description of the satellite image data. DATA AC DATE AC TIME PATH/ROW Landsat TM 24/12/2010 03:29:49 129/56 Landast TM 05/08/2011 03:28:41 128/56 Pre-processing The main aim of the pre-processing phase was to clean satellite images from atmospheric, radiometric, or geometric errors caused by satellite sensors. Once cleaned, the images were used in the classification phase. After the study area and image data were established, radiometric calibration21 on the image was applied. This process helped calibrate the satellite image to radiance (Lλ), as expressed by Equation (1). Lλ= Gain * Pixel value + Offset
(1)
The pixel value ranges from “0” to “255”, and the radiance for each image band depends on the gain and offset values. After calibrating the image data, dark subtraction method was applied to remove atmospheric scattering effects of the image data. This process used the minimum value of a band, 3
Fig.4. Training and testing set Arguably, in view of the high number of techniques available, choosing the right machine learning classifier for this domain problem was a challenging task. For this study, the researchers considered 11 machine learning techniques, namely Decision Tree (DT), Naïve Bayes (NB), k-Neareast Neighbour (kNN), Linear Discriminat Analysis (LDA), Quadratic Discriminant Analysis (QDA), Logistic Regression (LR), SGD Classifier, SVM-linear, SVM-RBF, SVMpolynomial, and MLP ANN. After the model had learned part of image regions from the training samples, it then classified them as a whole image. In addition, a lookup table (LUT) consisting of all color classes was used to render the classified image with appropriate colors.
Adv. Sci. Lett. 24, 1382–1387, 2018 Evaluation The assessment for accuracy of the satellite image classification was based on the overall accuracy, training time, and testing time. In particular, the overall accuracy was used as the primary performance indicator because it is widely used in the evaluation of satellite image classification methods. By contrast, the other two factors were treated as secondary or additional performance indicators in this study. K-fold cross-validation method was used to avoid overfitting. Essentially, the underlying mechanism of kfold cross-validation is that it randomly partitions an original sample into k subsamples of equal size. A single subsample is used to test the model as validation data, while the remaining (k-1) subsamples are used as training data23,24. Thus, in this study, the 10-fold cross-validation method was used to deal with overfitting. Post-processing In the last phase, the classified image was saved to a hard drive. Then, smoothing processes (e.g., sieve and clump) were performed to generalize the image data, such that the resultant polygons would be smooth, not curvy. Then, the resultant classified image underwent a conversion process, through which the original raster image was converted to a vector image. Finally, the polygon-to-line conversion process was carried out to ensure the image would be consisted of smooth lines, rather than curvy ones. 4. EXPERIMENTAL RESULT The analysis of the experimental data was carried out on a high-performance workstation, namely one 3620, 3.4GHz Intel i7-6700 Quad Core Processor, 16 GB RAM Dell Precision machine, running on Microsoft Windows 7, 64-bit operating system. Table 2 summarizes the results of image classification of the Landsat TM images. Evidently, MLP ANN was the most effective method for this problem domain, securing the highest overall accuracy at 99.55%. Equally impressive was LDA, which at 99.38% of the overall accuracy, came a close second. In contrast, QDA achieved 98.58% of overall accuracy, relegating it to the last place for this performance measure. However, LDA outperformed MLP ANN in terms of training and testing times, with the former spending 0.06 seconds and 0.534 seconds, respectively, and the latter 3.512 seconds and 4.067 seconds, respectively. For training time, QDA recorded the shortest time, which at 0.042 seconds was impressively fast, thus making it the most efficient method for this measure. In stark contrast, SVM-polynomial was the least efficient method, as attested by its dismaying training time of 26.890 seconds. For testing time, LDA recorded 0.534 seconds, which was the shortest testing time, thus establishing it as the best
