Kim et al. (2025) Assessment of deep learning models integrated with weather and environmental variables for wildfire spread prediction and a case study of the 2023 Maui fires

Identification

• Journal: Natural Hazards
• Year: 2025
• Date: 2025-12-22
• Authors: Jiyeon Kim, Y. Hu, Negar Elhami Khorasani, Kai Sun, Ryan Zhenqi Zhou
• DOI: 10.1007/s11069-025-07807-x

Research Groups

• GeoAI Lab, Department of Geography, University at Buffalo, Buffalo, NY, USA
• Department of Computer Science and Engineering, University at Buffalo, Buffalo, NY, USA
• Department of Civil, Structural and Environmental Engineering, University at Buffalo, Buffalo, NY, USA

Short Summary

This study assesses five deep learning models for wildfire spread prediction in Hawaii, comparing the best performers (ConvLSTM, ConvLSTM with attention) against the FARSITE model using the 2023 Maui fires as a case study. It finds FARSITE generally superior in accuracy but highlights the deep learning models' flexibility with widely available input data, and identifies key environmental factors influencing the Maui fires.

Objective

• RQ1: What are the advantages and limitations of five different deep learning models typically used for fire spread prediction?
• RQ2: What are the advantages and limitations of the deep learning models compared with a physics-informed and semi-empirical fire spread model?
• RQ3: What are the major weather and environmental factors, such as wind, vegetation, and topography, associated with the 2023 Maui fires?

Study Configuration

• Spatial Scale: State of Hawaii, with a case study on Maui. Fire data were processed into 375 meter (m) grid cells, forming 64 x 64 pixel raster images covering 576 square kilometers (km²). Input data resolutions varied: weather data at approximately 9 km (ERA-5 Land) or 31 km (ERA-5), topography at 10 m, Normalized Difference Vegetation Index (NDVI) at 500 m, and population density at 1 km, all resampled to 375 m.
• Temporal Scale: Wildfire data spanned from January 20, 2012, to August 12, 2023 (11 years and 7 months). Fire detection data had a temporal resolution of approximately 12 hours. Weather data were hourly, aggregated to 12-hour minimum and maximum values. NDVI data were 8-day, composited daily using a 16-day window. Population density data were updated annually. Models predicted fire spread for 12 hours, 24 hours, 36 hours, and 48 hours into the future.

Methodology and Data

• Models used:
  • Deep Learning Models: Long Short-Term Memory (LSTM), U-Net, U-Net with attention, Convolutional LSTM (ConvLSTM), ConvLSTM with attention.
  • Traditional Fire Spread Model: FARSITE (physics-informed and semi-empirical).
  • Explainable AI Method: Integrated Gradients (implemented using the Captum Python package).
• Data sources:
  • Fire Data: NASA Fire Information for Resource Management System (FIRMS) from Visible Infrared Imaging Radiometer Suite (VIIRS) on the Suomi National Polar-orbiting Partnership (NPP) satellite.
  • Weather Data: ERA-5 and ERA-5 Land (temperature, total precipitation, relative humidity, wind speed, wind direction, cloud cover).
  • Topography Data: National Elevation Dataset from the United States Geological Survey (USGS) (elevation, slope, aspect).
  • Vegetation Data: NASA VIIRS Vegetation Indices (VNP13A1) for NDVI; LANDFIRE for fuel model, canopy height, and canopy cover (for FARSITE).
  • Anthropogenic Activity Data: LandScan Global Population Database from Oak Ridge National Laboratory (ORNL) (population density).
  • Fuel Moisture Data (for FARSITE): Fire weather observation data from weather stations, processed using FireFamilyPlus software (based on the National Fire Danger Rating System - NFDRS).

Main Results

• Deep Learning Model Performance: ConvLSTM and ConvLSTM with attention achieved the highest F1-scores among the five deep learning models, with ConvLSTM showing higher precision and ConvLSTM with attention showing higher recall. U-Net models exhibited lower F1-scores due to lower precision, often predicting an overly expanding fire pattern. Training times ranged from 41.1 seconds (s) to 1056.1 s (17.6 minutes), while all models made predictions in less than 1 s.
• Comparison with FARSITE: FARSITE generally demonstrated higher precision, lower recall, and a higher F1-score than the best deep learning models (ConvLSTM and ConvLSTM with attention), particularly when provided with FARSITE-specific input data. However, deep learning models using widely available NDVI data achieved comparable performance to FARSITE, which requires more detailed vegetation and fuel moisture data. Multi-step predictions consistently resulted in decreased recall and F1-score across all models compared to single-step predictions.
• Feature Importance for 2023 Maui Fires: Weather-related factors (precipitation, wind speed, relative humidity, temperature) were consistently identified as highly important. Vegetation characteristics (canopy height, canopy cover, fuel model) also played significant roles. Topographic factors (elevation, aspect) were moderately important. Fuel characteristics (NDVI, fuel model) increased in importance for multi-step predictions. Population density was found to be the least important feature for 12-hour interval predictions.

Contributions

• Provided a comprehensive assessment of five deep learning models for wildfire spread prediction using over a decade of wildfire data from Hawaii.
• Conducted a direct comparison between the best-performing deep learning models (ConvLSTM, ConvLSTM with attention) and a widely-used physics-informed model (FARSITE) using the 2023 Maui fires as a critical case study.
• Demonstrated the practical advantage of deep learning models in their flexibility regarding input data availability, showing similar performance using widely available NDVI data compared to FARSITE's requirement for detailed vegetation and fuel moisture data, thus enhancing applicability in data-scarce regions.
• Identified key weather and environmental factors associated with the 2023 Maui fires using explainable AI, corroborating and complementing existing reports on the disaster.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Citation

@article{Kim2025Assessment,
  author = {Kim, Jiyeon and Hu, Y. and Khorasani, Negar Elhami and Sun, Kai and Zhou, Ryan Zhenqi},
  title = {Assessment of deep learning models integrated with weather and environmental variables for wildfire spread prediction and a case study of the 2023 Maui fires},
  journal = {Natural Hazards},
  year = {2025},
  doi = {10.1007/s11069-025-07807-x},
  url = {https://doi.org/10.1007/s11069-025-07807-x}
}