A Comparison Of Remote Sensing Approaches To Assess The Devastating May-June 2022 Flooding In Sylhet, Bangladesh

In May-June 2022, northeastern Bangladesh suffered devastating flooding affecting over 7 million people, to which various public agencies responded by producing maps of flooded areas derived from satellite images. In the wake of this and the growing availability of satellite flood algorithms and end-user-products, we compared surface water maps for the Sylhet District generated from various remote sensing approaches, focusing on (a) "local" versus "global" and (b) machine learning (ML) versus "traditional" (non-ML) methods. Specifically, we assessed (1) a "local" Sentinel-1 change detection algorithm calibrated on four recent floods in Bangladesh, (2) a pre-trained "globally applicable" ML Sentinel-1 algorithm, (3) the Copernicus Global Flood Monitoring (GFM) tool, (4) a deep learning (DL) "fusion" of MODIS and Sentinel-1 and (5) a MODIS algorithm used to produce the Global Flood Database (GFD). Evaluating the Sentinel-1 based approaches against hand-labels, we obtained accuracy scores of 82.0%, 82.4% and 73.4% for the "local", "global ML" and GFM maps, respectively, suggesting that the GFM may trade off accuracy in return for automated global coverage. The comparable performance of the "global ML" and the "local" algorithms suggests that ML can generalize effectively to region-specific flood events and could offer a better option than the (non-ML) algorithms currently used by GFM. Comparing all methods spatiotemporally shows that the MODIS-Sentinel-1 "fusion" overcomes the low accuracies encountered using only MODIS during persistent cloud cover, offering encouragement for the potential of multi-sensor fusion across a range of remote sensing applications.