This project analyzes the risk of forest fires in Algeria using meteorological data and fire weather indices collected from June to September 2012 across two forested regions: Bejaia and Sidi Bel-Abbas. Motivated by Algeria's susceptibility to forest fires due to its arid climate and landscape, the research repurposes binary fire occurrence data into continuous risk probabilities.
The dataset comprises 243 observations capturing discrete meteorological features (Temperature, Relative Humidity, Wind Speed, Rain) and continuous fire weather indices (FFMC, DMC, DC, ISI, BUI, FWI). A continuous response variable for regression analysis was derived using a k-nearest neighbors (kNN) approach to estimate fire risk probabilities from local neighborhood observations.
Three regression techniques were evaluated:
- Multiple Linear Regression
- Ridge Regression
- Generalized Additive Models (GAM)
- Fire weather indices outperform meteorological features in predicting fire risk.
- Ridge Regression with fire weather indices achieved robust predictive performance, maintaining lower residual trends and homoskedasticity compared to other methods.
- Generalized Additive Models (GAM) showed high predictive power but displayed notable residual clustering, suggesting unexplained variability.
- Meteorological features alone proved inadequate in effectively explaining forest fire occurrences.
Ridge regression based on fire weather indices is most effective for modeling forest fire risk. Future studies should investigate individual feature importance, interaction effects, and potential time-series patterns to enhance predictive accuracy.
This project demonstrates the potential of regression modeling to improve forest fire prediction, contributing to more effective prevention strategies.