Accuracy enhancement in land subsidence prediction using lightgbm and metaheuristic optimization

Land subsidence has caused numerous problems and dilemmas in agricultural lands, roads, and power and energy transmission lines for various reasons, including excessive groundwater extraction and climate change. Accurate prediction of land subsidence is crucial for effective risk management and mitigation strategies. However, existing models often struggle with limited predictive accuracy due to suboptimal algorithmic performance and the lack of effective hyperparameter optimization. This study addresses this gap by proposing a novel approach for land subsidence modeling in Arzuiyeh County of Iran, combining the LightGBM (Light Gradient Boosting Machine) algorithm with two advanced metaheuristic optimization techniques: the Imperialist Competitive Algorithm (ICA) and the War Strategy Optimization (WSO). The innovation of this research lies in the hybridization of LightGBM with ICA and WSO to optimize the hyperparameters and enhance the model’s predictive accuracy. To evaluate the effectiveness of the proposed models, comparisons were made with benchmark models, including Random Forest (RF) and Extreme Gradient Boosting (XGBoost). The results show that the LightGBM-ICA model outperforms the standalone LightGBM and LightGBM-WSO models, achieving an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) value of 0.981. The LightGBM-WSO, RF, XGBoost, and baseline LightGBM models achieved AUC values of 0.971, 0.970, 0.937, and 0.922, respectively. The findings demonstrate that combining LightGBM with metaheuristic optimization provides a robust solution to improve land subsidence prediction, offering higher accuracy than existing models. This research contributes to the field of natural hazard modeling by providing an innovative and accurate framework for land subsidence prediction, which can be applied in regions facing subsidence risks. © 2025 Elsevier B.V., All rights reserved.