Towards Sustainable Water Management: An Attention-Based Approach to Short-Term Snow Water Equivalent Forecasting

Water obtained from snow serves as a freshwater supply to more than 1.5 billion people. However, estimates of snow melts are often inaccurate due to rising global temperature. This study uses a deep learning approach to analyze remote sensing data--Snow Water Equivalent (SWE), Land Surface Temperature, Precipitation, and Water Vapor--from NASA between 2002 and 2011, which have been validated through numerous validation campaigns. Leveraging recent advancements in attention-based mechanisms for learning spatio-temporal dependencies, we develop a novel model incorporating advances in biomedical image segmentation (U-NET) and the LSTM and a novel feature: Attention-Augmented Skipp Connections (AASC). The model effectively forecasts monthly SWE, achieving an impressive RMSE of 0.0224 on a scale of 1, significantly outperforming traditional methods. This level of accuracy demonstrates the utility of attention-based approaches in analyzing remote-sensing data and related tasks to SWE prediction. Our methodology introduces a unique approach to SWE forecasting by facilitating streamlined visualization through map generation. This study presents a framework for accessible and intuitive insights for stakeholders across various domains. The model's versatility allows for implementation on smaller scales while utilizing more comprehensive data, highlighting its potential as a valuable tool for sustainable water monitoring: irrigation, flood control, power generation, and drought management decisions.