31 / 2023-03-29 14:33:47

Comparative study on the machine learning-based prediction models of developmental sensitivity to subsidence disasters

Subsidence disaster; Sensitivity assessment; Support Vector Machine; Random Forest; Prediction

The zoning prediction of developmental sensitivity to subsidence disasters is very important for the sustainable development of underground coal mining areas. Different sensitivity assessment models usually lead to various prediction results, showing that the model selection has greatly affected the prediction results. Datong Jurassic coal seam mining area in Shanxi Province, China, was selected as the study area. The 70% of 135 subsidence disasters developed in 2011 were used as the training set to construct a sensitivity prediction model, and the remaining data were used as the validation set. A total of 188 subsidence disaster points developed in 2014 were used as the test set for model assessment. The elevation, slope, aspect, topographic relief, distance from the fault, stratigraphic strata, and distance from the mine boundary were adopted as the sensitivity assessment factors. Two commonly used machine learning models of support vector machine (SVM) and random forest (RF) models were selected to carry out zoning prediction of developmental sensitivity to subsidence disasters. The area under the receiver operating characteristic curve (AUC) was used as the assessment index. The model accuracy, prediction accuracy, and rationality of the results were respectively analyzed. The results show that: (1) The polynomial kernel function-based SVM model has the highest prediction accuracy of 76.32% among the four kernel functions, which has been improved by 29.47%, 22.88%, and 43.57% than the linear kernel function (58.95%), radial basis kernel function (62.11%) and Sigmoid kernel function (53.16%), respectively. (2) The model accuracy of the SVM model (AUC=0.711) was 0.42% higher than the RF model (AUC=0.708), and the prediction accuracy of the RF model (AUC=0.745) was 12.88% higher than the SVM model (AUC=0.660). The prediction ability of the RF model is better when the accuracy of the SVM model and the RF model is similar. (3) In the sensitivity zoning results of the RF model, the areas of low sensitive area, moderate sensitive area, high sensitive zone, and very high sensitive area are respectively 150.904 km², 206.611 km², 207.802 km², and 159.090 km², and accordingly, the percentage of subsidence disaster points in 2011 accounted for 11.11%, 18.52%, 28.89% and 41.48% of the total disaster points, and those in 2014 accounted for 11.17%, 20.21%, 31.91% and 36.71%, respectively. (4) In the zoning of sensitivity results, more subsidence disaster points distribute in a smaller extremely sensitive zoning area for the RF model, which is a more suitable prediction model for zoning prediction of developmental sensitivity to subsidence disasters than SVM. The study can provide useful references for the selection of zoning prediction models of developing sensitivity to subsidence disasters in underground coal mining areas.