CHEN Junnan, WANG Haiguang
Most studies on plant disease image recognition are based on the images of individual diseases, and there are few related studies based on the images of mixed diseases acquired under the complex field environments. In this study, a total of 16 562 images of three wheat leaf diseases (stripe rust caused by Puccinia striiformis f. sp. tritici, leaf rust caused by P. triticina, and powdery mildew caused by Blumeria graminis f. sp. tritici) with single background and complex background were acquired during different wheat growth stages under both field and indoor environments, including the images of individual diseases and mixed diseases. According to the distribution of the labeled disease instances, a class-balanced mixed image dataset was constructed, then it was divided into a single-background image dataset and a complex-background image dataset, and subsequently, each of the three image datasets was divided into training set, validation set, and testing set based on the disease instance quantity ratio of 7∶2∶1, respectively. Nine object detection models, including YOLOv5n, YOLOv6n, YOLOv8n, YOLOv9t, YOLOv10n, YOLO11n, YOLO12n, SSD, and Faster R-CNN, were used to carry out wheat leaf disease detection based on the three image datasets, including the mixed image dataset, the single-background image dataset, and the complex-background image dataset. The results showed that on each image dataset, the comprehensive performance of each YOLO model was markedly better than the SSD and Faster R-CNN models, except that of YOLOv6n on the single-background image dataset. YOLO11n, YOLOv10n, and YOLOv8n had the best comprehensive performance on the mixed image dataset, the single-background image dataset, and the complex-background image dataset, respectively. By comprehensively considering the characteristics of the three constructed image datasets and the detection results, the mixed image dataset was regarded as the most suitable image dataset for building the image-based wheat leaf disease detection model. Furthermore, among all the target detection models, YOLO11n performed best on the testing set of the mixed image dataset, with the Precision, Recall, F1 score, mAP50, and mAP50-95 of 76.95%, 72.59%, 74.70%, 76.08%, and 54.19%, respectively. Meanwhile, YOLO11n demonstrated advantages in wheat leaf disease detection on both the single-background and complex-background image datasets. On the whole, the YOLO11n model was the optimal for wheat leaf disease detection among the nine selected target detection models. In this study, a method for the detection of multi-lesion images and mixed images of wheat leaf diseases and supports for plant disease detection in complex scenarios were provided.
