|
Support Vector classifiers for Land Cover Classification
 Mahesh Pal Lecturer, Department of Civil Engineering National Institute of Technology Kurukshetra 136119 Haryana (India) Email: mpce_pal@yahoo.co.uk Fax No. 01744 38050  Paul M Mather Prof., School of geography University of Nottingham University Park Nottingham, NG7 2RD, UK Email: paul.mather@nottingham.ac.uk
Abstract
Much research effort in the past ten years has been devoted to analysis of the performance of artificial neural networks in image classification (Benediktsson et al., 1990; Heermann and Khazenie, 1992). The preferred algorithm is feed-forward multi-layer perceptron using back-propagation, due to its ability to handle any kind of numerical data, and to its freedom from distributional assumptions. Although neural networks may generally be used to classify data at least as accurately as statistical classification approaches a number of studies have reported that users of neural classifiers have problems in setting the choice of various parameters during training (Wilkinson, 1997). The choice of architecture of the network, the sample size for training, learning algorithms, and number of iterations required for training are some of these problems. A new classification system based on statistical learning theory (Vapnik, 1995), called the support vector machine has recently been applied to the problem of remote sensing data classification (Huang et al., 2002; Zhu and Blumberg, 2002; Gualtieri and Cromp, 1998). This technique is said to be independent of the dimensionality of feature space as the main idea behind this classification technique is to separate the classes with a surface that maximise the margin between them, using boundary pixels to create the decision surface. The data points that are closest to the hyperplane are termed "support vectors". The number of support vectors is thus small as they are points close to the class boundaries (Vapnik, 1995). One major advantage of support vector classifiers is the use of quadratic programming, which provides global minima only. The absence of local minima is a significant difference from the neural network classifiers. Support vector classifiers are primarily designed for two class problems only. A number of methods are suggested to create multi-class classifiers using two-class methods (Hsu and Lin, 2002). For this study, a multi-class support vector classifier employing one against one approach of creating multi-class classifier is used for land cover classification using ETM+ multispectral and DAIS hyperspectral data. The two study areas used in the work reported here are located near the town of Littleport in UK and La Mancha region in Spain, respectively. For the Littleport area, ETM+ data acquired on 19th June, 2000, are used. The classification problem involves the identification of seven land cover types (wheat, potato, sugar beet, onion, peas, lettuce and beans). For the La Mancha area, hyperspectral data acquired on 29th June 2000 by the DAIS 7915 airborne imaging spectrometer were used. Eight different land cover types (wheat, water body, dry salt lake, hydrophytic vegetation, vineyards, bare soil, pasture lands and built up area) were used. Like neural network classifiers the performance of support vector classifier depends on some user defined parameters such as kernel type used to transform the data into high dimensional feature space, kernel specific parameters and the parameter C used in non-separable data case so as to control a trade-off between margin between separating hyperplanes and misclassification error. For this study a radial basis kernel with g (kernel specific parameter) value as two and C = 5000 is used for both data sets. Random sampling was used to collect the training and test pixels for both ETM+ and DAIS data set. Total selected pixels was divided in two parts using one as training and other as test data so as to remove any bias caused by using same pixels for training and testing the classifiers. A total of 2700 training pixels and 2037 test pixels for ETM+ data and a total of 1600 training (200 pixels/class) and 3800 test pixels were used for DAIS data. To avoid the situation in which the observed results may be classifier dependent, the data set used in the support vector experiment was inputted to both maximum likelihood (ML) and neural network (NN) classifiers. For this study, a standard back-propagation neural classifier was used. All user-defined parameters are set as recommended by Kavzoglu (2001). Results obtained using ETM+ data suggests that support vector classifier perform well in comparison with neural and statistical classifier. An accuracy of 87.9% ( 0.87) is achieved with support vector classifier in comparison with an accuracy of 85.1% (0.83) with NN and 82.9% (0.80) with ML classifier, while value in bracket indicates the Kappa value. The training time taken by support vector classifier is 0.30 minutes as compared to that of 58 minutes on a dual processor sun machine by NN classifier. These results suggest that support vector classifier performance is statistically significant in comparison with NN and ML classifiers and can be trained quickly. To study the behaviour of support vector classifier with DAIS Hyperspectral data a total of sixty five features (bands) was used, discarding seven features with severe striping out of the seventy two features. Beginning with five bands, an additional five bands were added at each cycle, thus generating thirteen accuracy values for the data set. Accuracy obtained with different number of features suggests that the performance of a support vector classifier is good with a small number of training data in comparison with ML and NN classifiers. Results also suggests that classification accuracy using support vector classifier increases almost continuously, with a fixed number of training data and the increasing number of features suggesting that support vector classifier is not affected by Hughes (1968) phenomenon, whereas accuracies produced by ML and NN declines slightly when the number of bands exceeds 50 or so. |