Yin et al. (2025) DACSA: deformable average channel and spatial attention model for wildfire prediction and drivers

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2025
• Date: 2025-09-12
• Authors: Ke Yin, Lifu Shu, Pengle Cheng, Mingyu Wang, Ying Huang
• DOI: 10.1007/s00477-025-03088-9

Research Groups

• School of Technology, Beijing Forestry University, Beijing, China
• Ecology and Nature Conservation Institute, Chinese Academy of Forestry Beijing, Key Laboratory of Forest Protection of National Forestry and Grassland Administration, Beijing, China
• Department of Civil, Construction and Environmental Engineering, North Dakota State University, Fargo, ND, USA

Short Summary

This study proposes the Deformable Average Channel and Spatial Attention (DACSA) model, integrated with Location-aware Adaptive Normalization (LOAN), for improved wildfire prediction and analysis of driving factors using remote sensing data. The model demonstrates superior performance over state-of-the-art methods and provides quantitative insights into the importance of various wildfire drivers.

Objective

• To propose and evaluate the Deformable Average Channel and Spatial Attention (DACSA) model, combined with Location-aware Adaptive Normalization (LOAN), for enhancing wildfire prediction accuracy and quantifying the driving factors of wildfires using remote sensing images.

Study Configuration

• Spatial Scale: Greece and parts of the eastern Mediterranean, covering an area of 1253 km by 983 km, with a resolution of 1 km by 1 km.
• Temporal Scale: Data from 2009 to 2021, with daily resolution for wildfire prediction.

Methodology and Data

• Models used: Deformable Average Channel and Spatial Attention (DACSA) model, Location-aware Adaptive Normalization (LOAN). Comparative models include TimeSformer, SwinTransformer, SimAM, SE, CA, CBAM, ScConv, RF, XGBoost, LSTM, and ConvLSTM.
• Data sources: FireCube dataset (2009–2021), comprising 90 variables at 1 km x 1 km spatial resolution and daily temporal resolution. Data sources within FireCube include:
  • Meteo (ERA5-Land): temperature, wind speed and direction, precipitation, relative humidity.
  • Satellite (MODIS): land surface temperature, Normalized Difference Vegetation Index (NDVI)/Enhanced Vegetation Index (EVI), Leaf Area Index (LAI)/Fraction of Absorbed Photosynthetically Active Radiation (FPAR), evapotranspiration.
  • Soil moisture (European Drought Observatory).
  • Topography (EU-DEM): elevation, slope, aspect.
  • Land Cover (Copernicus Corine Land Cover).
  • Population Density (WorldPop).
  • Roads Density (OpenStreetMap).
  • Burned areas (European Forest Fire Information System (EFFIS), MODIS active fires product).

Main Results

• The LOAN with DACSA model achieved superior performance in wildfire prediction compared to state-of-the-art deep learning and traditional machine learning models, with a notable 4.89% higher F1-score on the 2021 extreme wildfire season test set.
• DACSA, as an attention mechanism, outperformed mainstream attention methods (SimAM, SE, CA, CBAM, ScConv) across multiple evaluation metrics (Overall Accuracy, Mean Accuracy, Area Under the Receiver Operating Characteristic curve, Intersection over Union, F1-score).
• Ablation studies confirmed the effectiveness of deformable convolution, the use of average pooling, and the optimal positioning of the attention module within DACSA.
• Finer temporal resolution (10 days) led to more accurate prediction outcomes compared to coarser resolution (20 days).
• Analysis of DACSA's channel attention weights identified total precipitation, maximum wind speed, distance to roads, arable land area, forest coverage, and water-bearing areas as primary wildfire driving factors. Secondary factors include NDVI, maximum temperature, maximum dew point temperature, minimum relative humidity, pastures, and miscellaneous vegetation.
• The model's predicted wildfire locations closely matched actual occurrences, particularly evident in winter.

Contributions

• Proposed the Deformable Average Channel and Spatial Attention (DACSA) model, a novel attention mechanism specifically designed for wildfire prediction using remote sensing images.
• Developed and evaluated LOAN with DACSA, a new wildfire prediction model that integrates DACSA with Location-aware Adaptive Normalization (LOAN), demonstrating significant improvements in prediction accuracy.
• Provided a quantitative analysis of wildfire driving factors by leveraging DACSA's attention weights, offering insights into the relative importance of various dynamic and static variables.
• Demonstrated the superior performance of the proposed model and attention mechanism against existing deep learning and traditional machine learning approaches, and other mainstream attention methods.
• Offered a more accurate and reliable analysis of wildfire driving factors compared to traditional methods like SHAP analysis, particularly under consistent data conditions.

Funding

• National Key R&D Program of China (2023YFD2202001)
• National Natural Science Foundation of China (Grant No. 32171797)

Citation

@article{Yin2025DACSA,
  author = {Yin, Ke and Shu, Lifu and Cheng, Pengle and Wang, Mingyu and Huang, Ying},
  title = {DACSA: deformable average channel and spatial attention model for wildfire prediction and drivers},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2025},
  doi = {10.1007/s00477-025-03088-9},
  url = {https://doi.org/10.1007/s00477-025-03088-9}
}