CV4FEE: Flood Extent Estimation Using Consensus Voting in Ensemble of Methods for Change Detection in Sentinel-1 GRD SAR Images

There are several methods for finding flooded areas using change detection from C-band Sentinel-l temporal GRDH Synthetic Aperture Radar (SAR) images. Using these methods, an automatic flood extent mapping is obtained through differences between multiple time instances of SAR images, which itself is processed to remove noise and isolated pixels. While a few selected methods give comparable results using similar approaches, it is difficult to determine the best result amongst them. Hence, we propose CV4FEE, which is a consensus voting (CV) of outcomes from an ensemble of flood extent estimation (FEE) methods. We select three different state-of-the-art FEE methods, based on a specific criterion. We then use different consensus voting functions to compute deterministic as well as probabilistic flood extent maps, thus accounting for uncertainty in computation in selected methods. We demonstrate our results of flood area estimation in two different flooding events in India, namely the Hyderabad flood and the Nivar cyclone, in 2020. We show that the probabilistic maps give flood extent estimation that confirms the GEE4FLOOD results, qualitatively. Our method successfully demonstrates the uncertainty in the estimation process.