DLCAPL: Semi-Supervised Livestock Detection in the Ruoergai Alpine Grasslands Using Dual-View Localization Consistency and Reliability-Weighted Pseudo-Label Learning

Unmanned aerial vehicle (UAV) remote sensing provides an efficient approach for livestock monitoring in alpine grassland environments, particularly in regions where traditional field surveys are constrained by complex terrain and large spatial coverage. However, livestock targets in nadir-view UAV imagery are typically small, densely distributed, partially occluded, and visually similar to the surrounding vegetation background, which makes accurate detection challenging. In addition, manually annotating large numbers of individual animals in high-resolution UAV images is time-consuming and labor-intensive. To reduce dependence on extensive manual annotation while improving the utilization of unlabeled UAV imagery, this study proposes DLCAPL (Dual-View Localization Consistency and Reliability-Weighted Pseudo-Label Learning), a semi-supervised livestock detection framework based on YOLOv13s. Within an exponential moving average (EMA)-based teacher–student framework, a dual-view localization consistency mechanism is introduced to evaluate pseudo-label reliability based on localization stability across augmented views. Teacher predictions are further hierarchically divided into reliable, weak, and ignored pseudo-labels using medium- and high-confidence thresholds. For weak pseudo-labels, classification confidence and localization consistency are jointly used to estimate reliability and assign reliability-aware supervision weights, thereby retaining informative difficult targets while reducing the influence of noisy pseudo-labels. Experiments were conducted on a self-constructed UAV livestock dataset collected in the Ruoergai alpine grasslands containing 602 images and 57,390 livestock instances. With only 30% labeled data and 70% unlabeled data, DLCAPL achieved 73.28% mAP@0.5 and 50.68% mAP@0.5:0.95. Compared with the supervised baseline under the same annotation ratio, DLCAPL improved these two metrics by 17.96 and 16.43 percentage points, respectively, while approaching the performance of the fully supervised YOLOv13s model trained with 100% labeled data. These results demonstrate that DLCAPL can effectively exploit unlabeled UAV imagery, reduce annotation requirements, and maintain competitive detection performance for livestock monitoring in complex alpine grassland environments.