Wildfire Risk Prediction in Idaho: Multi-source Data Fusion and Machine Learning Model Construction

This article focuses on the county-level wildfire risk prediction project in Idaho, exploring how to integrate NOAA meteorological data, USGS geographic data, and NIFC fire data to build a prediction model using machine learning. The project covers feature engineering, model selection, spatial prediction methods, and the practical application value of wildfire risk management, providing technical references for natural disaster early warning.

Wildfire prediction is a complex spatio-temporal problem facing multiple challenges:

1. Multi-factor coupling: Affected by meteorology, terrain, human factors, and ignition sources together;
2. Spatio-temporal heterogeneity: Significant differences in risk patterns across regions/seasons;
3. Data scarcity: Insufficient historical records due to rare fire occurrences;
4. Prediction scale: County-level is suitable for resource allocation, grid-level for early warning—balance between design and application value is needed.

The project integrates three types of data sources:

• NOAA meteorological data: Temperature, humidity, precipitation, wind speed, drought indices (Palmer/KBDI);
• USGS geographic data: Terrain, land cover, soil, hydrological features;
• NIFC fire data: Historical fire location, size, time, cause, and suppression cost. Data preprocessing includes spatio-temporal alignment, missing value handling, feature construction, and spatial aggregation to county scale.

Feature Engineering:

• Meteorological features: 7/30-day averages, historical同期 anomalies, drought trends, thunderstorm frequency;
• Geographic features: Terrain complexity, fuel load, human accessibility, historical fire frequency;
• Temporal features: Season encoding, annual trends, climate change signals. Model Selection: Logistic regression (baseline), Random Forest (interpretable), XGBoost/LightGBM (high performance), neural networks (complex patterns). Validation Method: Spatial block cross-validation to avoid data leakage (adjacent counties have correlated risks).

Evaluation Metrics: AUC-ROC (discrimination ability), Precision-Recall (class imbalance), spatial accuracy (high-risk identification), calibration (consistency between probability and reality). Interpretability: SHAP values (important factors), partial dependence plots (non-linear relationships), spatial visualization (risk maps). Application Scenarios:

• Resource pre-positioning: Firefighting force deployment, air resource allocation;
• Fire prevention: Enhanced patrols, prescribed burns, public warnings;
• Insurance pricing: Risk assessment, premium adjustment;
• Urban planning: Firebreaks, development restrictions, building standards.

Current limitations and improvements:

1. Spatial resolution: County-level granularity is coarse; need to explore grid-level prediction;
2. Temporal dynamics: Dependent on static features; need to integrate real-time meteorological data streams;
3. Fire types: Lightning-caused and human-caused fires have different drivers; should model separately;
4. Climate change: Historical data cannot capture new risks; need to introduce climate scenarios;
5. Model update: Need to establish monitoring and retraining mechanisms to adapt to pattern evolution.

Scalability:

• Geographic expansion: Applicable to western U.S. states (California, Oregon) and global high-risk areas (Australia, Mediterranean);
• Disaster expansion: Extend to flood, hurricane, landslide prediction;
• Data source expansion: Integrate satellite remote sensing (MODIS/VIIRS), social media, power line data;
• Model improvement: Explore spatio-temporal deep learning (ConvLSTM, GNN) to capture spatial dependencies. Related Ecosystem: FWI (Canadian Fire Weather Index), NFDRS (U.S. Rating System), MODIS/VIIRS fire point data, Wildfire Risk to Communities (community assessment tool), FireMap (California real-time spread prediction).