A novel HHO-RSCDT ensemble learning approach for forest fire danger mapping using GIS

Abstract

Accurate prediction models for spatial prediction of forest fire danger play a vital role in predicting forest fires, which can help prevent and mitigate the detrimental effects of such disasters. This research aims to develop a new ensemble learning model, HHO-RSCDT, capable of accurately predicting spatial patterns of forest fire danger. The HHO-RSCDT method combines three distinct components, namely Random Subspace (RS), Credal Decision Tree (CDT), and Harris Hawks Optimizer (HHO). Herein, RS generates a series of subspace datasets, which are subsequently utilized to produce individual CDT classifiers. Then, HHO optimizes the ensemble model, enabling the model to achieve higher predictive performance. The model was trained and validated using a Phu Yen province, Vietnam dataset. The dataset includes 306 forest fire locations and ten influencing factors from the study province. The results showed the capability of the HHO-RSCDT model in predicting forest fire danger, with an accuracy rate of 83.7%, a kappa statistic of 0.674, and an AUC of 0.911. A comparison between the HHO-RSCDT model and two state-of-the-art machine learning methods, i.e., support vector machine (SVM) and random forest (RF), indicated that the HHO-RSCDT model could perform better, making it a valuable tool for modeling forest fire danger. The forest fire danger map produced using this novel model could be a new tool for local authorities in the Phu Yen province, assisting them in managing and protecting the forest ecosystem. By providing a detailed overview of the are as most susceptible to forest fires, the map can help authorities to develop targeted and effective forest management strategies, such as focusing on areas with high fuel loads or implementing controlled burning programs.