Fire Risk Predictor

A machine learning-based system for predicting wildfire risk in Brazil using environmental and atmospheric data. The project consists of a trained ML pipeline, a FastAPI backend for model inference, and a React frontend for interactive predictions.

Table of Contents

• Overview
• Project Structure
• Dataset
• Machine Learning Models
• API Reference
• Frontend
• Installation
• Deployment
• Environment Variables
• License

Overview

Brazil faces an increasing wildfire crisis, with 278,229 heat spots recorded in 2024 alone, the highest number in 14 years according to INPE. The Amazon region was the most affected biome, with over 140,000 spots representing a 77% increase compared to 2023.

This project addresses the challenge of reactive fire detection by using machine learning to predict where fires are most likely to occur based on environmental conditions, enabling preventive action before fires spread.

Key Features

• Three trained ML models: Random Forest, MLP (Neural Network), and XGBoost
• REST API for batch predictions and model comparison
• Threshold optimization for maximizing F1 score
• Web interface for interactive predictions
• Support for both file upload and sample data testing

Project Structure

fire-risk-predictor/
├── FireRiskPredictor.ipynb    # Training notebook with K-Fold validation
├── api/                        # FastAPI backend
│   ├── main.py                # API endpoints
│   ├── models.py              # Pydantic models
│   ├── model_manager.py       # Model loading and caching
│   ├── preprocessing.py       # Data preprocessing utilities
│   ├── config.py              # Configuration settings
│   ├── Dockerfile             # Container configuration
│   └── requirements.txt       # Python dependencies
├── frontend/                   # React frontend
│   ├── src/
│   │   ├── App.tsx            # Main application component
│   │   ├── components/        # UI components
│   │   └── services/          # API client
│   ├── package.json           # Node.js dependencies
│   └── vite.config.ts         # Vite configuration
└── README.md

Dataset

The project uses data from SISAM (Sistema de Informacoes de Saude Ambiental), containing heat spot records across Brazil since 2003.

Dataset Statistics

• Original dataset: 2.6+ million records
• Training sample: 250,000 balanced records (125k fire, 125k no-fire)
• Test sample: 50,000 records (separate from training)

Features (14 variables)

Feature	Description	Unit
datahora	Date and time of observation	datetime
longitude	Geographic longitude	degrees
latitude	Geographic latitude	degrees
co_ppb	Carbon monoxide concentration	ppb
no2_ppb	Nitrogen dioxide concentration	ppb
o3_ppb	Ozone concentration	ppb
pm25_ugm3	Particulate matter PM2.5	ug/m3
so2_ugm3	Sulfur dioxide concentration	ug/m3
precipitacao_mmdia	Daily precipitation	mm/day
temperatura_c	Temperature	Celsius
umidade_relativa_percentual	Relative humidity	%
vento_direcao_grau	Wind direction	degrees
vento_velocidade_ms	Wind speed	m/s
incendio	Fire occurrence (target)	0 or 1

Preprocessing

During inference, the following columns are excluded from model input:

• Location identifiers: datahora, longitude, latitude
• Categorical data: satelite, pais, estado, municipio, bioma
• Derived features: data_pas, numero_dias_sem_chuva, risco_fogo, frp, time_diff_hours

Machine Learning Models

Training Methodology

• K-Fold Cross Validation (K=3 and K=5)
• Balanced sampling to handle class imbalance
• StandardScaler preprocessing for MLP
• Grid search for optimal classification threshold

Models

Model	Description	Key Parameters
Random Forest	Ensemble of decision trees	Default sklearn parameters
MLP	Neural network with 2 hidden layers	100x100 neurons, max_iter=200
XGBoost	Gradient boosting	n_estimators=100, reg:squarederror

Evaluation Metrics

Models are evaluated using:

• Accuracy
• Precision
• Recall
• F1 Score (primary metric for threshold optimization)
• Confusion Matrix

API Reference

Base URL: http://localhost:8000

Endpoints

Health Check

GET /health

Returns API status and model loading state.

