2026 CSDMS meeting-009
From CSDMS
(if you haven't already)
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu
Browse abstracts
A Physics-Guided Input Framework for LSTM-Based Streamflow Modeling in Snow-Affected Watersheds
Mohamed Fathi Said,
Florida Gulf Coast University Fort Myers Florida, United States. m.fathi.said0@gmail.com
Accurate streamflow prediction in snow-dominated catchments remains a persistent challenge, primarily due to the nonlinear and temporally evolving interactions among precipitation phase, temperature variability, snow accumulation, and melt processes. Although machine learning approaches, and Long Short-Term Memory (LSTM) models in particular, have demonstrated strong capability in hydrologic modeling and streamflow prediction, their performance is often limited by the quality and physical relevance of the input forcings. This study investigates whether incorporating physics-informed, hydrologically meaningful inputs derived from standard meteorological variables can further enhance LSTM-based streamflow predictions. To evaluate this, we apply LSTM models across a subset of more than 85 snow-affected catchments from the CAMELS-US dataset. Instead of using raw precipitation time series as input, we partition precipitation into rain and snow components using temperature-dependent probability functions. The partitioned snow time series is also used to derive hydrologically meaningful variables, including snow accumulation and basin-scale snow storage. Finally, these enhanced physics-informed inputs are one-by-one incorporated within the LSTM framework. The results demonstrate a clear improvement in predictive performance, particularly in the historically low-performing catchments. The 25th percentile threshold of Nash-Sutcliffe Efficiency (NSE) increased from 0.51 using standard meteorological inputs to 0.62 with the physics-informed input set. Additionally, the median NSE across all basins improved by 0.07. These findings highlight the potential of integrating simple physical principles into data-driven models to bridge the gap between process understanding and machine learning, offering efficient pathways for improving hydrologic streamflow prediction in snow-affected watersheds.
