Enhancing land subsidence susceptibility mapping using deep tabular learning optimization with metaheuristic algorithms

Abstract

Land subsidence, a gradual sinking of the Earth’s surface, poses significant threats to urban and rural landscapes, leading to severe environmental, social, and economic consequences. Land subsidence monitoring and susceptibility mapping are important for urban planning and geohazard assessment. Previous research on Land Subsidence Susceptibility Modeling (LLSM) often relied on field survey data and traditional deep learning methods and lacked optimal hyperparameter tuning. To address these gaps, this study employed a Deep Tabular Learning algorithm, specifically TabNet (Attention interpretable tabular learning), optimized using three metaheuristic algorithms: Particle Swarm Optimization (PSO), Cuckoo Search (CS), and Whale Optimization Algorithm (WOA) and land subsidence detection in Kurdistan Province, Iran, from 2015 to 2022, utilizing Interferometric Synthetic Aperture Radar (InSAR) time analysis. The spatial database for modeling integrated land subsidence occurrence areas was derived from InSAR data with 15 critical criteria, including topographic, climatic, geological, and land-cover information. The modeling results and susceptibility maps revealed that the TabNet-CS model exhibited the highest accuracy in predicting land subsidence susceptibility, with a Root Mean Square Error (RMSE) of 0.223, Mean Absolute Error (MAE) of 0.125, and Area Under the Curve (AUC) of 0.956. The TabNet-PSO model demonstrated good performance, with RMSE = 0.241, MAE = 0.143, and AUC = 0.941. The TabNet-WOA model also showed promising results with RMSE = 0.255, MAE = 0.154, and AUC = 0.931. Finally, the standalone TabNet model yielded comparatively lower accuracy with RMSE = 0.297, MAE = 0.199, and AUC = 0.92. Integrating metaheuristic algorithms (CS, PSO, and WOA) improved the accuracy of the TabNet model by 3.6, 2.1, and 1.2 %, respectively.