Projects

Human-Factor-Based Wildfire Risk Prediction Model

Period: Mar 2025 – Jun 2025

Overview

Based on the fact that 71% of wildfires in the past 10 years were caused by human factors, this project aims to predict high-risk areas by training a CNN on fused data from satellite imagery and local residential (house-location) information.

Pipeline

1. Collect Sentinel-2 imagery for the target region.
2. Load and rasterize the residential (housing) shapefile
3. Stack the rasterized mask with the original imagery to form 4-channel images.
4. Generate sliding-window patches (64×64).
5. Split patches into positive/negative using historical wildfire coordinates.
6. Train the CNN on the labeled pathces and run inference to produce riks predictions.

Inference

• Generated a heatmap from the model’s predictions.

Recall / Confusion matrix

• Achieved a recall of 76.8% at a decision threshold of 0.9.

Planning

• Integrate additional human factors with environmental variables for better feature sets.
• Apply domain adaptation techniques to mitigate positive bias caused by data scarcity.