Detection of strawberry fruit and leaf diseases based on improved YOLOv8n and multi-scale collaborative attention

Manual detection of strawberry fruit and leaf diseases is susceptible to environmental constraints. In this study, an intelligent detection model (named YOLOv8n-SFLD) was proposed to integrate with multi-scale collaborative attention using an improved YOLOv8n framework. A high-quality dataset of 8,745 images was constructed with seven phenotypes of strawberry disease (angular leaf spot, anthracnose, blossom blight, gray mold, leaf spot, powdery mildew on leaves and fruits) from the Kaggle platform and field collections. Data augmentation was then performed with geometric transformations, illumination adjustments, noise addition, occlusion simulation, and Mosaic enhancement. A four-fold optimization was applied to the baseline YOLOv8n: 1) Universal Receptive Large Kernel Network (UniRepLKNet) module was integrated into the C2f structure to extract deep semantic features. 2) Channel Prior Convolutional Attention (CPCA) was introduced into the backbone to improve lesion detection accuracy under complex backgrounds. 3) Multi-Scale Collaborative Attention (MSCA) with parallel 3×3, 5×5, and 7×7 convolutional branches was incorporated into the neck architecture to capture multi-scale lesion features. 4) Wise-IoU loss function replaced CIoU to optimize bounding box regression for small targets via dynamic non-monotonic focusing. Experimental results demonstrate that the YOLOv8n-SFLD achieved superior performance. Attention mechanism showed that the embedding CPCA in the backbone obtained the best precision (88.6%), while the MSCA in the neck was achieved in the highest mAP50 (87.1%). Dataset analysis revealed that there was the significant variation in the performance over the different categories: The blossom blight, healthy fruit, angular leaf spot, and leaf spot were achieved in the mAP50 of 98.9%, 99.5%, 90.1%, and 95.0%, with the miss rates below 12.5%, while the anthracnose and healthy leaf shared the lower mAP50 (62.8%, and 58.8%) and higher miss rates (44.0%, and 37.2%), due to the sample imbalance. Ablation experiments demonstrated the synergistic effects of all modules: The complete YOLOv8n-SFLD (C2f-URBlock+CPCA+MSCA+Wise-IoU) was attained the precision of 91.9%, mAP50 of 87.0%, and mAP50-95 of 67.5%, which was improved by 7.4, 2.3, and 1.5 percentage points over baseline, with the model weight of 13.4 MB and FLOPs of 16.0 G. Furthermore, the YOLOv8n-SFLD outperformed all in the precision (5.2-10.6 percentage points higher) and mAP50 (2.0-5.8 percentage points higher), compared with the RT-DETR, YOLOv8n, YOLOv8s,YOLOv9t, YOLOv10n, YOLOv11, and YOLOv12. Visualization analysis showed that the superiority was achieved in the gray mold detection with multi-disease co-infection. All lesions were accurately detected, while the misdetection was observed in the dried leaves or miss fruit powdery mildew. In leaf spot with dense small lesions, the complete coverage was obtained with the tightly fitting boxes; The high precision and accurate localization were maintained in the powdery mildew under low-contrast backgrounds. The better performance was attributed to the synergistic optimization: C2f-URBlock was used to enhance the deep feature extraction, CPCA was used to strengthen the lesion discrimination under complex backgrounds, MSCA was suitable for the multi-scale variations, and Wise-IoU was used to improve the small-target regression. Furthermore, the improved YOLOv8n-SFLD effectively alleviated the multi-scale lesion variation, low-contrast features, and complex background interference. There was an optimal balance between accuracy and efficiency for deployment on the resource-constrained edge devices. High generalization was also obtained to detect similar diseases of the crops. The finding can provide reliable technical support for the precision pesticide in smart agriculture. Future work can focus on dataset expansion, model compression for mobile deployment, and UAV integration for real-time monitoring.