Fire Risk Prediction System

A machine learning system for wildfire risk prediction using ensemble methods that combine XGBoost and PyTorch LSTM models. The system analyzes weather time-series data to forecast fire risk 6 hours in advance.

Project Overview

This system predicts wildfire risk in California using:

• Weather station data from CIMIS (California Irrigation Management Information System)
• Historical fire records from Wikipedia and CAL FIRE
• Ensemble modeling combining XGBoost and attention-based LSTM
• 6-hour prediction horizon using 24-hour weather sequences

Architecture

1. XGBoost Classifier

• Gradient boosting for feature interactions
• Station-based splitting to prevent data leakage
• Cyclical time encoding for seasonal patterns (sin/cos)
• Lag features (previous day weather conditions)
• Class balancing for rare fire events (~5% positive class)

2. Attention-based LSTM (PyTorch)

• Bidirectional LSTM with custom attention mechanism
• 24-hour sequences → 6-hour prediction horizon
• Focal loss for imbalanced data
• Early stopping with validation monitoring

3. Ensemble Methods

• Weighted Average: Grid search optimization
• Stacked Meta-Learning: Logistic regression meta-classifier
• Calibration analysis and reliability curves

Quick Start

Prerequisites

pip install xgboost joblib scikit-learn matplotlib seaborn
pip install torch pandas numpy

Usage

1. Data Processing: dataset/colab-conditions-df-cleaning.ipynb
2. XGBoost Training: fire_risk_xgboost_implementation.ipynb
3. Ensemble Evaluation: ensemble_xgboost_lstm.ipynb

Project Structure

fire_risk_prediction/
├── fire_risk_xgboost_implementation.ipynb    # XGBoost model
├── ensemble_xgboost_lstm.ipynb               # PyTorch LSTM + Ensemble
├── dataset/
│   ├── colab-conditions-df-cleaning.ipynb   # Data preprocessing
│   └── conditions_df.csv                     # Processed weather/fire data
└── README.md

Key Features

Feature Engineering

• Cyclical Encoding: Month/day-of-year as sin/cos pairs
• Station Statistics: Historical fire rates per weather station
• Lag Features: Previous 24-hour weather conditions
• Region Encoding: Geographic information via one-hot encoding

Data Processing

• Station-based splitting: 80% stations for training, 20% for testing
• Class balancing: scale_pos_weight for XGBoost, focal loss for LSTM
• Sequence construction: 24-hour windows for LSTM input
• Missing value imputation: Median imputation for numerical features

Model Performance

Model	Approach	Key Strength
XGBoost	Feature-based	Complex feature interactions
LSTM	Sequential	Temporal pattern recognition
Ensemble	Combined	Best of both approaches

Note: Run the notebooks to see actual performance metrics

Technical Details

XGBoost Configuration

xgb_params = {
    "n_estimators": 500,
    "max_depth": 5,
    "learning_rate": 0.05,
    "subsample": 0.8,
    "colsample_bytree": 0.8,
    "tree_method": "hist",
    "scale_pos_weight": "auto"  # Handles class imbalance
}

LSTM Architecture

• Input: 24-hour weather sequences
• Hidden: 64-unit bidirectional LSTM
• Attention: Custom attention mechanism
• Output: Single probability (fire risk in next 6 hours)
• Loss: Focal loss for imbalanced classification

Data Sources

• Weather Data: CIMIS (California Irrigation Management Information System)
• Fire Records: Wikipedia fire tables + CAL FIRE historical data
• Time Period: 2018-2020
• Geographic Coverage: California weather stations

Use Cases

• Early Warning: 6-hour advance fire risk alerts
• Resource Planning: Firefighting resource deployment
• Research: Weather pattern analysis for fire prediction

Evaluation

The system includes comprehensive evaluation:

• ROC and Precision-Recall curves
• Confusion matrices and classification reports
• Feature importance analysis (XGBoost + SHAP)
• Attention weight visualization (LSTM)
• Model calibration analysis

Future Enhancements

• Real-time weather data integration
• Extended prediction horizons (12-24 hours)
• Additional weather variables (soil moisture, drought indices)
• Spatial modeling for fire spread prediction