geek-mayank/algerian-forest-fire-prediction

### Algerian Forest Fires — EDA ➜ Modeling (End-to-End) - **Dataset**: 244 daily records from two Algerian regions (Bejaia, Sidi Bel-abbes), Jun–Sep 2012, 11 weather/FWI features + class. - **Cleaning**: - Fixed headers/whitespace and types; added `Region` flag. - Dropped NaNs and a separator row → **243 rows** retained. - Exported `Algerian_forest_fires_dataset_CLEANED.csv`. - **EDA Highlights**: - **Distributions**: Weather and FWI components show meaningful spread; `FWI` is right-skewed with outliers. - **Correlations**: Strong relationships among FWI components; `FWI` aligns highly with `FFMC`, `ISI`, `BUI`, `DMC`. - **Class view**: Converted `Classes` → 0 (not fire), 1 (fire) for EDA; seasonal/region patterns visible in monthly counts. - Visuals: Histograms, correlation heatmap, boxplots (FWI), region-wise monthly fire counts. - **Feature Prep for Modeling**: - Dropped `day`, `month`, `year`. - Removed highly correlated predictors (|r| > 0.85) to reduce multicollinearity. - Standardized features with `StandardScaler`. - **Task**: Predict **Fire Weather Index (FWI)** as a regression problem. - **Models & Test Performance**: | Model | MAE | R² | |------------------|-----|-------| | Linear Regression| 0.547 | 0.9848 | | Ridge | 0.564 | 0.9843 | | Lasso | 1.133 | 0.9492 | | LassoCV | 0.620 | 0.9821 | | ElasticNet | 1.882 | 0.8753 | | ElasticNetCV | 0.658 | 0.9814 | - **Outcome**: Simple linear models (Linear/Ridge) performed best with very high explanatory power on test data. - **Artifacts**: Saved `scaler.pkl` and `ridge.pkl` for deployment. FWI helps quantify fire risk; with lightweight preprocessing and regularized linear models, we achieved strong, stable performance suitable for operationalization. #MachineLearning #DataScience #EDA #Regression #MLOps #Algeria #ForestFires #FWI