Implementations of A Robust Deep Learning-Based Approach for Early Crop Disease Detection in Smart Agriculture

The importance of agriculture in securing food security cannot be understated. However, various kinds of diseases that affect leaves and fruits of tomatoes often lead to lower yields and reduced quality, causing significant financial losses. Therefore, fast and accurate identification of tomato diseases is essential in crop management and controlling diseases spread. In this study, an automatic system based on the YOLO11n-cls model trained using transfer learning and fine tuning will be used for the identification of tomato diseases. As the models efficiency needs to be measured, multiple YOLO models such as YOLOv8n, fine-tuned YOLOv8n, transfer learning YOLOv8n, YOLOv8n-cls, and YOLO11n-cls were applied on a customized tomato disease dataset that includes healthy and unhealthy classes of tomatoes. Accuracy, precision, recall, the number of model parameters, and its computational cost (GFLOPS) were used to evaluate models. Based on the obtained results, the YOLO11n-cls model outperformed all others with accuracy, precision, and recall values of 98.4%, 97.9%, and 97.4%, respectively, being lightweight with 1.55 million parameters and only 3.3 GFLOPS. Development and testing of the model and experimentation were carried out in Google Colab and Kaggle platforms with NVIDIA T4 GPU support. Furthermore, a website powered by the Gradio library was created for uploading tomato leaves and fruit images and their subsequent automatic classification with confidence scores. The results of experimental research, including confusion matrices and model efficiency assessment, proved the effectiveness of the suggested model.