California Wildfire Prediction: A Practical Comparison Between Traditional Machine Learning and Multimodal Deep Learning

The George Washington University team conducted research on the California wildfire early warning system, comparing traditional tabular machine learning (e.g., Random Forest) with multimodal deep learning models to explore the optimal solution for predicting wildfires 16 days in advance. The study found that well-designed feature engineering is key, and Random Forest performed best under both evaluation strategies, providing a rigorous benchmark and practical insights for the development of wildfire early warning systems.

In recent years, global warming has led to frequent wildfires, and multiple large-scale wildfires in California from 2020 to 2024 have caused significant losses. The George Washington University team aims to predict whether a wildfire will occur within the next 16 days based on the past 16 days of environmental data, securing valuable time windows for emergency management.

The dataset covers California from 2020 to 2024, including 609,102 records, 5,343 9km×9km grid cells, and 114 16-day windows. Data sources include multi-source heterogeneous information such as CAL FIRE wildfire history, Landsat 8 NDVI images from Google Earth Engine, ERA5-Land meteorological data, NASADEM terrain data, infrastructure and census data, etc.

Feature Engineering is divided into four categories: historical fire features (lagged records, summaries), meteorological derived features (drought index, etc.), composite risk features (interaction terms, ignition risk), and seasonal features (cycle encoding, fire season indicators). Model Architecture: Traditional tabular models (Logistic Regression, Random Forest, XGBoost, Decision Tree); Multimodal DL models (ResNet18+MLP, EfficientNet-B2+MLP, UNet+MLP, ResNet18+Random Forest late fusion).

Evaluation Strategies: Time-based split (train before 2022, validate in 2023, test in 2024) to simulate real scenarios; Random split (60/20/20) to test ideal performance. Experimental Results: Random Forest was optimal (PR-AUC of 0.371 in time-based split, 0.715 in random split); Multimodal models performed similarly but did not surpass it; SHAP analysis showed that lagged fire features contributed the most, while NDVI increments were limited; Late fusion models improved fire recall rate.

Limitations: 9km resolution smooths small-scale events, class imbalance (1% fire rate), only using a single NDVI band, no generalization to other regions. Future Directions: Higher resolution data, cost-sensitive learning, multi-spectral band fusion, cross-region validation.

Practical Insights: Feature engineering is more critical than complex models; time-based split evaluation is more reliable; multimodal fusion requires complementary information. Summary: The success of Random Forest reflects the value of domain knowledge and feature engineering, providing a rigorous benchmark for wildfire prediction. Future efforts need to combine technological advancements to enhance system performance.