Wildfire Spread Prediction: A Deep Learning System Integrating Multimodal Geospatial and Meteorological Data

This project aims to build a deep learning system integrating multimodal geospatial and meteorological data to achieve 64×64 pixel-level prediction of the next-day wildfire spread mask. The system integrates multi-dimensional data including meteorology, vegetation, terrain, and population, offers three model options: physics-enhanced UNet, ResNet-18 UNet, and logistic regression baseline, and supports optimization strategies like EMA and Polyak averaging, providing decision support for scenarios such as emergency management and risk assessment.

Wildfires are among the most destructive natural disasters globally, and climate change is increasing their frequency and intensity. Accurate prediction of short-term spread trends is crucial for decision-making, but wildfire behavior is influenced by non-linear interactions of multiple factors such as meteorology, vegetation, and terrain. Traditional physical models have high computational costs and struggle to handle complex relationships. Deep learning, through data-driven learning patterns, provides new ideas to address this challenge.

Multimodal Data Fusion

The system integrates meteorological data (temperature, humidity, wind speed), vegetation index (NDVI), fuel drought indicators (ERC/PDSI), terrain (slope/aspect), human-made barriers (land use, population density), and historical fire data, designing a 9-channel standardized input (previous day's fire mask, wind speed components, temperature and humidity, NDVI, slope, aspect, barrier information).

Model Architecture

1. Logistic Regression Baseline: Pixel-level regression provides a performance reference, supporting a complete training process and parameter configuration.
2. Physics-Enhanced UNet: Core model, introducing the PhysicsPrior module to encode physical mechanisms of wildfire spread (wind direction alignment kernel, slope driving, vegetation humidity damping, barrier suppression), using designs like SiLU activation and batch normalization.
3. ResNet-18 UNet: A variant based on the pre-trained ResNet encoder, enhancing feature extraction capabilities.

Training Optimization

Adopts mixed-precision training, cosine learning rate scheduling, and a composite loss of Focal Loss and Focal Tversky Loss, supporting two weight optimization strategies: EMA (Exponential Moving Average) and Polyak averaging.

Evaluation and Visualization

Core metrics include Average Precision (AP), F1 score, Intersection over Union (IoU), etc.; provides diagnostic tools like PR curves, confusion matrices, and training history curves; uses Test-Time Augmentation (TTA) during inference to improve robustness.

Practical Applications

• Emergency Management: Optimize fire-fighting resource allocation and plan evacuation routes.
• Insurance Assessment: Dynamic risk pricing and identification of high-risk assets.
• Research and Policy: Analyze climate change impacts and evaluate land management strategies.

Technical Highlights

1. Physics-Data Fusion: Embeds domain knowledge of wildfire spread into neural networks, balancing flexibility and generalization ability.
2. Modular Design: Decoupled components, facilitating expansion of data sources and model architectures.
3. Reproducibility: Ensures reproducibility of experimental results through configuration files and fixed seeds.

Limitations and Future Directions

• Limitations: 64×64 resolution may be insufficient, time granularity is only next-day, and data coverage is limited.
• Future: Explore higher resolution, multi-model integration, real-time data fusion, causal inference, and uncertainty quantification.

This project demonstrates the potential of machine learning to address major social challenges, improving prediction accuracy through multimodal fusion and physical priors. The project is fully open-source, sharing code, documentation, and best practices, aiming to promote community collaboration, accelerate technological progress, and contribute to protecting life, property, and ecological security. Researchers and practitioners in the fields of geospatial AI and disaster prediction are welcome to follow and participate.