Advancing the LightGBM approach with three novel nature-inspired optimizers for predicting wildfire susceptibility in Kauaʻi and Molokaʻi Islands, Hawaii

Abstract

This study developed three hybrid light gradient boosting machine (LightGBM) models using novel metaheuristic algorithms (golden jackal optimization [GJO], the pelican optimization algorithm [POA], and the zebra optimization algorithm [ZOA]) to predict wildfire susceptibility on Kauaʻi and Molokaʻi islands, Hawaii. Thirteen geo-environmental variables were employed as potentially influential variables to predict wildfire susceptibility, while 1641 and 136 recorded wildfire ignition points on Kauaʻi and Molokaʻi islands, respectively, and 1641 and 136 randomly generated non-wildfire locations on the same islands were used as output data. The impact of the independent variables on wildfire susceptibility was interpreted using the Shapley additive explanations (SHAP) method. It was found that ZOA-LightGBM had the highest accuracy (AUC = 0.9314) for the prediction of wildfire susceptibility on Kaua‘i island, followed by GJO-LightGBM (0.9308), POA-LightGBM (0.9303), and LightGBM (0.9228). On Molokaʻi island, ZOA-LightGBM also achieved the highest AUC (0.858), outperforming POA-LightGBM (0.8488), GJO-LightGBM (0.8426), and LightGBM (0.8395). Analysis of the hybrid models revealed that the use of metaheuristic algorithms resulted in the setting of the optimal hyperparameters for the LightGBM model, enhancing its performance in modeling wildfire susceptibility on both islands. The SHAP results indicated that the distance from roads, elevation, land surface temperature, and average annual wind speed had the strongest impact on wildfire susceptibility on Kauaʻi Island, whereas distance from roads, slope, and average annual rainfall were the most important variables on Molokaʻi island.