Mohanty et al. (2025) Machine Learning Insights for Fire Impacts on Snow Disappearance Predictability in Northern California

Identification

• Journal: Mendeley Data
• Year: 2025
• Date: 2025-11-14
• Authors: Mohanty, Yashnil, Abolafia-Rosenzweig, Ronnie, He, Cenlin
• DOI: 10.17632/4xrk8g6458

Research Groups

Not explicitly stated in the provided text.

Short Summary

This study investigates the impact of wildfires on snow disappearance predictability in Northern California using machine learning models. It finds that wildfires significantly alter snow dynamics, and including post-fire landscapes in model training data is crucial for improving snow disappearance predictions in burned areas.

Objective

• To examine the influence of fires on snow disappearance predictability using a suite of machine learning experiments.

Study Configuration

• Spatial Scale: Regional (Northern California), down to pixel level (e.g., pixels with >75% burn fraction).
• Temporal Scale: Seasonal (day of snow disappearance) and multi-year (lasting perturbations from wildfires).

Methodology and Data

• Models used: Random Forest, Multiple Linear Regression, Long Short-Term Memory (LSTM), Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost).
• Data sources: Not explicitly detailed in the provided text, but includes hydrometeorological predictors, burn fraction, and vegetation greenness.

Main Results

• Machine learning models trained solely on no/low-burn areas perform significantly worse when applied to moderate-to-high burn areas.
• For random forest models, the bias standard deviation for snow disappearance day (DSD) is 12.0 days and R² is 0.86 in no/low-burn areas, compared to 14.0–15.3 days and R² of 0.72–0.79 in heavily burned categories.
• A systematic overestimation of DSD in very high burn areas suggests that post-fire-enhanced snow ablation is not captured by models trained on unburned data.
• Including samples from burned areas in model training 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 by over 50% (mean bias from 4.0 days to 1.6 days).
• Model performance in moderate-to-high burn areas remains lower than in no/low-burn regions (R² lower by 0.05–0.06; bias standard deviation higher by 1.8–2.3 days).
• Adding explicit fire-related predictors (burn fraction and vegetation greenness) does not yield further accuracy gains.

Contributions

• Demonstrates that wildfires significantly impact snowpack dynamics and that machine learning models need to account for these changes.
• Establishes that including post-fire landscapes in training data is essential for improving machine learning-based snow predictions in burned areas.
• Shows that implicitly incorporating fire effects through training data is more effective than explicitly adding fire-related predictors for the models tested.

Funding

Not explicitly stated in the provided text.

Citation

@article{Mohanty2025Machine,
  author = {Mohanty, Yashnil and Abolafia-Rosenzweig, Ronnie and He, Cenlin},
  title = {Machine Learning Insights for Fire Impacts on Snow Disappearance Predictability in Northern California},
  journal = {Mendeley Data},
  year = {2025},
  doi = {10.17632/4xrk8g6458},
  url = {https://doi.org/10.17632/4xrk8g6458}
}