Algerian Forest Fires — FWI Prediction

A regression-based machine learning project that predicts the Fire Weather Index (FWI) using meteorological and fire-index features from the Algerian Forest Fires dataset. Four regression models are trained, tuned, and compared.

---

Table of Contents

• Overview
• Dataset
• Project Structure
• Setup
• Usage
• Models
• Results

---

Overview

The Fire Weather Index (FWI) is a numeric rating of fire intensity. This project:

1. Loads and preprocesses the Algerian Forest Fires dataset
2. Removes highly correlated features (threshold: 0.85)
3. Scales features using StandardScaler
4. Trains and evaluates four regression models with cross-validated hyperparameters
5. Produces visualisations for correlation analysis, scaling effects, and model comparison

---

Dataset

File: Algerian_forest_fires_dataset_cleaned.csv

The dataset covers two regions of Algeria (Bejaia and Si-Bel-Abbes) and contains daily weather observations and fire index components. The cleaned version used here has had null values and formatting issues resolved.

Feature	Description
Temperature	Noon temperature (°C)
RH	Relative Humidity (%)
Ws	Wind speed (km/h)
Rain	Total rainfall (mm)
FFMC	Fine Fuel Moisture Code
DMC	Duff Moisture Code
DC	Drought Code
ISI	Initial Spread Index
BUI	Buildup Index
FWI	Fire Weather Index (target)

The columns day, month, and year are dropped as they are not used as predictive features.

---

Project Structure

.
├── app.py                                      # Flask prediction app
├── data/
│   └── Algerian_forest_fires_dataset_cleaned.csv
├── models/
│   └── train_model.py                         # Main training script
├── src/                                       # Saved scaler, models, and feature list
├── static/                                    # Frontend assets
├── templates/                                 # Flask templates
├── requirements.txt                           # Python dependencies
└── README.md

---

Setup

Prerequisites: Python 3.9 or higher

1. Clone or download this repository.

2. (Recommended) Create and activate a virtual environment:

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

3. Install dependencies:

   pip install -r requirements.txt

4. Place Algerian_forest_fires_dataset_cleaned.csv inside the data/ directory.

---

Usage

Run the training script:

python3 models/train_model.py

The script will:

• Print MSE, RMSE, and R² for each model to the console
• Display the best cross-validated hyperparameters for Ridge, Lasso, and ElasticNet
• Show a series of plots (correlation heatmap, scaling comparison, model comparison charts)
• Save trained artifacts in src/, including features.json for the Flask app

---

Models

Model	Hyperparameter Tuning
Linear Regression	None
Ridge Regression	alpha selected via 5-fold CV from [0.1, 1.0, 10.0]
Lasso Regression	alpha selected via 5-fold CV from [0.1, 1.0, 10.0]
ElasticNet Regression	alpha and l1_ratio selected via 5-fold CV

Regularisation hyperparameters are tuned automatically using scikit-learn's RidgeCV, LassoCV, and ElasticNetCV — no separate grid search step is needed.

---

Results

After training, a summary table is printed to the console sorted by R², and two comparison charts are displayed:

• RMSE bar chart — lower is better
• All-metrics grouped bar chart — MSE, RMSE, and R² side by side

These make it straightforward to pick the best-performing model for this dataset.