Changes detection in remote sensing multitemporal image data by apllying Support Vector Machines with the use of polynomial kernel and RBF kernel (Radial Basis Function kernel)

Abstract

This paper investigates an approach to the problem of detecting changes in multitemporal remote sensing images using Support Vector Machines (SVM) with the use of the polynomial kernel and RBF kernel (kernel based radial function). For the experiments two Landsat 5-TM images were used covering the same area, located in the State of Roraima, Brazil (61°37’W–61°49’W of longitude and 3°40’N–3°52’N of latitude). The methodological proposal is based on the difference of fraction images. The difference in soil and vegetation fractions in natural scene images tends to have a symmetrical distribution around the origin and this fact is used to model two normal multivariate distributions: change and non-change. The Expectation-Maximization (EM) algorithm was implemented to estimate the parameters associated with these two distributions. Random samples were extracted from the distributions and used to train the SVM classifier. Two procedures were used to assess the accuracy of the methodology. First, the qualitative analysis carried out through the production of the change map. Then, the quantitative analysis carried out through the construction of the confusion matrix using a synthetic image. It was observed that the RBF kernel presented very similar results for all sets of test samples, regardless of the size of the training sample set, which does not occur with the polynomial kernel. The experiments developed in this work show the adequacy of the proposed methodology, producing acceptable results in the detection of changes in soil cover, since the SVM is a robust method, handles the problem of dimensionality and with noisy samples and requires a number small sample of training samples for the classification process.