Deep Learning-Based Fruit Disease Detection and Severity Prediction Using VGG-16 and VGG-19 Architectures

Abstract

Fruit diseases cause major economic and agricultural losses worldwide, affecting crop yield and food quality. Conventional visual inspection methods are labor-intensive, subjective, and incapable of early-stage disease detection. The increasing availability of digital imaging equipment and advances in deep learning have opened promising pathways for automated, reliable, and scalable fruit disease detection systems. Identifying not only the type of disease but also its severity is critical for timely intervention, loss minimization, and precision agriculture applications. Despite recent progress, significant challenges persist in fruit disease detection. These include high intra-class variability in disease appearance, limited annotated datasets, sensitivity to lighting and imaging conditions, and the computational cost of deploying large-scale models on resource-constrained farm devices. Furthermore, distinguishing subtle early-stage symptoms from healthy tissue and estimating a continuous severity gradient demands more sophisticated modeling beyond simple binary classification. This study proposes a deep learning framework employing two pre-trained convolutional neural network architectures, VGG-16 and VGG-19, fine-tuned on a curated multi-class fruit disease image dataset comprising five disease categories and healthy instances. A systematic pipeline involving image acquisition, preprocessing, data augmentation, transfer learning, and a dual-output classification and regression head is designed. Models are trained with the WuC-Adam optimizer, and Grad-CAM is used to visualize the regions of activation to provide explainability. Three core algorithms are implemented: (1) a Transfer Learning-based disease classification algorithm using VGG-16/VGG-19 feature extraction, (2) a Severity Regression algorithm using multi-layer perceptron output on extracted features, and (3) a Grad-CAM visualization algorithm for model interpretability. Mathematical formulations including convolution operations, softmax classification, and mean squared error loss are provided for each stage. The proposed VGG-19 model achieved a peak classification accuracy of 97.4%, with a precision of 96.8%, recall of 97.1%, and F1-score of 96.9% on the test dataset. The severity prediction module attained a Mean Absolute Error (MAE) of 2.3% and a Root Mean Squared Error (RMSE) of 3.1%. VGG-19 consistently outperformed VGG-16 across all evaluation metrics and converged faster under the same training conditions. The experimental findings confirm the viability of deep transfer learning for automated fruit disease detection and severity quantification. The proposed dual-output framework offers a practical, interpretable, and highly accurate tool for deployment in precision agriculture contexts. Future work will explore lightweight model variants and real-time mobile deployment.