Remote Sensing Image Classification with Multiple Classifiers Based on Support Vector Machines

Classification accuracy is one of major factors influencing the application of classified image. This Paper proposes a SVM-based multiple classifiers fusion method for remote sensing image classification. We use both spatial Gabor wavelet texture feature and spectral feature to construct SVM classifier separately. then taking advantage of characteristic of SVM, namely for a given sample, the larger is the distance to the hyper plane, the more reliable is the class label. so the most reliable classification result is thus the one that gives the largest distance. This is our decision fusion rule. Using Landsat ETM+ satellite image as test data, the experimental results indicate that all classes including water, mountain, gobi, vegetation, desert and resident area could be well classified, and the overall accuracy achieved 86.5%, more than other each separate SVM classifier.