|
Support Vector classifiers for Land Cover Classification
3. Data
The first of the two study areas used in the work reported here are located near the town of Littleport in eastern England. The second is a wetland area of the La Mancha region of Spain. For the Littleport area, ETM+ data acquired on 19th June 2000 is used. The classification problem involves the identification of seven land cover types (wheat, potato, sugar beet, onion, peas, lettuce and beans) for the ETM+ data set. For the La Mancha study area, hyperspectral data acquired on 29th June 2000 by the DAIS 7915 airborne imaging spectrometer were available. Eight different land cover types (wheat, water body, dry salt lake, hydrophytic vegetation, vineyards, bare soil, pasture lands and built up area) were specified. The DAIS data show moderate to severe striping problems in the optical infrared region between bands 41 and 72. Initially, the first 72 bands in the wavelength range 0.4 µm to 2.5 µm were selected. All of these bands were examined visually to determine the severity of striping. Seven bands displaying very severe striping problems (bands 41 - 42 and 68 - 72) were removed from the data set. The striping in the remaining bands was removed by automatically enhancing the Fourier transform of each image (Cannon et al., 1983; Srinivasan et al., 1988). The input image is first divided into overlapping 128-by-128-pixel blocks. The Fourier transform of each block is calculated and the log-magnitudes of each FFT block are then averaged. The averaging process removes all frequency domain quantities except those which are present in each block; i.e., some sort of periodic interference. The average power spectrum is then used as a filter to adjust the FFT of the entire image. When an inverse Fourier transform is performed, the result is an image with periodic noise eliminated or significantly reduced. 4. Result and discussions Random sampling was used to collect the training and test pixels for both ETM+ and DAIS data set. Total selected pixels were divided into two parts, one for training and one for testing the classifiers, so as to remove any possible bias resulting from the use of same set of pixels for both the testing and training phases. 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. Kappa values and overall classification accuracies are calculated for each of the classifiers used in this study with ETM+ data, while overall accuracy is calculated for the DAIS hyperspectral data. The Z statistic is also used to test the significance of apparent differences between the three classification algorithms, using ETM+ data. For this study, a standard back-propagation neural classifier was used. All user-defined parameters are set as recommended by Kavzoglu (2001). Like neural network classifiers the performance of support vector classifier depends on a number of user-defined parameters which may influence the final classification accuracy. For this study a radial basis kernel with g (kernel specific parameter) value as two and C = 5000 is used for both data sets. The values of these parameters were chosen after a number of trials and the same parameters are used with DAIS data set. This study also suggests that, in comparison with the NN classifier, it is easier to fix the values of the user defined parameters for SVM. As mentioned earlier, SVM involves in solving a quadratic programming problem with linear equality and inequality constraints which has only a global optimum. In comparison the presence of local minima is a significant problem in training the neural network classifiers. Results obtained using ETM+ data suggests that support vector classifier perform well in comparison with neural and statistical classifier (Tables 1 and 2).
Further, the training time taken by support vector classifier is 0.30 minutes in comparison of 58 minutes by the NN classifier on a dual processor sun machine. Results suggest that support vector classifier performance is statistically significant in comparison with NN and ML classifiers. To study the behaviour of support vector classifier with DAIS Hyperspectral data a total of sixty five features (bands) was used, a total of 65 features (spectral bands) were available, as seven features with severe striping were discarded, as explained above. The initial number of features used was five, and the experiment was repeated with 10, 15, …, 65 features, giving a total of 13 experiments. Figure 1 suggests that, in comparison to the other classifiers, the performance of the support vector classifier is quite good with small training data set irrespective of the number of features used.  Figure 1. Classification accuracies obtained with DAIS hyperspectral data using different classification algorithms. The training data set size is 200 pixels/class. Results obtained from analysis of the hyperspectral data suggest that classification accuracy using SVM increases almost continuously as a function of the number of features, with the size of the training data set held constant, whereas the overall classification accuracies produced by the ML, DT and NN classifiers decline slightly once the number of bands exceeds 50 or so. Thus, suggesting that the ‘Hughes phenomenon’ (Hughes, 1968) of decreasing classifier performance as the dimensionality of the feature space increases beyond a threshold is not supported by the experiments using the support vector classifiers. |