Machine learning based fire danger assessment framework for Indian forests

We developed a comprehensive fire risk assessment framework for Indian forests, divided into five distinct forest zones (Himalayan, Northeast, Central India, Deccan, and Western Ghats) characterized by diverse climatic conditions and forest types. This framework focused on three primary triggering factors: weather, fuel availability, and anthropogenic ignition.

For the weather factor, we considered the Fire Weather Index (FWI) module of the Canadian Forest Fire Danger Rating System with ECMWF's ERA5 reanalysis as meteorological inputs over the period 2003-2021. As fire weather is a dominant factor in causing fires, we developed a robust system to predict fire weather danger. We evaluated the simulated FWI against MODIS active fire data and observed that FWI was a good enough metric for fire weather danger assessment. FWI was categorized into five danger classes through an ensemble approach based on logistic regression, FWI percentiles, percentage of fires, and K-means clustering. We introduced machine learning techniques to reduce the subjective decisions in these methods. This increased the efficiency of the danger rating system to detect fire probability well by 30-50%. A rigorous evaluation of the danger classes revealed that there was no overlap of central tendencies between different methods in the ensemble. The defined danger classes demonstrated coherent values for evaluative parameters, with a consistently high hit rate, low hits due to chance, moderate correct rejections, and an acceptable false alarm ratio.

Addressing fuel availability, we used vegetation indices (MODIS normalized difference and enhanced vegetation indices) and topographic features (aspect, elevation and slope from FLDAS land surface model). The anthropogenic ignition factor consisted of population density and land use information. In India, fragmented forests cohabitate with human settlements and agricultural lands. To quantify the impact of anthropogenic ignition on fire occurrences, we computed the percentage of built-up and agricultural area within each grid cell. We used machine learning predictive algorithms such as multiple linear regression with interactions, support vector machines, decision trees and neural networks to integrate these triggering factors with fire count as the target variable, We selected the highest-performing system as the risk assessment framework.

This country-scale fire risk assessment provides insights into regional exposure variations and serves as a foundational step towards establishing an operational fire risk assessment system for India. This framework will be of help to operational fire management agencies, enabling enhanced prediction of fire danger and informed decision-making.