کاربرد روش بینایی ماشین در تشخیص بیماری‌های توت‌فرنگی با YOLO 11

Introduction
Strawberries are one of the most important and valuable garden crops, widely supplied to global agro-markets, as well as food and pharmaceutical industries, due to their high nutritional and economic value, antioxidant compounds, and favorable taste. However, strawberry production is often accompanied by challenges, one of the most important of which is prevalence of plant diseases. Diseases of fungal or bacterial origin typically cause damage to the crop growth, which may reduce the yield, and in some cases, lead to the complete harvest failure and significant financial losses. Timely and accurate identification of these diseases plays a crucial role in effective farm management. Traditional methods, such as visual inspection, require considerable expertise, are time-consuming, prone to errors, and often yield suboptimal results. In recent years, advancements in technologies related to artificial intelligence and machine learning, particularly in machine vision models, have made it possible to automatically identify plant diseases with higher speed and accuracy. In this study, the YOLO algorithm, one of the most widely used and advanced methods in digital object recognition, was employed to identify various strawberry diseases, including angular leaf spots, anthracnose, gray mold, and powdery mildew. To improve the accuracy of the model, modifications were made to the network architecture and training process. In addition to high accuracy and appropriate speed, this method may reduce the costs related to monitoring and managing strawberry farms. The results obtained from this study demonstrate that the YOLO algorithm can be effectively utilized in smart agriculture, in conjunction with specific equipment and tools such as drones and image sensors, to control diseases and enhance production. This research represents a practical step toward utilizing modern technologies to manage plant diseases and enhance agricultural productivity.
Materials and Methods
The dataset used consisted of 2902 images of strawberry leaves and fruits collected from the Roboflow database. The whole dataset was divided into three distinct portions: training (70%), validation (20%), and test (10%). All images were set to 640×640 pixels, and the labeling process was performed according to YOLO standards. Eight disease classes, including angular leaf spot, anthracnose rot, blossom blight, gray mold, healthy leaves and fruits, leaf spot, powdery mildew on fruits, and powdery mildew on leaves, were included in the dataset. The YOLO11-Large (YOLO11L) model was trained using pre-trained weights and an object detection task. The Optuna algorithm was used to optimize the hyperparameters. The training process consisted of 200 epochs, utilizing an early stopping mechanism to prevent overfitting. The training batch size was set to 4, and other settings, such as data augmentation, including random rotation, scaling, horizontal and vertical inversion, random cropping, and image blending, were also applied. Finally, a deep learning model based on YOLO11L was used to identify and classify strawberry plant diseases. The final model is consists of 190 layers and approximately 790,000 trainable parameters, which are distributed among three main parts of the network: about 480,000 parameters in the backbone (feature extraction), 285,000 in the neck (feature aggregation), and 25,000 in the head (detection output). The model’s total computational complexity is approximately 6.86 GFLOPs. The processing speed of the model was measured to be 0.5 ms for preprocessing, 23.6 ms for inference, and 2.3 ms for postprocessing per image.
Results and Discussion
The results show that the evaluation accuracy of the model is the best in case of blossom blight class, where a precision of 0.951 and a full recall of 1.000 were obtained. This indicates the ability of the model to identify this disease without any omission errors. Additionally, the mAP@50 and mAP@50-95 values for this class are 0.995 and 0.882, respectively, which confirm the model's accuracy at all Intersection over union (IoU) thresholds. The angular leaf spot class also demonstrated good performance, falling just short of balance between precision (0.905) and recall (0.904). Additionally, the mAP@50 and mAP@50-95 values for this class are 0.927 and 0.760, respectively, indicating the practical identification of this disease at various levels of overlap. The leaf spot class with the highest number of samples (223) also exhibits strong performance, with a precision of 0.907, recall of 0.914, and mAP@50 of 0.943, confirming that the model has experienced improved learning and generalization with increasing data volume. On the other hand, some classes, such as anthracnose and gray mold, suffer from an imbalance in precision and recall. In the anthracnose class, the high precision (0.952) indicates the ability of the model to avoid type I errors, but the lower recall (0.800) indicates the possibility of undetected samples. Similarly, in the gray mold class, the precision is 0.897, and the recall is 0.812, indicating some challenges in extracting the unique features of this disease. The lower value of mAP@50-95 in this class (0.628) indicates that the model suffers from performance degradation at different levels of spatial accuracy. In the healthy class, the model has a perfect recall (0.942), indicating that almost all healthy samples are correctly identified. However, the lower precision in this class (0.799) suggests that some diseased samples are falsely diagnosed as healthy, which can be risky in real applications, especially in prevention processes. Finally, the fruit and leaf powdery mildew classes have the weakest performance among all available classes. The precision of 0.816 and recall of 0.725 for the fruit powdery mildew class indicate a serious challenge for the model in accurately diagnosing this disease. In particular, the mAP@50-95 value of 0.689 also highlights that the model lacks the necessary stability across different detection scales. Possible reasons for this poor performance may include the lack of data in the relevant classes, the apparent similarity with other classes, e.g., leaf powdery mildew and leaf spot, and the insufficient visual diversity in the dataset.
Conclusion
This study led to the development of an advanced strawberry disease detection system based on an updated YOLO11L architecture, which achieved an average precision of 90.9% in the mAP@50 benchmark. The proposed model performed very well in identifying diseases with distinct visual symptoms, such as blossom blight (99.5% precision) and leaf spot (94.3% precision). However, a relative decrease in accuracy was observed when classifying diseases with similar visual symptoms, such as fruit powdery mildew (82.7% precision) and gray mold (89.1% precision). This was mainly due to two key factors: (1) insufficient training data for recent, classes and (2) high overlap in visual patterns between them. From an applied perspective, the presented model has significant potential in improving plant disease management solutions, through applications such as intelligent monitoring of farms and greenhouses, integration with unmanned aerial systems for large-scale surveillance, and reducing untargeted pesticide use through accurate and situational disease detection.