Flood monitoring and change detection based on unsupervised image segmentation and fusion in multitemporal SAR imagery

This paper presents an unsupervised method for change detection and analysis of the behavior of floods using multitemporal SAR (Synthetic Aperture Radar) images. First, images were filtered using the Enhanced Frost Filter in order to reduce the effect of speckle noise. Fuzzy Clustering Means (FCM) and k-means algorithms were used for unsupervised segmentation, and both results were fused using PCA. Subsequently, a Boolean image was created from the change information using a thresholding algorithm. Finally, the area of changes in the scene was calculated with spatial resolution information from the images. For the experiment phase, synthetic images were first created with varying levels of speckle noise, making it possible to evaluate the performance of the proposed method. The results showed an overall accuracy of approximately 99% and a kappa index of 0.76 for images whose Equivalent Number of Looks (ENL) equals 0.7. This shows that the proposed method is efficient in detecting changes in SAR images with an ENL greater than or equal to 0.7. Finally, two SAR images were tested, one before and one after a flood that covered an area of the Magdalena River in Colombia called Plato-Magdalena. Our method found that the river flooded approximately 131.51 hectares of terrain in the case of the studied images.