Subsidence vulnerability indexing using convolutional neural networks based on clustering and regression modeling strategies

Abstract

Land subsidence triggered by groundwater over-abstraction is a topical research activity, and this study investigates the Subsidence Vulnerability Index (SVI) using Convolutional Neural Networks (CNNs). The study predicts SVI using the clustering and regression modeling strategies, which have identical aims of decreasing the inherent subjectivity within the basic ALPRIFT but distinct formulations. Both strategies incorporated ALPRIFT data layers, referring to hydrogeological and physical settings of a study area. Also, subsidence maps were incorporated in CNN formulations using 15 SAR (ALOS-2 PALSAR-2 Synthetic Aperture Radar) images. The formulations were implemented in the Tabriz aquifer, northwest of Iran, suffering a severe decline in the water table. The results provide evidence that both CNNs increase the accuracy compared to the basic ALPRIFT as per the ROC curve and AUC values. The Clustering-CNN has slightly better performance (AUC = 0.87) compared to the Regression-CNN (AUC = 0.83). Also, both CNNs have approximately identical spatial patterns and hotspots. The trained and tested CNNs on pixels with high-reliability of InSAR (Interferometric Synthetic Aperture Radar) results show their capability in predicting SVI on pixels with low-reliability.