Machine learning insights for fire impacts on snow disappearance predictability in Northern California

Snowpacks are a critical water resource in the western United States, where effective water management depends on accurate monitoring and forecasting. Wildfires cause lasting perturbations to atmosphere–vegetation–snow interactions that are increasingly altering snow dynamics and affecting snowpack predictability. This study examines the influence of fires on snow disappearance predictability using a suite of machine learning (ML) experiments. Models trained solely on no/low-burn areas to predict the day of snow disappearance (DSD), using hydrometeorological predictors without explicit fire information, perform significantly worse when applied to moderate-to-high burn areas. Specifically, across no/low-burn areas, random forest models achieve a bias standard deviation of 12.0 d and R ^2 of 0.86, compared to 14.0–15.3 d and R ^2 of 0.72–0.79 in more heavily burned areas, with qualitatively consistent results across other ML models with variable complexity (multiple linear regression, long short-term memory, multilayer perceptron, and extreme gradient boosting). The systematic overestimation of DSD across pixels with >75% burn fraction likely indicates that post-fire-enhanced snow ablation is not captured by the models. Including samples from burned areas in model training—implicitly incorporating fire effects—improves DSD predictions in burned landscapes without degrading performance in unburned regions. The largest accuracy improvements occur in pixels with >75% burn fraction, where DSD overestimation is reduced from 4.0 to 1.6 d. Nevertheless, model performance in moderate-to-high burn areas remains lower than in no/low-burn regions ( R ^2 lower by 0.05–0.06; bias standard deviation higher by 1.8–2.3 d). Adding explicit fire-related predictors, namely burn fraction and vegetation greenness, does not yield further accuracy gains. These results demonstrate that wildfires significantly impact snowpack dynamics and that including post-fire landscapes in training data is essential to improving ML-based snow predictions in burned areas. Fully capturing fire impacts, however, may require more detailed or process-specific fire information to account for complex post-fire snow dynamics.