Precise Delineation of Small Water Bodies from Sentinel-1 Data using Support Vector Machine Classification

Abstract

SAR images are increasingly used for delineating water bodies. The high revisit frequency of Sentinel-1A and -1B make these images ideal for studies in hydrology that require data during the wet seasons and intense precipitation events. The aim of our study consisted in determining the attainable accuracy in delineating water bodies using Sentinel-1 data and propose a scheme for the extraction of the water surface. We chose a study site of three small reservoirs in the Pampulha area of Belo Horizonte, Brazil. The Support Vector Machine (SVM) classifier was used to separate water surfaces from the land using a probability threshold of 95%. In our scheme, we combined a SVM classification with a probability map to estimate the partial water area of the pixels within the transition zone between water and land. The VV polarization provided the best separation between water and land and the best geometric accuracy for the study area. Following the proposed approach, adding the water area of the marginal pixels to the water surface yielded accuracies better than 90% and above 80% when the geometry errors were also considered. Our results demonstrated the applicability of using high-resolution Sentinel-1 SAR data for the accurate mapping of water surfaces.