Automated detection and counting of redbanded stink bugs in soybean using an improved computer vision model.

The redbanded stink bug (RBSB), Piezodorus guildinii, is a major economic pest of soybean, with feeding damage that leads to significant yield losses and increased reliance on pesticide applications. Current management practices depend on manual identification and repeated field counting, which are labor-intensive, time-consuming, and prone to human error, particularly across large production areas. To address these limitations, this study evaluated the potential of computer vision models to automatically detect and count RBSB adults using imagery. A total of 2,281 field images were collected under varying stink bug densities using different sensors. These images were used to train multiple YOLOv8 model variants for automated detection and counting. The best-performing model was further enhanced by integrating a convolutional block attention module (CBAM) and adaptive spatial feature fusion (ASFF) to improve detection accuracy and counting performance, which are critical for informed pest management decisions. The improved YOLOv8m model achieved 96.30% precision, 76.40% recall, F1-score of 85.20, 82.00% mean average precision (mAP50), and 75.30% mean average precision (mAP50-95) for detecting and counting RBSB adults in images, outperforming the baseline model, which showed lower performance with 95.68% precision, 76.69% recall, F1-score of 85.13, 78.70% mAP50, and 66.00% mAP50-95. The enhanced model was deployed in a prototype web application to evaluate its practical capabilities. This application allows users to upload images and visualizes detection bounding boxes and RBSB counts, resulting in low misdetection error. Overall, the framework and results presented in this study provide a practical alternative to manual identification of RBSB adults in soybean systems using images and have the potential to improve traditional pest monitoring practices.