Machine Learning and Deep Learning for Snowmelt Prediction: Long-Term Trends and Implications for Water Resource Management

Snowmelt (SM) is a critical component of the hydrological cycle, particularly in regions dependent on SM for irrigation and drinking water. This study addresses the pressing need for precise SM predictions by analyzing historical SM data from 1948 to 2014 in Northern Pakistan and examining their correlation with 38 independent variables (IVs). Leveraging advanced Machine Learning (ML) and Deep Learning (DL) algorithms, we developed predictive models to capture the complex dynamics of SM. Among the ML models, Random Forest (RF) outperformed Linear Regression (LR), achieving an R² value of 0.9263 and a Root Mean Square Error (RMSE) of 0.0094 kg m⁻², highlighting its ability to model non-linear relationships. For DL-based predictions, the Long Short-Term Memory (LSTM) model demonstrated superior performance, with an R² of 0.9697 and a Mean Absolute Error (MAE) of 0.0035 kg m⁻², showcasing the efficacy of DL in capturing high-dimensional interactions. Feature importance analysis identified key drivers of SM, including Storm Surface Runoff, Soil Moisture Content, and Ground Heat Flux, underscoring the need to integrate diverse datasets for accurate predictions. These findings have significant implications for water resource management, flood forecasting, climate change monitoring, and avalanche prediction. By enhancing the precision of SM forecasts, this study supports the development of adaptive strategies to address water scarcity and mitigate flood risks, thereby advancing the Sustainable Development Goals (SDGs) related to clean water and climate action. The integration of ML and DL techniques provides a robust framework for understanding and managing the complex interplay of factors influencing SM, offering a foundation for sustainable water resource planning in vulnerable regions.