Abstract

The classification of color satellite images is presented using Multilayer Perceptron Neural Network and Support Vector Machine. Multilayer Perceptron is used for non-linear classification with 10 hidden layers using different number of epochs. A multiclass SVM is chosen for classification using radial basis function (RBF) kernel. Before performing classification, the image enhancement and feature extraction steps are carried out. The image enhancement is done using contrast stretching. The color features are extracted by using Principal Components Analysis (PCA). Classification results are obtained and testing is done by varying the number of images in the training and test datasets, the number of features and different classifiers. 100 images each obtained from Landsat satellite of NASA, US and Bhuvan geoportal of NRSC, Hyderabad are used in classification. Seven class categories, residential land, commercial land, grasslands, evergreen forest, mixed forest, sediments and clear water are identified. The results are analyzed and it is observed that SVM provides better results as compared to Multilayer Perceptron (MLP). Performance analysis is carried out with respect to classification accuracy and time.