Change Detection in Satellite Images Using Self-Organizing Maps

In several applications it is necessary to detect changes from aerial images analysis. The goal, which has been addressed in several studies, is to develop an automatic method with low computational cost. However, it is a difficult task and isolated contributions have been presented. The main purpose of this paper is present a proposal to detecting change using aerial images obtained by satellite. Kohonen´s Self-Organizing Maps (SOM) is used to identify the classes of the images and the change detection is based on post-classification technique. In this paper, a set of images from different period of time from Brazil has been employed. As input of the self-organizing maps, used as an usupervised and nonparametric artificial neural network, it was obtained different features. The self-organizing map is evaluated by varying the type of topological neighborhood function, as a Gaussian, truncated and square Gaussian. The results are obtained through the comparison between the outputs obtained with these maps of colors. The experiments performed on the satellite image have shown that SOM is efficient and have better results for the area of study were obtained using Gaussian and Truncated Gaussian functions.