Consistent Surface Water Monitoring by Fusing Sentinel-1 and -2 Through Convolutional Neural Networks

This study presents a modular approach to monitor surface water dynamics independently of cloud cover, by leveraging the complementary strengths of Sentinel-1 and Sentinel-2. First, scene-based water segmentation masks are derived through a U-Net architecture. Different input features are assessed and the impact of clouds is investigated. Combining both sensors, F1 scores of 0.82 and 0.76 are obtained for cloud-free and clouded test data respectively, compared to 0.61 using the Sentinel-2-based NDWI. Next, a CatBoost architecture is used to aggregate these scene-based predictions into monthly products. An assessment of the input feature space puts the average and standard deviation of the three scene-based models forward as best performing. For the latter, an F1 score of 0.76 is obtained.