Land degradation in semi-arid grasslands has intensified under the combined effects of climate change and human disturbance, highlighting the need for accurate and interpretable monitoring approaches. This study proposes a process-based framework integrating unmanned aerial vehicle (UAV) imagery with deep learning to quantify desertification using the Soil Pixel Proportion (SPP)—a novel ecological indicator that quantifies the proportion of exposed soil relative to total surface area. The framework combines MSCA-U-Net for multi-scale soil–vegetation segmentation and ViT-B/16 for degradation-level classification, establishing a link between remote sensing features and ecological processes. Field validation showed strong correlations between SPP, NDVI (r = –0.76), and soil organic carbon (r = –0.68), confirming its ecological validity. After segmentation-assisted relabeling, ViT achieved an F1-score of 95.3% and maintained cross-regional accuracy of 91.7% from semi-arid to semi-humid regions. Robustness tests under cloud, shadow, and snow disturbances verified stable performance (F1 > 85%), ensuring reliable detection of early-stage degradation. The SPP-based framework provides a reliable and interpretable approach for large-scale grassland degradation assessment, offering valuable support for restoration planning and progress evaluation toward Land Degradation Neutrality (LDN) within the United Nations Convention to Combat Desertification (UNCCD) framework.

This paper introduces a disparity-independent framework for 3D localization in heterogeneous stereo vision systems. Unlike conventional disparity-based approaches, the method employs independent monocular calibration, cooperative geometric projection, and polynomial focal length fitting with integrated error compensation to directly compute spatial coordinates. This reformulation reduces the reconstruction problem to a constant-time linear algebra solution, achieving real-time efficiency. The framework accommodates variations in focal length, installation height, and tilt angle, thereby overcoming the homogeneous camera assumption. Extensive experiments under diverse configurations demonstrate an average accuracy of 97.9%, substantially outperforming SGBM, PSMNet, and RAFT-Stereo. The proposed method exhibits strong robustness to camera heterogeneity and environmental perturbations, making it suitable for robotic applications requiring cost-effective, high-precision localization in dynamic and resource-constrained settings.