RESEARCH ARTICLE method for this measure. Again, SVM-polynomial fared relatively sluggish, spending 28.211 seconds in testing, thus making it as the least efficient method. The differences in training and testing times between SVM-RBF and k-NN were 11.109 seconds and 7.832 seconds, respectively, which were highly significant. However, no such significant differences were detected among SGD Classifier, LDA and DT, which recorded differences of 0.473 seconds, 0.474 seconds, and 0.478 seconds, respectively. Table.2. Experimental results of satellite image classification. METHOD OVAC(%) TRAIN(s) TEST(s) DT 99.28 0.098 0.576 NB 98.75 0.051 0.713 k-NN 99.21 0.378 8.210 LDA 99.30 0.060 0.534 QDA 98.58 0.650 0.042 LR 99.36 0.256 0.750 SGD Classifier 99.01 0.125 0.598 SVM-linear 99.31 1.032 3.311 SVM-RBF 99.08 8.758 19.867 SVM-polynomial 98.91 26.890 28.211 MLP ANN 3.512 4.067 99.55 For the MLP ANN, the experiments were conducted using three types of activation functions, namely rectified linear unit (relu), logistic sigmoid function (logistic), and hyperbolic tan function (tanh). However, only relu activation function is discussed in this paper given its excellent results, as compared with the other two functions. For these functions, three algorithms for weight optimization were considered, namely an optimization algorithm in the family of quasi-Newton methods (l-bfgs), stochastic gradient descent (sgd), and a stochastic gradientbased optimization algorithm as proposed by Kingma, Diederik, and Jimmy Ba (adam). Table 3 summarizes the experimental results of the proposed MLP ANN algorithm for this problem. For the hidden layers, the experiments were conducted on 1-to-3 hidden layers, with each layer consisting of several nodes (ranging from one to 10 nodes). However, only the three-hidden-layer configurations are discussed in this paper given their excellent results, as compared with the others. The hidden layer configurations were considered based on the three aforementioned weight optimization algorithms, such as one hidden layer with four nodes, two hidden layers with 5 nodes plus 3 nodes, and two hidden layers with 5 nodes plus 2 nodes. Clearly, MLP ANN of one input layer and two hidden layers was the most effective and efficient method, recording 99.58% of overall accuracy, 1.732 seconds for training time, and 2.329 seconds for testing time. The configuration of the network used five and three nodes for the first and second hidden layers, respectively. The hidden layers’ configuration used relu activation 4
RESEARCH ARTICLE
Adv. Sci. Lett. 24, 1382–1387, 2018
function for hidden layer and l-bfgs algorithm for weight optimization. This algorithm yielded the fastest training time and testing time, recording 1.732 seconds and 2.329 seconds, respectively. Interestingly, MLP of a hidden layer with four nodes and sgd optimization algorithm achieved 99.58% of the overall accuracy, the performance of which was similar with the best method (i.e., MLP 5n+3n (relu, lbfgs)). However, the former performed poorly in terms of training and testing time, as it took substantially longer time than the latter (see Table 3.) Table.3. Experimental results of proposed algorithm of satellite images classification. METHOD
OVAC(%) TRAIN(s) TEST(s)
MLP 4n (relu, l-bfgs) MLP 4n (relu, sgd) MLP 4n (relu, adam) MLP 5n+2n (relu, l-bfgs) MLP 5n+2n (relu, sgd) MLP 5n+2n (relu, adam) MLP 5n+3n (relu, l-bfgs) MLP 5n+3n (relu, sgd) MLP 5n+3n (relu, adam)
99.54 99.58 99.36 99.55 99.57 99.36 99.58 99.57 99.56
0.991 18.256 12.995 3.512 22.991 24.465 1.732 10.017 8.363
1.534 18.785 13.561 4.067 23.533 25.018 2.329 10.615 8.926
Fig.6. Shoreline output Finally, the Landsat image was overlaid with the shoreline output to demarcate the boundary of the map, which could be used for further shoreline map analysis. Figure 7 depicts the shoreline, as seen in red, vividly outlining the original Landsat TM image.
