Flood Forecasting with Sentinel-2 Images Using Machine Learning

This paper proposes a methodology for detecting flooded areas using Sentinel-2 images, followed by flood forecasting based on a combination of the deep neural network U-Net and a support vector machine (SVM). The U-Net architecture classifies a given Sentinel image into the two classes “water” and “no water”, the SVM subsequently performs a near-future prediction of flooded areas based on the U-Net results and additional information (DEM, land use information, precipitation data etc.). Experimental results demonstrate that for a test site in Ukraine the U-Net/SVM model achieves the highest overall accuracy (98.8%), slightly outperforming other models, including Random Forest and SVM. The resulting flood maps provide valuable information for planning rescue operations and territory management, allowing for rapidly identifying areas of flooding. It can thus contribute to a significant reduction in economic losses and an increase in emergency preparedness.