Abstract: Rapid and accurate early identification of regional landslides and their potential hazards is fundamental to landslide disaster prevention and mitigation. However, traditional field investigation methods are limited by low efficiency and insufficient spatial coverage, highlighting the need for automated landslide identification approaches applicable to large areas. Taking Sai Kung East Country Park in Hong Kong, China as the study area, a regional landslide automatic identification framework is developed in this paper based on the Mask Region-based Convolutional Neural Network (Mask R-CNN). Digital terrain model (DTM) data are integrated with satellite imagery as model inputs, transfer learning is adopted to complete model training, and confidence scores are employed to quantify prediction uncertainty. Results that(1) After incorporating DTM data, the precision on the training set increases from 73.24% to 86.75%, and the model's ability to distinguish landslides from roads is significantly enhanced. (2) The recall on the testing set reaches 71.43%, demonstrating satisfactory generalization capability. (3) The confidence scores of most landslide samples exceed 0.7, indicating high reliability of the identification results. The proposed Mask R-CNN approach integrating DTM and satellite imagery enables effective automatic detection and instance segmentation of landslides over large areas, providing reliable technical support for regional cluster landslide identification and geological hazard early warning.