Machine Learning-Based Los Angeles Wildfire Risk Assessment and Evacuation Route Optimization System

This project addresses the high incidence of wildfires in Los Angeles County. Using the Random Forest algorithm, it classifies wildfire risks across all 2493 census tracts in the county. By integrating the Social Vulnerability Index (SVI), infrastructure distribution, and geographic features, it constructs a community vulnerability assessment and evacuation route optimization plan, providing scientific decision support for disaster prevention and emergency response.

Los Angeles County is located in southern California, with a dry climate and dense vegetation. Wildfire threats are severe from summer to autumn, and their frequency and intensity have increased in recent years due to climate change. Traditional risk assessments rely on empirical judgment, with limitations such as single indicators failing to fully reflect comprehensive vulnerability and lack of systematic integration of key factors like infrastructure and population characteristics. There is an urgent need for data-driven, refined analysis tools.

Data Integration

We use the CDC's Social Vulnerability Index (SVI) to measure community response capacity from four dimensions: socioeconomic status, household characteristics, racial/ethnic composition, and housing/transportation conditions. We also construct spatial features, including infrastructure accessibility (distance to fire stations/hospitals, fire hydrant density) and traffic network analysis (road length/density, area of the region).

Model Selection

We selected the Random Forest algorithm (200 decision trees, maximum depth of 10 layers, balanced class weights), with the target variable being three-class risk labels (low/medium/high). Generalization ability is ensured through 5-fold cross-validation and an 80/20 training-test split. SHAP values are used to analyze feature contributions and enhance interpretability.

The system includes a complete toolchain: data preprocessing scripts integrate SVI, infrastructure, and traffic data; feature engineering generates standardized features and GeoJSON spatial files; after model training, it is serialized into a pickle file; visualization outputs interactive HTML maps such as risk heatmaps, infrastructure overlay maps, and evacuation route maps.

Value for Los Angeles County: Precisely identify high-risk communities to deploy resources in advance; optimize evacuation routes by combining road networks and medical facilities; support land use planning and infrastructure construction. The methodology can be transferred to other disaster-prone areas, requiring the construction of localized feature datasets and model iteration.

Developed using Python 3.10+, relying on libraries such as pandas, scikit-learn, and geopandas. Reproduction steps: Configure environment → Data exploration → Feature engineering → Model training → Visualization generation; provides functions for historical wildfire comparison and random fire point simulation verification. For applications in other regions, attention should be paid to data quality control, feature set adjustment, and model parameter optimization.

Facing the challenges of climate change, data-driven disaster risk assessment is an important tool for urban resilience construction. This project demonstrates the application potential of machine learning in the field of public safety, providing a useful exploration for building a more intelligent and equitable urban emergency management system.