Sensitivity study of multi-field information maps of typical landslides in mining areas based on transfer learning

Abstract

The main purpose of this study is to analyze the main influencing factors of the landslide in the coal mine area and, on this basis, establish the sensitivity zoning model of the landslide. Considering the difficulty to obtain the expected results by using machine learning under the condition of lacking data, the typical landslide is used as the data basis, that is, the Fenxi coal mine and Xishan Bujiu coal mine are selected as the coal mining landslide points. Various factors, such as goaf, land subsidence, slope structure, formation lithology, and various indicators are used as input data sources, and artificial neural network (ANN) datasets are used for training to establish a pre-training model. Using the pre-training model, the mining landslide sensitivity evaluation model based on transfer learning is established. In order to demonstrate the performance of transfer learning more intuitively, the neural network is introduced to evaluate the evaluation model. The test results show that transfer learning can achieve a transfer effect higher than 0.95, and the regional distributions of highest landslide sensitivity calculated based on self-transfer learning, direct push transfer learning, and inductive transfer learning are 31.33, 35.50, and 33.75%, respectively, which further deduced that inductive transfer learning can be used for evaluating an LSP model.