An End-to-End Framework for Multimodal and Multitemporal Analysis of Water Bodies Using the SEN12-WATER Dataset

Water detection and monitoring have increasingly gained immense attention in recent years for their relevance in providing insights into water dynamics, anticipating dangerous events over time with high accuracy. This article aims to tackle these challenges by introducing a novel multimodal and multitemporal dataset, representing the evolution of SEN2DWATER. This new dataset, named SEN12-WATER, integrates spatio-temporal information from various data sources, namely SAR polarizations, elevation, slope, and multispectral optical data. The dataset is benchmarked through an end-to-end deep learning framework, combining a U-Net architecture for water body segmentation and a Time-Distributed Convolutional Neural Network for next-frame prediction. The experiments demonstrate how the proposed dataset and methodology offer efficient tools enabling accurate water resource management and water loss prediction, with extensive validation against data acquired on-ground. The proposed framework is demonstrated through a use case focused on drought-prone regions in Italy and Spain, chosen for their well-documented water anomalies during 2016–2022. Notably, what is proposed is of significant interest to both the remote sensing and environmental science community, supporting ongoing research and possibly being of support for decision-makers to manage water resources more effectively.