Integrating Machine Learning with Metaheuristic Optimization for Accurate and Interpretable Land Subsidence Susceptibility Mapping in Semi-Arid Regions

Land subsidence, a significant environmental and geotechnical hazard, poses severe threats to infrastructure, agriculture, and water resources, particularly in semi-arid regions. Traditional approaches to subsidence mapping often face challenges in effectively capturing the complex interactions between natural and human-induced factors. This study overcomes these limitations by integrating advanced machine learning with metaheuristic optimization techniques, significantly enhancing the accuracy and reliability of land subsidence susceptibility mapping. The primary purpose of this research is to develop more reliable and interpretable models for land subsidence prediction in Arzuiyeh County of Iran by combining K-Nearest Neighbor (KNN) with two powerful metaheuristic algorithms: the Invasive Weed Optimization (IWO) and the Pelican Optimization Algorithm (POA). The innovation lies in the use of these algorithms to optimize the KNN model, enhancing its predictive power and interpretability. Three models—KNN, KNN-IWO, and KNN-POA—were developed and tested in a semi-arid region. The KNN-POA model consistently outperformed the others, achieving an Area Under the Curve (AUC) of 0.957, which indicates a high level of accuracy in predicting land subsidence-prone areas. This was followed by the KNN-IWO model with an AUC of 0.929, while the base KNN model showed the lowest AUC of 0.867. The reduction in RMSE and MAE values across the models further highlights the improvements brought by the metaheuristic optimizations, with KNN-POA achieving the lowest errors, both in the training and testing phases. © 2025 Elsevier B.V., All rights reserved.