Automatic in-trap pest detection using deep learning for pheromone-based Dendroctonus valens monitoring

Dendroctonus valens LeConte, the red turpentine beetle (RTB) is an invasive pest which is severely threatening pine species in China. Traditional pheromone-based RTB monitoring depends on a manual count, which is not only time consuming but also error-prone. This study proposes a deep learning detection method for counting adult RTBs directly in the pheromone trap. The images of bark beetles were captured by a digital camera embedded in the collection cup of the pheromone trap. The state-of-the-art one-stage deep learning detector RetinaNet was further downsized by the depthwise separable convolution and feature pyramid tailoring, which was feasible to run on embedded devices. To enable the fine-grained identification of bark beetles, the downsized detector was enhanced by k-means anchor optimisation and residual classification subnet. The models were trained end-to-end on a GPU workstation with data augmentation and then deployed on embedded devices with minimal preprocessing. Experiments demonstrated that the object level average precision (AP) was 0.746 and the average runtime on the Jetson TX2 and Raspberry Pi 3b were 0.448s and 23.44s, respectively. The in-trap detector was capable of distinguishing the most aggressive RTBs from other five species of bark beetles attracted by pheromone with unconstrained size, pose, orientation, integrity, and position. The proposed method showed promising performance both qualitatively and quantitatively within limited computational budgets, which has introduced a practical and applicable solution for early warning of RTB outbreaks.