Forecasting fire risk with machine learning and dynamic information derived from satellite vegetation index...

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Forestry and Fire Management

Background:

• Traditional fire risk maps rely on static data, neglecting crucial long-term vegetation dynamics.
• Understanding vegetation's cumulative dryness and density is vital for accurate fire prediction.

Purpose of the Study:

• To assess the impact of two satellite-derived metrics, mean woody vegetation density (NDVI_W) and its trend (NDVI_T), on fire risk mapping.
• To determine if these metrics improve the accuracy of fire risk predictions using machine learning models.

Main Methods:

• Decomposed satellite-derived NDVI time-series for Mediterranean woodlands to derive NDVI_W and NDVI_T.
• Employed three machine learning algorithms (Logistic Regression, Random Forest, XGBoost) to test the predictive power of the new metrics.
• Analyzed the 2007 wildfires in Greece to validate the proposed methodology.

Main Results:

• The XGBoost model, accounting for variable interactions, yielded the best fire risk mapping performance.
• NDVI_W significantly improved model performance, while NDVI_T was only significant in areas with high NDVI_W.
• An interaction effect was observed: long-term dryness impacts fire risk primarily in dense vegetation areas.

Conclusions:

• The proposed method, utilizing NDVI_W and NDVI_T, offers more accurate fire risk maps compared to conventional approaches.
• These dynamic vegetation metrics provide valuable data for improving fire behavior models and pre-fire management strategies.