Abstract: Karst collapse is a common dynamic geological hazard in karst development areas, characterized by high suddenness and concealment, posing severe threats to site stability and engineering safety. To accurately identify the primary triggering factors of karst collapse in the study area, this paper constructs both a deep neural network (DNN) model and a traditional analytic hierarchy process (AHP) model based on regional geological conditions to evaluate and predict collapse susceptibility, followed by comparative validation of their accuracy. The results demonstrate that the DNN model outperforms the AHP model in both predictive accuracy and reliability, achieving an overall evaluation accuracy of 87.9%, significantly higher than the 66.7% attained by the AHP model. The DNN model exhibits distinct advantages in assessing karst collapse susceptibility and can provide a scientific basis for disaster prevention and planning in karst regions.

Key words: karst collapse, deep learning, neural network, analytic hierarchy process, susceptibility assessment