The classification process produced the image classification map as shown in Figure 5, which clearly highlights the differentiation between the land class (in brown) and the water class (in blue).
Fig.7. Shoreline map 5. CONCLUSIONS
Fig.5. Classification result Later, post-classification processes were performed on the image classification map to extract the shoreline from the satellite image. Figure 6 shows the shoreline output after the completion of the extraction processes.
5
In this study, a series of experiments involving 11 machine-learning techniques was carried out. The results showed that MLP ANN was the most effective image classification method, having achieved the highest overall accuracy. However, this method fared moderately in terms of training time and testing time. Nonetheless, its performances in these measures could be improved by making appropriate adjustments to a network topology. More revealingly, the results indicated that LDA was the most optimal method for all the three measures (i.e., overall accuracy, training time, and testing time) — put simply, LDA was highly accurate and exceedingly fast. In contrast, SVM-polynomial was the least effective method among the methods tested. Taken together, the results suggest that choosing a particular image classification method for shoreline extraction depends on a number of factors, notably the
RESEARCH ARTICLE
Adv. Sci. Lett. 24, 1382–1387, 2018 network topology and performance parameters, such as overall accuracy, training time, and testing time. Furthermore, MLP ANN method depends on inputs, outputs and network training methods. This study could be extended by applying pan-sharpening process on satellite images to increase spatial resolution. In addition, such a study could include object-based classification involving assemble machine learning approach or deep learning neural network to further improve existing methods for shoreline extraction.
[12] [13]
[14]
ACKNOWLEDGMENTS We would like to thank the Ministry of Science, Technology and Innovation of Malaysia for the ScienceFund Research Grant (No: 01-01-04-SF2291), which funded this study; the Department of Survey and Mapping of Malaysia, the Malaysian Remote Sensing Agency, the Federal Department of Town and Country Planning Peninsular Malaysia, and the United States’s Geological Survey for providing us the images data; and Universiti Putra Malaysia for the logistical assistance rendered.
[15]
[16]
[17]
[18]
REFERENCES [1]
[2]
[3]
[4] [5] [6] [7] [8] [9]
[10]
[11]
MatAmin AR, Ahmad F, Mamat M, Rivaie M, Abdullah K. Sediment Variation along the East Coast of Peninsular Malaysia. Ecol Quest. 2012;16(1):99-107. doi:10.2478/v10090-012-0010-6. Mustaqim M, Mohd B, Ghazali ZZ. Wind Influence Towards Shoreline Movement At Pantai Sabak, Kelantan Darulnaim. In: NATIONAL GEOSCIENCE CONFERENCE 2014. ; 2014. geology.um.edu.my/gsmpublic/.../P083- Mustaqim paper.pdf. Mohamad MF, Lee LH, Samion MKH. Coastal Vulnerability Assessment towards Sustainable Management of Peninsular Malaysia Coastline. Int J Environ Sci Dev. 2014;5(6):533-538. doi:10.7763/IJESD.2014.V5.540. Zainora A, Ainaa NI. The Perception of Community on Coastal Erosion Issue in Selangor, Malaysia. J Clean Energy Technol. 2013;1(3):164-168. doi:10.7763/JOCET.2013.V1.38. Dolan, Robert, S. Fenster, Michael, and J. Holme S. Temporal Analysis of shoreline Recession and Accretion. J Coast Res. 1991;7(3):723-744. Boak EH, Turner IL. Shoreline Definition and Detection: A Review. J Coast Res. 2005;214:688-703. doi:10.2112/030071.1. Lira C, Taborda R. Advances in Applied Remote Sensing to Coastal Environments Using Free Satellite Imagery.; 2014. doi:10.1007/978-3-319-06326-3. Sarwar MGM, Woodroffe CD. Rates of shoreline change along the coast of Bangladesh. J Coast Conserv. 