Identification and severity classification of tomato diseases based on two-stage network model

Recognizing crop disease types and assessing severity are crucial for ensuring food security, managing disease spread, and estimating economic impacts. Deep learning techniques have been successful in this field, but approaches relying only on a single model have limitations due to the complexity of the task. Therefore, Therefore, this study introduced a two-stage model that combines MobileNetV3 and U-Net for identifying tomato diseases and classifying their severity. In the first phase, for tomato disease recognition, this study redesigned the attention mechanism and activation function of the MobileNetV3 model and introduced transfer learning and Focal loss. The identified diseases are then grouped and entered into the next stage. In the second stage, various deep learning algorithms with different backbone structures were first used to segment the tomato leaf and lesion regions. Finally, the severity of the disease was classified by calculating the ratio of disease spots pixel area to leaf pixel area. The results showed that the classification accuracy of the improved MobileNetV3 model for diseases reached 97.28%, with a model size of 4.06 MB, which was better than other models. The U-Net with ResNet backbone performed the best, achieving MIoU and MPA of $\mathbf{9 0 . 4 8 \%}$ and $\mathbf{9 5 . 3 9 \%}$, respectively. The average accuracy for disease severity of tomato leaf mold and bacterial spot disease reached $92.87 \%$ and $91.48 \%$, respectively. This study proves the feasibility of rapid identification of tomato diseases and assessment of severity, which provides technical support for precision application in agriculture.