关键词: 黄土陷穴, 区域预测, 多源数据, 卷积神经网络, 深度神经网络

Abstract: Loess sinkholes are a unique type of geological hazard that is widespread across the Loess Plateau. Their prevention and control are essential considerations in construction projects within this region. Based on a modified RUSLE, this study extracts 12 different types of feature factors from multiple data sources, including DEM, precipitation, surface cover, and vegetation index. Two prediction models, CNN and DNN are constructed to predict areas prone to loess sinkhole development. The results of the two models are compared and analyzed to provide reference for the prevention and control of sinkhole hazards, construction projects, and soil and water conservation in loess areas. The findings show that both the CNN and DNN models achieve an accuracy rate of over 80% and an F1 score of over 83%, indicating their effectiveness in predicting areas prone to loess sinkholes. The CNN model achieves an accuracy of 83.25% and an F1 score of 85.18%, which are 2.63% and 1.56% higher than those of the DNN model, respectively. This demonstrates the superior generalization ability and detailed performance of the CNN model. Analysis of the prediction results indicates that loess sinkholes develop more strongly in valley areas, less so on flat terrain, and are influenced to some extent by human activities.

Key words: loess sinkhole, regional prediction, multi-source data, CNN, DNN