Development of Continuous Spatiotemporal Flood Masks using Deep Learning and Remote Sensing

Abstract

Sentinel-1 is a satellite with a synthetic aperture radar(SAR) instrument. It is an active microwave satellite that provides data up to 10m resolution in all weather conditions and day/night, making it suitable for detecting floods. Water bodies appear dark in sentinel-1 due to microwaves' high absorbance, making detecting the water from the background possible. Satellites operating in the visible range of EM spectra are well-suited for detecting water bodies. However, clouds and shadows block visible light from reaching the satellite, making it challenging to get satellite images during flood events. In this work addressing this problem, we chose 20 different places on different kinds of terrains worldwide, like large complex rivers, high urban areas, small rivers, large lake areas, etc., to detect the water in all types of terrains. We have evaluated the performance of Logistic Regression, XGBoost, and U-Net that use sentinel-1 image as input and dynamic worl data(first 10m resolution near-real time land use land cover data developed by Google) as the target to detect the water bodies. After testing different models, U-Net outperformed all other models in various terrains and has an average F1 Score greater than 0.9.