Multisource classification using support vector machines: An empirical comparison with decision tree and neural network classifiers

Remote sensing image classification has proven to be attractive for extracting useful thematic information such as landcover. However, often for a given application, spectral information acquired by a remote sensing sensor may not be sufficient to derive accurate information. Incorporation of data from other sources such as a digital elevation model (DEM), and geophysical and geological data may assist in achieving more accurate land-cover classification from remote sensing images. Recently, support vector machines (SVM) have been proposed as an alternative for classification of remote sensing data, and the results are promising. In this paper, we employ the SVM algorithm to perform multi-source classification. An IRS-1C LISS III image along with normalized differenced vegetation index (NDVI) image and DEM are used to produce a land-cover classification for a region in the Himalayas. The accuracy of SVM-based multi-source classification is compared with several other non-parametric algorithms namely a decision tree classifier, and back propagation and radial basis function neural network classifiers. The well-known kappa coefficient of agreement is used to assess classification accuracy. The differences in the kappa coefficient of classifiers have been statistically evaluated using a pairwise Z-test. The results show a significant increase in the accuracy of the SVM based classifier on incorporation of ancillary data over classification performed solely on the basis of spectral data from remote sensing sensors.