A Novel Statistical-Based Scale-Independent Approach to Unsupervised Water Segmentation of SAR Images

In this work, a novel approach to water segmentation of Synthetic Aperture Radar images is presented. The proposed methodology, suitable for any VV-VH dual-pol data source, it is based only on the statistical and morphological content of the two polarimetric channels. No external information or supervision is required to successfully complete the segmentation task, allowing a complete automation of the process. The methodology has been applied to Sentinel-1 Interferometric Wide-Swath data and validated using very high-resolution optical images acquired by an Unmanned Aerial System over two Italian rivers, during different seasons. The performances of the algorithm have been estimated carrying out a pixel-based validation, deriving from the binary confusion matrix two classical figures of merit such as Global accuracy and F1-score. In the validation procedure, it has been assessed an overall Global Accuracy of 0.92 and an overall F1-score of 0.81, suggesting that the presented methodology applied to S-1 data is well-suited for the monitoring of rivers characterized by a wet channel width greater than 60 m.