Supervised Cross-Scene Transfer and Boundary-Aware Optimization for UAV-Based Landslide Surface Fissure Segmentation

High-resolution unmanned aerial vehicle (UAV) imagery enables detailed mapping of landslide surface deformation， yet automated extraction of fine-scale fissures remains challenging because of cross-scene domain shifts and boundary ambiguity. This study proposes a supervised cross-scene transfer framework that integrates source-domain pre-training， target-domain fine-tuning， and boundary-aware optimization. A Residual U-Net (ResU-Net) was pre-trained on a generalized multi-scene fissure dataset and subsequently fine-tuned on site-specific UAV orthophotos of a reservoir landslide. A hybrid loss combining binary cross-entropy， Dice loss， and boundary loss was employed to jointly reduce pixel-wise classification errors， region-level mismatch， and boundary displacement. On the primary target-domain dataset， the proposed framework achieved a mean IoU of 0.5631 and a mean Dice coefficient of 0.7205 across three independent runs， representing an absolute IoU improvement of 0.3606 over the baseline ResU-Net. Ablation experiments demonstrated the combined contributions of cross-scene transfer and boundary-aware supervision， while the proposed configuration showed clearer delineation of thin fissures in the displayed qualitative examples. The resulting fissure maps were integrated into a prototype visual analytics system for expert-assisted inventory construction and engineering interpretation. The framework provides a practical approach for fine-scale fissure mapping from UAV imagery， although its transferability requires validation at independent landslide sites.