Abstract:
Over the past four decades, rising demand for livestock products in Africa has led to increased stocking rates
resulting in overgrazing and land degradation. As the population is projected to rise, the need for sustainable
livestock management is more urgent than ever, yet efforts are hindered by the lack of accurate, up-to-date
livestock counts. Recent advances in remote sensing and deep learning have made it possible to count livestock from space. However, the extent to which models trained on aerial imagery can enhance livestock detection
in satellite images and across diverse landscapes remains limited. This study assessed the transferability of YOLO,
Faster R-CNN, U-Net, and ResNet models for livestock detection across three contrasting landscapes, Choke
bushland (Pleiades Neo), Kapiti savanna (WorldView-3), and LUMO open grassland (WorldView-3), using satellite imagery with 0.3 m and 0.4 m spatial resolution. Additionally, we applied a multi-stage transfer learning to
evaluate the effectiveness of aerial imagery (0.1 m) trained models in improving livestock detection in satellite
imagery. Results indicate that YOLOv5 consistently outperformed other models, achieving F1 scores of 0.55,
0.67, and 0.85 in Choke, Kapiti, and LUMO, respectively, demonstrating robustness across varying land cover
types and sensors. Although segmentation models performed moderately on 0.3 m imagery (F1 scores of 0.51 and
0.40 for Choke and LUMO), their performance dropped significantly on the coarser resolution (0.4 m) Kapiti
imagery (F1 score of 0.14). In addition, multi-stage transfer learning improved segmentation models recall by
9.8 % in heterogeneous bushland site. Our results highlight that the integration of multi-source imagery and deep
learning can help in large scale livestock monitoring, which is crucial in implementing sustainable rangeland
management.
