A Deep Neural Network Slope Reduction Model on Sentinel-1 Images for Water Mask Extraction

The monitoring of water bodies from space is one of the main challenges for flood risk assessment, food security and climate change monitoring applications. A common issue in Synthetic Aperture Radar (SAR) images water-bodies masks generation problem is falsely identifying as water, areas that are characterised by a complex morphology. Existing works on water bodies estimation either remove the steep slope areas or not, on a case-by-case manner. The deep learning era allows for automatic adaptation to Sentinel 1 images where the false removal of water, in cases of rivers and lake shores, can be avoided.
This paper proposes a DNN model that generates water-bodies masks for Sentinel-1 satellite data by fusing the SAR backscatter coefficients and the Digital Elevation Model (DEM) data. Hence, disregarding the steep sloped areas can eradicate these false positives, in the cost of removing at the same time some actually inundated areas.
Our proposed method uses as input the different polarisation bands of Sentinel-1 images and combines them with the corresponding DEM information to estimate the output class, filtering out the highly sloped areas, leading to improved delineation accuracy. The model involves a training stage using machine learning algorithms including neural networks. The algorithm decides upon the effectiveness of slope removal for the specific pixel or block of pixels.
In order to validate these findings and to quantify the impact of high slope removal, we run experiments on three major Italian lakes and their surrounding territories. Specifically, when filtering out these water areas, the F-Score, capturing the correct identification of water areas, increased from 81.1% to 94.1% for Garda lake, from 75.33% to 92.77% for Maggiore lake and from 75.54% to 88.91% for Trasimeno lake.