Abstract: Talus slopes in high-altitude cold mountain regions, characterized by extensive distribution, complex morphology, and elevated geohazard risks, pose significant challenges for intelligent identification due to harsh environmental conditions and data scarcity. This study addresses this gap by constructing the first high-resolution semantic segmentation dataset for talus slope detection. Utilizing Gaofen-2 (GF-2) satellite imagery, we developed a multidimensional interpretation framework incorporating morphological, spectral, and environmental criteria, culminating in an open-access benchmark dataset comprising 3,811 standardized annotated samples. Through systematic evaluation of four convolutional neural network architectures and two Transformer-based models under unified experimental protocols, we validated the technical superiority of the Mask2Former model - integrating Transformer architecture with dynamic Mask attention mechanisms - in complex terrain scenarios, achieving a mean Intersection over Union (mIoU) of 75.72% and F1-score of 77.62%. The proposed methodology has demonstrated exceptional generalization capability and robustness in precise talus slope delineation. This research not only addresses the critical data gap for talus slope studies in alpine environments but also provides scientific guidelines for model selection in intelligent recognition tasks across complex geomorphological settings. The established dataset and multidimensional interpretation system offer valuable references for advancing geohazard monitoring through intelligent remote sensing technologies in cold mountainous regions.