Publication

Abstract : The land cover classification and urban analysis of remotely sensed images has become a challenging problem, hence efficient classifiers are required in order to combat the problem of classifying the huge remote sensing aerial datasets. In this paper we have proposed the use of Random Kitchen Sink (RKS) algorithm and Regularized Least Squares (RLS) classifier for the classification of aerial image. The new machine learning algorithm RKS, primarily engages in mapping the feature data to a higher dimensional space and thereby generates random features. These randomized data are then adopted by RLS classifier for the classification task. It is observed that the randomization of the data reduces the computation time needed for training. The experiment is performed on five classes of the UC Merced Land Use Aerial Imagery Dataset. The efficiency of the proposed method is estimated by comparing the accuracy results with the conventional classifier namely, Support Vector Machine (SVM). Experimental result shows that the proposed method produces a high degree of classification accuracy i.e. 94.4%, when RBF kernel with LOO (Leave One Out) cross-validation was used, when compared to SVM. In this paper, statistical features show better precision and accuracy in classifying different set of classes, compared to textural features in both the classification approaches. Hence, better accuracies could be attained for multi class classification when compared to other classification technique like, SVM since, the random features reduces computation time and enhance the performance of kernel machines.