关键词: 无人机机场, 野生动物监测, 目标检测, 多物种数据集, YOLO

Abstract

Aims: Unmanned aerial vehicles (UAVs) provide high-spatial-resolution imagery for wildlife monitoring in nature reserves and offer new opportunities for repeated surveys over large or difficult-to-access areas. With the increasing use of UAV-airport systems, large volumes of patrol imagery can be collected during routine monitoring. However, such imagery is often heterogeneous in species composition, target scale, and background conditions, and multi-species datasets derived from real patrol operations remain relatively limited. In addition, the applicability of commonly used object-detection models to these data has not been sufficiently evaluated under unified experimental settings. This study therefore aimed to construct a multi-species wildlife image dataset from routine UAV-airport patrol imagery collected in a nature reserve, systematically compare the performance of different YOLO models and model configurations, and evaluate their applicability across target scales and wildlife species. 

Method: Visible-light images were collected using a DJI Dock 2 UAV-airport system and a DJI M3TD UAV in Xilin Gol Grassland National Nature Reserve, Inner Mongolia, China. Thermal infrared imagery was used to assist in locating suspected wildlife targets, after which visible-light images were acquired for confirmation and detailed recording. After perceptual-hash-based deduplication, manual annotation, and category screening, the final dataset contained 5,838 images, 11,992 annotated instances, and 16 wildlife categories, including six mammal species, nine waterbird species, and the Eurasian magpie. To reduce information leakage caused by highly similar consecutive images, the dataset was divided into training, validation, and independent test sets by continuous shooting sequence rather than by randomly assigning individual images. Seven YOLO baseline models were first compared under standardized training conditions. Based on the baseline results, YOLO11s and YOLO26s were further evaluated at five input resolutions ranging from 640 × 640 to 1,536 × 1,536 pixels. Eight candidate improvement strategies targeting class sampling and data augmentation, feature enhancement, multi-scale feature fusion, small-object detection, and bounding-box regression were also evaluated. Target instances were divided into large, medium, and small scales according to the ratio of bounding-box area to image area, and additional validation and test subsets containing at least one small-scale target were constructed to evaluate overall performance in complete image scenes containing small targets. 

Results: Among the seven baseline models, YOLO26s achieved the highest mAP@50:95 on the full validation set, reaching 0.558 ± 0.008, although the differences among several lightweight YOLO models were relatively small. Increasing input resolution produced a clearer improvement than changing the model architecture alone. When the input resolution of YOLO26s was increased from 640 × 640 to 960 × 960 pixels, mAP@50:95 increased to 0.603 ± 0.018 on the full validation set and to 0.679 ± 0.014 on the validation subset containing at least one small-scale target. Further increasing resolution yielded diminishing gains while substantially increasing computational cost. Scale-stratified evaluation showed that detection performance differed among large, medium, and small targets, with small-scale targets generally remaining more difficult to detect. The eight candidate improvement strategies showed no clear overall advantage during single-seed screening, and none of the four strategies subjected to four-seed reevaluation produced a stable overall improvement over the YOLO26s@960 baseline, although some strategies yielded local improvements at particular target scales. Considering validation accuracy and computational cost, YOLO26s@960 was selected as the recommended configuration for the present study-area and data-acquisition conditions. On the independent test set, it achieved an mAP@50:95 of 0.625 ± 0.007, with mAP@50:95 values of 0.611, 0.442, and 0.375 for large-, medium-, and small-scale targets, respectively. Species-specific AP@50:95 ranged from 0.019 to 0.914, indicating substantial variation in detection performance among wildlife species. 

Conclusion: The results demonstrate the feasibility of applying deep-learning-based object detection to multi-species wildlife imagery obtained during routine UAV-airport patrols. Compared with simply replacing YOLO model generations, increasing input resolution provided a more pronounced improvement under the present dataset conditions, particularly for smaller targets, although the benefit decreased as resolution continued to increase. The results also show that improvements observed for individual target scales do not necessarily translate into consistent gains in overall detection performance, emphasizing the value of evaluating both overall and scale-specific metrics. Under the study-area and data-acquisition conditions examined here, YOLO26s@960 provided a suitable balance between detection accuracy and computational cost and can serve as a reference configuration for computer-assisted screening of similar UAV-airport wildlife imagery. The marked differences among target scales and species further indicate that practical applications should consider target-specific performance rather than relying solely on an overall mAP value. Expanding the dataset across seasons, regions, and patrol conditions will further help evaluate the stability and transferability of the recommended configuration.

Key words: UAV airport, wildlife monitoring, object detection, multi-species dataset, YOLO