Shield tunneling construction can cause ground settlement. In order to explore the law of surface subsidence, this study is carried out based on actual engineering. This study uses Convolutional Neural Networks (CNN) combined with ArcGIS geographic tools to make judgments and predictions based on survey data from actual engineering projects. By analyzing the principles of neural networks, build a neural network model suitable for this study; Analyze soil layer information, select suitable data for the input model, input the model for prediction, and compare the accuracy using an adaptive genetic algorithm optimized BP neural network (AGA-BP) and a combination model of deterministic coefficient logistic regression (CF-LR). The research results were compared using ROC curves. After comparison, the accuracy of the CNN model is 0.862, the accuracy of the AGA-BP model is 0.808, and the accuracy of the CF-LR model is 0.778. The standard errors of the three models are 0.037, 0.046, and 0.048, respectively. Compared with the other three, CNN has the highest accuracy and the smallest error. By combining CNN model with ArcGIS for visualization processing, the research results show that the CNN model can identify the risk level of settlement at different positions within the shield tunnel section with high accuracy. The high-risk areas for settlement disasters in the shield tunnel section account for a large proportion and preventive measures need to be taken.