Susceptibility assessment of earth fissure related to groundwater extraction using machine learning methods combined with weights of evidence

The susceptibility of a region to the occurrence of earth fissures is often used to assess the probability of geohazards across an area. The main objective of this study is to discuss and explore machine learning methods for earth fissure susceptibility assessment, including the single machine learning method and the ensemble model. A total of ten affecting factors including elevation, slope, topographic wetness index, rainfall, drawdown of groundwater level, the thickness of Quaternary sediments, distance from rivers, distance to faults, normalized difference vegetation index, and land use were selected. The weight of evidence (WoE) method was first used to determine the quantitative relationship between an earth fissure and its related parameters. The WoE, support vector machine learning combined with the WoE (SVM +WoE), and the random forest combined with the WoE (RF+ WoE) model were then used to classify earth fissure susceptibility. The area under the curve and root-mean-squared error was used to evaluate the three methods and to determine the most optimal approach for earth fissure susceptibility map. The results indicated that the RF+ WoE model had the highest predictive accuracy, followed by the SVM+WoE and the WoE models. The study area was finally classified into regions with very high, high, moderate, low, and very low susceptibility, accounting for 11.20%, 15.66%, 24.13%, 32.60%, and 16.07% of the area. Susceptibility mapping can apply machine learning methods combined with the WoE method for earth fissure assessment.