Search

IntroductionRodent outbreak is the main biological disaster in grassland ecosystems. Traditional rodent damage monitoring approaches mainly depend on costly field surveys, e.g., rodent trapping or hole counting. Integrating an unmanned aircraft system (UAS) image acquisition platform and deep learning (DL) provides a great opportunity to realize efficient large-scale rodent damage monitoring and early-stage diagnosis. As the major rodent species in Inner Mongolia, Brandt’s voles (BV) (Lasiopodomys brandtii) have markedly small holes, which are difficult to identify regarding various seasonal noises in this typical steppe ecosystem.MethodsIn this study, we proposed a novel UAS-DL-based framework for BV hole detection in two representative seasons. We also established the first bi-seasonal UAS image datasets for rodent hole detection. Three two-stage (Faster R-CNN, R-FCN, and Cascade R-CNN) and three one-stage (SSD, RetinaNet, and YOLOv4) object detection DL models were investigated from three perspectives: accuracy, running speed, and generalizability.ResultsExperimental results revealed that: 1) Faster R-CNN and YOLOv4 are the most accurate models; 2) SSD and YOLOv4 are the fastest; 3) Faster R-CNN and YOLOv4 have the most consistent performance across two different seasons.DiscussionThe integration of UAS and DL techniques was demonstrated to utilize automatic, accurate, and efficient BV hole detection in a typical steppe ecosystem. The proposed method has a great potential for large-scale multi-seasonal rodent damage monitoring.