Research and Analysis on Severity Grading of Early Blight in Tomato Based on Improved MobileNetV3

Early blight of tomato is a major foliar disease in tomato production. Traditional methods relying on expert-based manual diagnosis and severity assessment are inefficient and subjective, making them unsuitable for real-time field monitoring. This paper constructs a severity classification dataset for early blight of tomato and proposes a lightweight automatic classification model based on an improved MobileNetV3-Large. By reconstructing the classification head and introducing intermediate fully connected layers and double Dropout, combined with an improved loss function, sampling strategy, and various data augmentations, class imbalance is mitigated and the model’s generalization performance is improved. Results show that the model achieves an overall accuracy and weighted F1 score of 94.9% on the four severity test set, with recalls of 1.00, 0.93, 0.90, and 0.74 for healthy, mild, moderate, and severe cases, respectively. False positives are mainly concentrated between moderate and severe. This method achieves high classification accuracy while maintaining a lightweight model, providing technical support for field monitoring and precise control of early blight of tomato.