Tomato leaf disease classification and parameter optimisation using a hybrid deep learning framework

Abstract

Agriculture underpins national economies, and tomatoes are among the most widely grown vegetables, yet tomato crops are highly susceptible to leaf diseases that reduce yields and farm incomes. Single-backbone convolutional neural networks and transfer-learning approaches often rely on generic features and struggle to separate visually similar diseases. This study proposes a hybrid deep learning framework in which two frozen, ImageNet-pretrained convolutional neural networks (CNNs), EfficientNetB0 and DenseNet121, extract complementary features that are concatenated and passed to a lightweight classifier. The learning rate, number of dense units, and dropout rate of the classifier were selected with a customised, population-based honey badger optimisation search. Images from a public ten-class tomato leaf dataset were resized and normalised, and a balanced subset of 2,000 images was used for training and tuning. Over 35 repeated training runs on a fixed training/validation split, the model achieved a mean validation accuracy of 82.1% (standard deviation 1.3 percentage points). The final model, trained for 15 epochs, reached 87.0% on the validation split and 87.5% (95% confidence interval 85.3--89.4%) on 1,000 independent images that were never used for training or tuning, with a macro-averaged F1-score of 0.875. Viral and bacterial diseases were recognised most reliably, whereas early blight and target spot remained the most difficult classes. The results show that frozen dual-backbone feature fusion provides a compact baseline for tomato leaf disease classification, although the benefit of the hyperparameter search over default settings was not established.