Abstract: In landslide mapping from remote sensing imagery, complex and highly variable target morphologies often lead to fragmented segmentation, while complex backgrounds may cause false positive detections and jagged boundaries. To address these issues, a Shape-Aware and Boundary Refinement Network (SABR-Net) is proposed. The proposed model employs a multi-branch heterogeneous convolutional structure to enhance the feature representation of narrow, elongated, and irregular landslide morphologies. Simultaneously, a gated boundary refinement module is introduced, which utilizes high-level semantic information to dynamically suppress background noise in low-level features, thereby improving the delineation quality of landslide boundaries. The SABR-Net model is validated on the rainfall-induced landslide dataset in Zixing City, Hunan Province. Experimental results demonstrate that SABR-Net achieves a mean Intersection over Union (mIoU) of 75.23% and an F1 score of 70.87%. Its comprehensive performance outperforms seven mainstream models, including Trans-UNet, while maintaining a total parameter count of only 1.50 M. The proposed method achieves a favorable balance between extraction accuracy and computational overhead, effectively alleviating the issues of morphological fragmentation and jagged boundaries. It provides a feasible solution for disaster emergency monitoring and deployment on edge devices.