Land subsidence hazard modeling: Machine learning to identify predictors and the role of human activities

Land subsidence caused by land use change and overexploitation of groundwater is an example of mismanagement of natural resources, yet subsidence remains difficult to predict. In this study, the relationship between land subsidence features and geo-environmental factors is investigated by comparing two machine learning algorithms (MLA): maximum entropy (MaxEnt) and genetic algorithm rule-set production (GARP) algorithms in the Kashmar Region, Iran. Land subsidence features (N = 79) were mapped using field surveys. Land use, lithology, the distance from traditional groundwater abstraction systems (Qanats), from afforestation projects, from neighboring faults, and the drawdown of groundwater level (DGL) (1991-2016) were used as predictive variables. Jackknife resampling showed that DGL, distance from afforestation projects, and distance from Qanat systems are major factors influencing land subsidence, with geology and faults being less important. The GARP algorithm outperformed the MaxEnt algorithm for all performance metrics. The performance of both models, as measured by the area under the receiver-operator characteristic curve (AUROC), decreased from 88.9-94.4% to 82.5-90.3% when DGL was excluded as a predictor, though the performance of GARP was still good to excellent even without DGL. MLAs produced maps of subsidence risk with acceptable accuracy, both with and without data on groundwater drawdown, suggesting that MLAs can usefully inform efforts to manage subsidence in data-scarce regions, though the highest accuracy requires data on changes in groundwater level.