Learning Discriminative Representations of Landslides via Deep Metric Learning Under Limited Supervision
Abstract
Detecting low-contrast targets from complex backgrounds in remote sensing, such as landslides, poses fundamental challenges for learning discriminative representations under limited expert annotations. Deep metric learning (DML) provides a principled framework for learning discriminative representations by explicitly shaping the embedding space to enhance intraclass compactness and interclass separability. However, most existing DML paradigms are developed for image-level settings, where entire images are treated as a single category, and fail to explicitly address the supervision requirements of instance-level object detection, because contrastive supervision is constrained to labeled foreground instances, lacking an explicit mechanism to leverage background regions. We propose OSG-UAHNM, a plug-and-play framework that learns discriminative landslide representations through loss augmentation. A batch-wise online sample generator (OSG) constructs instance-level contrastive pairs from both labeled instances and unlabeled background regions, while an uncertainty-aware hard negative miner (UAHNM) identifies high-quality hard negatives by jointly modeling classification confidence and localization precision. With negligible parameter overhead and zero inference cost, OSG-UAHNM can be integrated into diverse architectures, albeit with noticeable additional training computation. Experiments on a field-surveyed loess landslide dataset demonstrate consistent AP50:95 gains of 7.6, 6.5, 2.7, 1.1, and 1.8 percentage points on Faster R-CNN, YOLOv8n, YOLO11n, YOLO26n, and DETR, respectively. Cross-dataset validation on the publicly available Bijie landslide dataset further demonstrates the robustness of the framework across distinct geological settings and landslide scales. Embedding-space analysis shows that interclass separation increases by more than two orders of magnitude while maintaining intraclass compactness.