List Models

GET /models

Returns available models with metadata.

Batch Prediction

POST /predict/batch

Upload a CSV file for batch predictions with a single model.

Parameters:

• file: CSV file (multipart/form-data)
• model_name: RF, MLP, or XGBoost (default: RF)
• threshold: Classification threshold 0.0-1.0 (optional)

Prediction with Sample Data

POST /predict/sample

Run predictions using cached sample dataset.

Parameters:

• model_name: RF, MLP, or XGBoost
• size: Sample size 100-50000 (default: 10000)
• threshold: Classification threshold (optional)

Compare Models

POST /compare

Compare all models on the same dataset.

Parameters:

• file: CSV file (multipart/form-data)
• threshold: Classification threshold (optional)

Compare Models with Sample

POST /compare/sample

Compare all models using cached sample dataset.

Parameters:

• size: Sample size 100-50000
• threshold: Classification threshold (optional)

Optimize Threshold

POST /optimize-threshold

Find optimal classification threshold for a model.

Parameters:

• file: CSV file with ground truth (multipart/form-data)
• model_name: RF, MLP, or XGBoost
• threshold_min: Minimum threshold (default: 0.1)
• threshold_max: Maximum threshold (default: 0.9)
• threshold_step: Step size (default: 0.05)

Optimize Threshold with Sample

POST /optimize-threshold/sample

Find optimal threshold using cached sample dataset.

Body (JSON):

{
  "model_name": "RF",
  "sample_size": 10000,
  "threshold_min": 0.1,
  "threshold_max": 0.9,
  "threshold_step": 0.05
}

Get Sample Data

GET /sample

Retrieve a random sample from the reference dataset.

Parameters:

• size: Sample size 100-50000 (default: 10000)

Feature Columns

GET /feature-columns

Returns the list of required feature columns for prediction.

Frontend

The web interface is built with React 19, TypeScript, and Tailwind CSS v4.

Features

• Model status monitoring
• Data source toggle (file upload or sample data)
• Single model prediction
• Multi-model comparison
• Threshold optimization with visualization
• Confusion matrix display
• F1 score vs threshold charts

Running Locally

cd frontend
npm install
npm run dev

The frontend will be available at http://localhost:5173.

Installation

Prerequisites

• Python 3.11+
• Node.js 18+
• Trained model files (RF.pkl, MLP.pkl, XGBoost.pkl)

Backend Setup

cd api
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

Frontend Setup

cd frontend
npm install

Running Locally

Backend:

cd api
hypercorn main:app --bind 0.0.0.0:8000

Frontend:

cd frontend
npm run dev

Deployment

The project is configured for deployment on Railway with the following services:

• API: Docker container with Hypercorn ASGI server
• Frontend: Static site deployment
• Storage: MinIO bucket for model files and sample dataset

Docker Build (API)

cd api
docker build -t fire-risk-predictor-api .
docker run -p 8000:8000 --env-file .env fire-risk-predictor-api

Environment Variables

Backend (api/.env)

Variable	Description	Default
MINIO_BASE_URL	MinIO server URL	http://bucket.railway.internal:9000
MINIO_ACCESS_KEY	MinIO access key	(required)
MINIO_SECRET_KEY	MinIO secret key	(required)
MINIO_REGION	MinIO region	us-east-1

Frontend (frontend/.env)

Variable	Description	Default
VITE_API_URL	Backend API URL	http://localhost:8000

Tech Stack

Backend

• FastAPI 0.100.0
• Hypercorn 0.14.4
• Pandas 2.1+
• NumPy 2.0+
• Scikit-learn 1.3+
• XGBoost 2.0+
• Pydantic 2.3.0

Frontend

• React 19.2
• TypeScript 5.9
• Vite 7.2
• Tailwind CSS 4.1
• Axios 1.13
• Lucide React (icons)

Contributors

• Gustavo Lelli
• Otavio Coletti

License

This project is developed by MonDesa.