wildfirez

Wildfire Size Classification Project

Predicting wildfire size classes using machine learning on the FPA FOD (Fire Program Analysis Fire-Occurrence Database) containing 1.88 million US wildfire records from 1992-2015.

Project Overview

This project builds an ordinal classification model to predict fire size categories:

• Small (0-9.9 acres): Original classes A + B
• Medium (10-299 acres): Original classes C + D
• Large (300+ acres): Original classes E + F + G

Key Features

• Ordinal-aware classification: Leverages the natural ordering of fire size classes
• Geospatial features: Coordinate clustering, regional binning, distance metrics
• Temporal features: Cyclical encoding of month/day, fire season indicators
• Class imbalance handling: Balanced class weights for rare large fire events
• Interpretable results: SHAP feature importance analysis

Project Structure

wildfires/
├── config/
│   ├── __init__.py            # Package init
│   └── config.py              # Configuration settings
├── data/
│   └── processed/             # Processed parquet files (train/test splits)
├── models/                    # Saved model artifacts
│   ├── best_params.json       # Tuned hyperparameters
│   ├── model_metadata.joblib  # Feature names and metrics
│   └── wildfire_model.txt     # Trained LightGBM model
├── reports/
│   └── figures/               # Visualizations and metrics
├── scripts/
│   ├── 01_extract_data.py     # Extract SQLite → Parquet
│   ├── 02_eda.py              # Exploratory data analysis
│   ├── 03_preprocess.py       # Data preprocessing
│   ├── 04_feature_engineering.py  # Feature creation
│   ├── 05_train_model.py      # Model training
│   ├── 06_evaluate.py         # Model evaluation
│   └── 07_predict.py          # Prediction pipeline
├── run_pipeline.py            # Run full or partial pipeline
├── requirements.txt           # Dependencies
├── .gitignore                 # Git ignore rules
└── README.md

Getting Started

Prerequisites

• Python 3.9+
• SQLite database file (FPA_FOD_20170508.sqlite)

Installation

1. Clone/download the repository
2. Create a virtual environment:

   python -m venv venv
   venv\Scripts\activate  # Windows
   # source venv/bin/activate  # Linux/Mac
   

3. Install dependencies:

   pip install -r requirements.txt
   

4. Place the SQLite database file in the project root

Running the Pipeline

Using the pipeline runner (recommended):

# Run full pipeline
python run_pipeline.py

# Skip EDA step
python run_pipeline.py --skip-eda

# Run with hyperparameter tuning
python run_pipeline.py --tune

# Resume from a specific step (1-7)
python run_pipeline.py --from-step 5

Or execute scripts individually:

# 1. Extract data from SQLite
python scripts/01_extract_data.py

# 2. Exploratory data analysis (generates plots)
python scripts/02_eda.py

# 3. Preprocess data
python scripts/03_preprocess.py

# 4. Feature engineering
python scripts/04_feature_engineering.py

# 5. Train model (add --tune for hyperparameter tuning)
python scripts/05_train_model.py
# python scripts/05_train_model.py --tune  # With Optuna tuning

# 6. Evaluate model
python scripts/06_evaluate.py

# 7. Make predictions
python scripts/07_predict.py --lat 34.05 --lon -118.24 --state CA --cause "Lightning"

Model Details

Features Used

• Temporal: Month, day of week, season, fire season indicator (cyclically encoded)
• Geospatial: Lat/lon coordinates, regional clusters (K-means), coordinate bins
• Categorical: State, fire cause, reporting agency, land owner
• Year: Fire year, years since 1992

Algorithm

• LightGBM gradient boosting for multi-class classification
• Class weights to handle imbalanced data (~90% small fires)
• Linear weighted Cohen's Kappa for ordinal evaluation

Expected Performance

• Balanced Accuracy: ~65-75%
• Macro F1 Score: ~0.45-0.55
• Large fire detection is challenging due to class imbalance

Evaluation Metrics

For ordinal classification, we prioritize:

• Macro F1: Equal importance to all classes
• Balanced Accuracy: Accounts for class imbalance
• Linear Weighted Kappa: Penalizes predictions far from true class

Output Files

After running the pipeline:

• data/processed/: Parquet files for train/test splits
• models/wildfire_model.txt: Trained LightGBM model
• models/model_metadata.joblib: Feature names and metrics
• reports/figures/: Visualizations (confusion matrix, SHAP plots, etc.)

Data Source

Fire Program Analysis Fire-Occurrence Database (FPA FOD)

• 1.88 million geo-referenced wildfire records
• Period: 1992-2015
• 140 million acres burned
• Source: US federal, state, and local fire organizations

License

This project uses publicly available government data.