Improved snow water equivalent prediction with a hybrid SNOW17-LSTM model
Aamir Raj Lamichhane1 and Andrew Bennett1
1Department of Hydrology and Atmospheric Sciences
The University of Arizona, Tucson, AZ
Snowpack regulates streamflow and seasonal water availability in mountainous watersheds, making it a crucial component of the hydrologic cycle. SNOW17 is one of the most popular process-based models used to simulate snow ablation and accumulation. However, traditional process-based models like SNOW-17 rely on site-specific calibration, limiting their ability to generalize across diverse snow conditions. On the flip side, purely data-driven models such as LSTMs offer flexibility but lack physical interpretability. We hypothesize that a data-driven parameterization better captures site-specific snow dynamics than static calibration, while SNOW-17’s process representation constrains predictions to remain physically consistent and interpretable. Our framework uses LSTM networks to predict daily SNOW-17 parameters, enabling dynamic parameterization without explicit site-wise calibration.
This study uses daily data from over 730 SNOTEL sites across the western United States, with temperature and precipitation as inputs and snow water equivalent (SWE) as the target variable. We compare the hybrid model against the standalone SNOW-17 and LSTM models under three experimental setups: site-wise models, regional model using a temporal split, and spatial holdout testing based on snow-based site clustering. The preliminary results show that the pure data-driven approach (LSTM) achieves superior overall accuracy across experiments (median NSE > 0.9) compared to the hybrid model (median NSE ~0.8), but the hybrid approach demonstrates improved timing consistency. This highlights the value of process-based constraints in machine learning applications.
This work demonstrates how hybrid physics-ML approaches can balance predictive performance with physical consistency in snowpack modeling, providing insights crucial for operational use where both accuracy and realism matter.