Detection of unsupervised SAR images changes for the monitoring of the behavior and analysis of floods in Colombia

Abstract

This paper proposes an unsupervised method of detecting changes for studying the behavior of the floods in Colombia using images SAR (Synthetic Aperture Radar). The proposed change detection approach is a pre-classification method, in which first images are classified and then changes are found through relational operators. First, SAR images were filtered with filter-Enhanced Frost in order to eliminate the effect of Speckle noise. For unsupervised segmentation of SAR images, the Fuzzy Clustering (FCM) approach and k-means were used. The results from both approaches were fused in order to improve accuracy in segmentation using the method of image fusion PCA (principal Component Analysis). For the detection of changes, operators ratio Mean-Log-Ratio Ratio were implemented, the results of both relational operators were also fused using PCA in order to improve accuracy in this phase of the proposed method. For the experiment, first, synthetic images created from homogeneous areas of SAR images of Colombia were used. With these images, it was possible to evaluate the performance of the proposed method. The results showed an overall accuracy about 99%, which concludes that the proposed method is highly efficient on the data set used. Then, two real SAR images were experimented, one before and one after a flood, which were provided by the IDEAM and cover the area of Magdalena in the PlatoMagdalena. The proposed method found in these images about 77.46 hectares in the municipality of Magdalena plate were flooded by the Magdalena River..