Development of a Prediction Model for Potential Forest and Land Fires using Machine Learning Algorithms Based on Patrol Data

Abstract

Indonesia allocates 120 million hectares or 64% of its land area as forest areas. Indonesia's forests continue to experience deforestation; one of the causes is forest and land fires (karhutla). The government conducts forest and land fire prevention through integrated patrols with the Forest and Land Fire Prevention Patrol Information System (SIPP Karhutla) facility for patrol data management. However, the patrol data are still primarily used for data observation and simple spatial analysis in the spatial module. Patrol data has not been used for further forest and land fire prevention studies. Based on these problems, this research aims to build a prediction model of potential forest and land fires using SVM, Random Forest, and XGBoost algorithms and compare model performance to get the best model. The preprocessing stage uses the SMOTE-ENN method to handle data class imbalance, and the k-fold cross-validation stage and hyperparameter tuning use the random search method. The confusion matrix evaluation method to see the model performance in terms of accuracy is XGBoost (94.81%), Random Forest (90.23%), SVM-linear (79.58%), SVM-polynomial model (73.99%), SVM-rbf (74.26%), and SVM-sigmoid (35.04%). Therefore, the best prediction model is XGBoost (94.81%) with boosting technique. The results of this study have implications for helping early prevention of forest and land fires on the islands of Sumatra and Kalimantan.