Geographic-based sampling of absence data for enhanced machine learning model with multi-strategy reinforcement GWO in land subsidence susceptibility mapping

The spatial sampling of absence data can improve the performance of machine learning models in predicting land subsidence susceptibility (LSS). Existing methods face challenges in quantifying the reliability of absence data, as well as in sampling in a way that generates a representative dataset for the mapped land subsidence occurrence areas. In this study, a new hybrid sampling approach based on the First and Third Laws of Geography is proposed to generate more reliable absence data. To evaluate the proposed sampling approach, Support Vector Regression (SVR) and Convolutional Neural Network (CNN) were employed to generate LSS maps, and the models were optimized through simultaneous feature selection and hyperparameter tuning. For this purpose, a multi-strategy reinforcement Grey Wolf Optimizer guided by Q-learning (MSQLGWO) was employed. Additionally, a transformation function was integrated into this algorithm to enable joint optimization of feature subsets and hyperparameters (MSQLTGWO). The results indicate that training the optimized models with the dataset generated using the proposed sampling approach generally yielded better performance than the other sampling datasets. The SVR-MSQLTGWO model achieved an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.9853, Root Mean Squared Error (RMSE) of 0.2449, and R² of 0.7600. The CNN-MSQLTGWO model trained using the dataset generated by the proposed sampling approach achieved the lowest RMSE of 0.2328 and the highest R² of 0.7831, with an AUC-ROC of 0.9833. The generated LSS maps also showed strong spatial agreement with the mapped land subsidence occurrence areas in the study area, with 97.54% and 96.57% of these areas located in the high and very high susceptibility classes for the SVR-MSQLTGWO and CNN-MSQLTGWO models, respectively. Finally, the most influential features were identified through SHapley Additive exPlanations (SHAP) analysis and comparison with the feature-selection results. Therefore, the findings of this study suggest that employing the proposed sampling approach, combined with an improved metaheuristic algorithm, can enhance the accuracy and reliability of LSS mapping within the evaluated study area.