Automatic floods detection with a kernel k-means approach

Abstract

The important information for flooding crises management is to have a map showing a contour of damaged areas in a few times as possible. The remote sensing imagers, especially the Synthetic Aperture Radar (SAR) in a high spatial resolution can offer a global view of the situation. Indeed, detection of flooded areas will become a challenge since the reaction time of the teams on the ground should be as short as possible. Such method should avoid a complex parameterization, large time of compilation and long intervention of the operator. An automatic method based on an unsupervised clustering done in three steps is proposed. First of all, a Digital Elevation Model (DEM) is used as a prior information to localize high probability of floods. Then, the separation of the wet and dry pixels is done by a method called non-linear clustering kernel k-means. Finally, to isolate the flooded pixels from the permanent water, a non linear clustering with a log ratio image is applied in the features space. Two images polarized Vertical-Vertical (VV) with a high spatial resolution from RADARSAT 2 were used in this work. The study area is localized in the South-west part of Madagascar (Toliary). The Haruna hurricane was passed in this region on February 22nd, 2013. The final result of this study is a map showing the flooded areas. Because of lack of ground truth data, we couldn't valid our result with a confusion matrix. But we have compared it with the results obtained by current methods as the manual and the color composite methods. The comparison has shown that our approach has had a good compromise on flood detection.