CES-YOLO: An efficient fallen fragrant pear detection and counting algorithm in complex post-disaster orchard environment

Accurately detecting and quantifying the number and distribution of fallen fruits after a disaster is of great significance for the rapid loss assessment of agricultural fruit insurance and orchard yield estimation. In view of the problems in the unstructured orchard environment, such as the small target size of fragrant pear fruits in the aerial images obtained by drones, the similar color between fruits and leaves, and susceptibility to occlusion by ground debris. This study proposed the CES-YOLO model for detecting and counting fallen fragrant pears after hail disasters, which is based on the YOLOv8s model. First, CAS-ViT was used as the backbone feature extraction network to improve the ability of CES-YOLO to perceive global information. Second, an efficient multiscale attention mechanism (EMA) is embedded before the output detection head to capture more detailed information and enhance the correlation between local features. Finally, to mitigate the problems of false detection and localization inaccuracies caused by overlapping occlusions in fragrant pear fruit clusters, we introduced SIoU for bounding box regression, focusing on the difference between the directions of the predicted box and the ground truth box, improving both the localization accuracy and convergence speed. The experimental results showed that the [email protected], accuracy, and recall rate of the CES-YOLO model in detecting and counting fallen fragrant pears in complex environments were 93.9%, 89.9%, and 86.4% respectively, with improvements of 0.5%, 2.5% and 0.9%, respectively, compared to the YOLOv8s model. When the CES-YOLO model was transferred to the fragrant pear gale disaster dataset for cross-disaster generalization experiments, [email protected] reached 95.5%. The CES-YOLO model proposed in this study has a high counting accuracy for fallen fragrant pears in complex post-disaster environments, and could provide a reliable technical support for improving the accuracy and efficiency of agricultural fruit insurance disaster loss assessment.