Landslide detection in remote sensing images on the basis of feature recursion and sample balance

To increase the accuracy of landslide identification via high-resolution optical remote sensing images, this paper presents an improved landslide identification model based on the Mask R-CNN. First, a novel feature image extraction and fusion structure, ResNeSt-RFP, is constructed to enhance the model's capacity for deep landslide feature expression. Second, the improved model incorporates a one-stage object recognition structure, separates the segmentation and classification modules, and leverages the important sample balancing training capabilities inherent in the one-stage structure. This enables the model to effectively utilize the limited number of landslide samples, thereby improving the accuracy of the suggested bounding boxes in the first stage and enhancing the overall model performance. The experimental results conducted on the self-compiled Yunnan landslide dataset and the publicly available Bijie landslide dataset demonstrate that the improved Mask R-CNN effectively adapts to landslide remote sensing image datasets acquired from different sensors. The detection performance metrics achieve 91.4% in Bounding Box Average Precision (BBox AP)50 and 63.5% in BBox AP75, while the segmentation metrics reach 87.4% in Segmentation Mask Average Precision (Mask AP)50 and 52.6% in Mask AP75. Compared with the baseline model, these metrics exhibit respective improvements of 7.9%, 5.5%, 8.1%, and 6.5%. Furthermore, comparative experiments on these datasets show that the proposed method outperforms models such as YOLOX, Cascade R-CNN, Deformable DETR, Mask2Former, and DINO. To further validate its generalizability across different datasets, we conducted additional experiments on the Global Distributed Coseismic Landslide Dataset (GDCLD). This study demonstrates that the proposed model achieves notable results in automated landslide extraction tasks, providing accurate landslide identification outcomes that can aid in emergency response and disaster assessment efforts.