Sujinsudevan/Algerian-Forest-Fire-Risk-Modeling

# Algerian Forest Fires – Regression Analysis on FWI ## Project Overview This project predicts the **Fire Weather Index (FWI)** — a continuous measure of fire risk — from daily meteorological and fire-weather variables for two regions in Algeria. **Objective:** Build and evaluate regression models to estimate **FWI** given meteorological and fire index data, while dealing with strong feature correlations. **Dataset:** - Year: 2012 - Regions: **Bejaia** (coastal) and **Sidi-Bel Abbes** (inland) - Source: Algerian Forest Fires Dataset (UCI Repository) - Features include: - Temperature, Relative Humidity (RH), Wind Speed (Ws), Rainfall (Rain) - Fire-weather indices: FFMC, DMC, DC, ISI, BUI - Region code (0 or 1) Target variable (**y**): `FWI` (Fire Weather Index) --- ## Data Cleaning & Preparation Steps performed: 1. Removed repeated header rows from the merged dataset. 2. Dropped extra text rows and NaN entries. 3. Added `Region` column → **0** = Bejaia, **1** = Sidi-Bel Abbes. 4. Converted `Classes` to numeric (not used as `y` in this regression task). 5. Ensured all features were numeric. 6. Verified no duplicate rows remained. **Final dataset shape**: `243 rows × 13 columns` --- ## Exploratory Data Analysis – Key Insights - Strong positive correlation with `FWI`: `ISI`, `BUI`, `FFMC`, and `Temperature` - Negative correlation: `RH` (humidity) and `Rain` — higher moisture reduces fire risk - High multicollinearity among predictors → suitable for regularized regression models - Inland region exhibits slightly higher extreme FWI values --- ## Model Building ### Why Regularized Linear Models? - Strong predictor correlations → plain Linear Regression can overfit - **Lasso (L1)** → feature selection - **Ridge (L2)** → keeps all variables, shrinks coefficients - **Elastic Net (L1 + L2)** → balances both approaches ### Models Used: 1. Linear Regression 2. Lasso Regression (`Lasso` & `LassoCV`) 3. Ridge Regression (`Ridge` & `RidgeCV`) 4. Elastic Net (`ElasticNe …