An Unsupervised Probabilistic Method for Large Scale Flood Mapping: Exploring Archive of Sentinel-1A/B Satellites over India
Synthetic aperture radar (SAR) imaging provides an all-weather sensing technique that is suitable for near-real-time mapping of disasters such as floods. In this article, I use SAR data acquired by Sentinel-1A/B satellites to investigate a flood event that affected the Indian state of Kerala in August 2018. I apply a Bayesian approach to generate probabilistic flood maps, which contain for each pixel its probability to be flooded rather than binary flood information. I find that the extent of the flooded area begins to increase throughout Kerala after August 8, with the highest values on August 9 and August 21. I observe no apparent correlation between the spatial distributions of the flooded areas and the rainfall amounts at the district level of the study area. Instead, larger flooded areas are in the districts of Alappuzha and Kottayam, located in the downstream floodplain of the Idduki dam, which released a significant volume of water on August 16. The lack of apparent correlation is likely due to two reasons: first, there is often some delay between the rainfall event and the flooding, especially for rather large catchments where flood waves need some time to reach floodplains from higher elevations. Second, rainfall is more abundant at overhead catchments (hills and mountains), whereas flood occurs further downstream in the floodplains. Further comparison of our SAR-based flood maps with optical data and flood maps produced by moderate resolution imaging spectroradiometer highlights the advantages of our data and approach for rapid response purposes and future flood forecasting.