Abstract
Accurate wildfire spread prediction is essential for disaster mitigation and resource allocation. While deep learning models have shown promise in processing complex remote sensing data, their "black box" nature often limits their utility in high-stakes decision-making. This study provides a comparative analysis of four deep learning architectures—Autoencoder, ResNet, UNet, and Swin-UNet—evaluating their performance on the "Next Day Wildfire Spread" dataset. By employing Explainable AI (XAI) techniques such as SHAP, Grad-CAM, and Integrated Gradients, we investigate the internal logic of these models. Our findings highlight a critical trade-off between precision and recall and demonstrate that while Transformer-based models like Swin-UNet offer high precision, traditional CNN-based UNets remain highly efficient for real-time applications.
Introduction
Predicting the trajectory of a wildfire is a race against time where every hour of lead time can save lives and property. Modern satellite imagery provides a wealth of environmental data, yet the models used to interpret this data are frequently opaque. For fire managers, a binary prediction is often insufficient; understanding the underlying drivers—such as wind velocity, drought conditions, or historical burn patterns—is vital for building trust in automated systems. This paper addresses the gap in comparative, interpretable analysis between traditional Convolutional Neural Networks (CNNs) and modern Transformer architectures in the context of wildfire forecasting.
Methodology
To evaluate the efficacy of different spatial modeling approaches, we implemented four distinct architectures:
- Autoencoder: Served as a baseline for dimensionality reduction and feature reconstruction.
- ResNet (Residual Network): A standard CNN utilizing skip connections to mitigate signal degradation in deep layers.
- UNet: An encoder-decoder architecture designed for image segmentation, utilizing long-range skip connections to preserve spatial detail.
- Swin-UNet: A hierarchical Vision Transformer that employs shifted windows for self-attention, allowing the model to capture long-range dependencies across the landscape.
To move beyond standard performance metrics, we integrated a suite of Explainable AI (XAI) tools. SHAP (SHapley Additive exPlanations) was used to determine global feature importance, while Grad-CAM and Integrated Gradients (IG) provided local attributions, visualizing which specific pixels and environmental variables (e.g., humidity, elevation) most influenced individual predictions.
Data and Features
The study utilized the Next Day Wildfire Spread dataset, covering California from 2012 to 2020 at a 1 km resolution. The models were trained on 12 input variables, including:
- Topography: Elevation.
- Meteorology: Wind direction/velocity, temperature extremes, humidity, and precipitation.
- Environmental Indicators: Drought severity, Normalized Difference Vegetation Index (NDVI), and the Energy Release Component (ERC).
- Human Factors: Population density.
- Temporal Context: The "Previous Fire Mask" (PFM), representing the fire's location at time t to predict its state at t+1.
Results and Discussion
Performance Comparison
The results indicate that architectures utilizing extensive skip connections—UNet and Swin-UNet—significantly outperform standard Autoencoders and ResNets. Swin-UNet achieved the highest AUC-PR (0.2803) and Precision (0.3686), making it particularly effective at minimizing false alarms. Conversely, UNet achieved the highest Recall (0.3859) with a competitive AUC-PR of 0.2739, suggesting it is better suited for "early warning" systems where missing a potential fire spread is the primary concern.
Computational Efficiency
A stark contrast was observed in computational requirements. UNet achieved its high performance with a load of only 0.024 GFLOPs, whereas Swin-UNet required 1.146 GFLOPs. This nearly 50-fold difference in efficiency suggests that while Transformers offer superior precision, UNet remains the more viable candidate for resource-constrained or real-time edge deployments.
Interpretability Insights
XAI analysis revealed that the most successful models prioritized the Previous Fire Mask (PFM). Models that failed to weight the PFM heavily, such as the baseline Autoencoder, performed poorly. Furthermore, Swin-UNet’s self-attention mechanism allowed it to integrate broader climate factors like drought and vegetation more flexibly than the CNN-based models, which focused more heavily on immediate local spatial patterns.
Conclusion
This research demonstrates that while Vision Transformers like Swin-UNet provide state-of-the-art precision in wildfire prediction, the UNet architecture offers a more computationally efficient alternative with superior recall. The use of XAI techniques confirms that the integration of historical fire data (PFM) is the single most important factor in predictive accuracy. Future work will focus on developing hybrid models that combine the local efficiency of CNNs with the global context-awareness of Transformers, as well as testing these architectures on diverse global datasets to ensure regional robustness.
Summary updated