Detection of Land Subsidence Using Hybrid and Ensemble Deep Learning Models

Land subsidence (LS) is among the most prominent forms of subsurface erosion and geomorphological hazards. This study used two deep learning (DL) models consisting of the hybrid CNN-RNN and ensemble DL (EDL) merged with two dense models. The main variables controlling LS (consisting of environmental, hydrological, hydrogeological, digital elevation model, and soil characteristics), were used as the input for the predictive DL models. Likewise, to establish the degree of performance of each parameter, different control points have been established. We then trained and tested our DL models using the receiver-operating characteristic-area under curve (ROC-AUC) and precision-recall plots. The measures based on the game theory consisting of permutation feature importance measure (PFIM) and SHapley Additive exPlanations (SHAP) were employed to assess the features relative importance and interpretability of the predictive model output. Our findings show that the ensemble CNN-RNN model performed well with the ROC-AUC curve (0.95) of class 1 (land subsidence) for training data for detecting and mapping land subsidence compared to EDL with the ROC curve (0.93) of class 1 (land subsidence) for training datasets. The CNN-RNN also performed well with the precision-recall curve (0.954) of class 1 for testing data for detecting and mapping land subsidence compared to the EDL model with the precision-recall curve (0.95) of class 1. The results of this research revealed that much of the study area is susceptible to land subsidence. The results of the model sensitivity analysis suggested that the groundwater drop rate is the most sensitive for the model. Based on the SHAP values, the groundwater drop rate was identified as the most contributed feature to the model output. The importance of this study is at a broader level, especially in arid and semiarid environments with similar geomorphological, and climatic conditions.