2013;17(3):515-526. doi:10.1007/s11852-013-0251-6. Zhang Y, Li X, Zhang J, Song D. A Study on Coastline Extraction and Its Trend Based on Remote Sensing Image Data Mining. Abstr Appl Anal. 2013;2013(Article ID 693194):6. doi:10.1155/2013/693194. Choung Y-J, Jo M-H. Shoreline change assessment for various types of coasts using multi-temporal Landsat imagery of the east coast of South Korea. Remote Sens Lett. 2016;7(1):91-100. doi:10.1080/2150704X.2015.1109157. Semenov SM, Abushenko NA, Chichigin AS. Discrimination
[19]
[20] [21]
[22]
[23] [24]
of Shorelines on Satellite Images from Boundary-Point and Halftone Information. Mapp Sci Remote Sens. 2016;36(4):245255. doi:10.1080/07493878.1999.10642126. Zhao L, Fan L, Wang C, Tang Y, Zhang B. A Non-Supervised Method For Shoreline Extraction U sing High Resolution SAR Image. 2012;(2009):317-322. Rigos A, Andreadis OP, Andreas M, Vousdoukas MI, Tsekouras GE, Velegrakis A. Shoreline Extraction from Coastal Images Using Chebyshev Polynomials and RBF Neural Networks. In: AIAI 2014. ; 2014:593-603. Yu S, Mou Y, Xu D, You X, Zhou L, Zeng W. A New Algorithm for Shoreline Extraction from Satellite Imagery with Nonseparable Wavelet and Level Set Method. Int J Mach Learn Comput. 2013;3(1):158-163. doi:10.7763/IJMLC.2013.V3.293. Rokni K, Ahmad A, Solaimani K, Hazini S. A new approach for surface water change detection: Integration of pixel level image fusion and image classification techniques. Int J Appl Earth Obs Geoinf. 2015;34(September):226-234. doi:10.1016/j.jag.2014.08.014. Sekovski I, Stecchi F, Mancini F, Del Rio L. Image classification methods applied to shoreline extraction on very high-resolution multispectral imagery. Int J Remote Sens. 2014;35(10):3556-3578. doi:10.1080/01431161.2014.907939. Yousef A, Iftekharuddin K. Shoreline extraction from the fusion of LiDAR DEM data and aerial images using mutual information and genetic algrithms. Proc Int Jt Conf Neural Networks. 2014:1007-1014. doi:10.1109/IJCNN.2014.6889863. Mustapha MR, Lim HS, Mat Jafri MZ. Comparison of Neural Network and Maximum Likelihood Approaches in Image Classification. J Appl Sci. 2010;10(22):2847-2854. Ojaghi S, Ebadi H, Ahmadi FF. Using artificial neural network for classification of high resolution remotely sensed images and assessment of its performance compared with statistical methods. Am J Eng Technol Soc. 2015;2(1):1-8. Samat N. Assessing Land Use Land Cover Changes in Langkawi Island : Towards Sustainable Urban Living. Malaysian J Environ Manag. 2010;11(1):48-57. Mao Z, Pan D, He X, et al. A Unified Algorithm for the Atmospheric Correction of Satellite Remote Sensing Data over Land and Ocean. Remote Sens. 2016;8(7):536. doi:10.3390/rs8070536. Bochow M. Improving Class Separability - A Comparative Study of Transformation Methods for the Hyperspectral Feature Space. In: Proceedings of 4th EARSeL Workshop on Imaging Spectroscopy. New Quality in Environmental Studies.; 2005:439-447. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.381. 9791. Kohavi R. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection. Int Jt Conf Artif Intell. 1995;14(12):1137-1143. doi:10.1067/mod.2000.109031. Bengio Y, Grandvalet Y. No Unbiased Estimator of the Variance of K-Fold Cross-Validation. J Mach Learn Res. 2004;5:1089-1105.
Received: 22 September 2016. Accepted: 18 October 2016
6
