Lower-Boundary-Aware Sky–Ground Segmentation for Robust Horizon Estimation in Low-Altitude UAV Remote Sensing

Low-altitude unmanned aerial vehicles (UAVs) provide flexible platforms for remote-sensing data acquisition, but their image sequences are frequently affected by attitude variation, illumination degradation, cluttered sky–ground structures, and partial occlusion. To improve horizon perception under these conditions, this paper proposes LBA-YOLO-Seg, a lightweight lower-boundary-aware sky–ground segmentation framework that explicitly links mask prediction with downstream geometric recoverability. The method combines Ghost-based feature extraction, Coordinate Attention, and a high-resolution boundary pathway to preserve boundary localization and directional continuity, while a boundary-completeness-adaptive Kalman estimator improves temporal stability. Experiments on a self-collected UAV dataset show that the proposed method outperforms traditional horizon detectors, representative segmentation networks, boundary-aware methods, and YOLO family baselines in geometric accuracy and temporal stability. Zero-shot evaluation on the public CH1 dataset further demonstrates cross-domain generalization, while embedded deployment on Rhino Pi-X1 achieves an end-to-end processing rate of 47.53 FPS. These results indicate that optimizing segmentation representation for lower-boundary recoverability provides an effective and deployable solution for robust UAV horizon perception.