Human-Interpretable Rules for Predicting Vegetation Fires in Kalimantan Using Meteorological Data with Machine Learning

Vegetation fires are recurring issues in Kalimantan, Indonesia. Real-time predictions are essential to protect the environment. This study aims to determine the important environmental parameters, design machine learning models for predicting fires, and generate human-interpretable rules based on meteorological data and satellite imagery in Kalimantan, Indonesia. Important features of meteorological data become the input variables, and the fire radiative power from satellite-sourced data becomes the target variable. Support vector machine, decision tree, K-nearest neighbor, logistic regression, random forest, and gradient boosting models were tested for predicting fires. Based on the features' importance, the maximum temperature, average relative air humidity, sunshine duration, average wind speed, and precipitation are important for fire prediction. The gradient boosting model performs best, with a score of 96% for training data and 91% for test data. Using the model, 68 human interpretable rules of important features with a high fire risk were generated. Using these rules, the fire occurrence prediction can be done manually without machine learning expertise. This finding may need to be verified by the expert for its accuracy. We hope that this research can be expanded to cover broader geographical scopes.