Modeling the Influence of Groundwater Exploitation on Land Subsidence Susceptibility Using Machine Learning Algorithms

Groundwater over-exploitation in arid and semiarid environments has led to many land subsidence cases. Immense economic losses incurred from land subsidence occurrences prompted many scientists to model this phenomenon. To this end, we used three machine learning models, boosted regression trees (BRTs), generalized additive model (GAM), and random forest (RF), together with four anthropological and geo-environmental predictors, to produce a spatial prediction map across land subsidence-prone area in the south of Iran. The inventory map and preparatory thematic layers were generated through extensive field surveys, using Google Earth images, local information, and organizational archives. The results revealed that the GAM significantly out-performs the BRT in terms of high goodness of fit (84.3% vs. 80.2%) and predictive power (81.6% vs. 70.1%). The RF model, as a benchmark model, showed slightly higher goodness of fit (85.45%) compared to the GAM; however, its prediction power was evidently lower than the GAM. Hence, the GAM was found as the best susceptibility model in the study area. According to the relative contribution test, the drawdown of groundwater level with 77.5% contribution was found to be the main causative predictor of land subsidence occurrence, followed by lithology (19.2%), distance from streams (2.5%), and altitude (0.8%). The results of the GAM suggest that almost 31.6% of the study area is highly susceptible zone to land subsidence occurrence, which can be of interest for further pragmatic actions.