Detection of floods in SAR images with non-linear kernel clustering and topographic prior

After a major flood catastrophe, a precious information is the delineation of the affected areas. Remote sensing imagery, especially synthetic aperture radar, allows to obtain a global and complete view of the situation. However, the detection of the flooded areas remains a challenge, especially since the reaction time for ground teams is very short. This makes the application of automatic detection routines appealing. Such methods must avoid complex parametrization, heavy computational time and long intervention by the operator. We propose an automatic three steps strategy, starting by rebalancing the different types of pixels (non-water, permanent water and flooded) using digital elevation model information, then isolating water pixels and finally separating flooded from permanent water pixels using non-linear clustering in dedicated feature spaces. Experiments on two sets of ASAR images show the effectiveness of the method competing with supervised standard log-ratio thresholding.