To address the limitations of traditional landslide prediction methods that rely on expert experience and struggle to adapt to complex underground engineering scenarios, an improved random forest method was proposed to enhance prediction accuracy and generalization capability. Key parameters of landslides in underground engineering environments were collected using specialized equipment, and a high-quality dataset was created through incomplete data processing. By comparing the global and local classification performance of classifiers such as k-nearest neighbors, support vector machines, and decision trees, random forest was identified as the optimal base classifier. An innovative multi-decision tree correlation measurement method was introduced, which quantifies redundancy between trees through inner product calculations in feature space and optimizes the model by threshold filtering. Experimental results demonstrated that, the improved random forest achieved a classification accuracy of 93.05%, with Precision, Recall, and F1-score maintaining the highest stability across different labeled data, validating its dual advantages in both global and local classification performance. Practical engineering applications confirmed its effectiveness in multidimensional complex scenarios, providing a reliable solution for intelligent geological hazard prediction.