Comparative Evaluation of Lightweight and Heavyweight Convolutional Neural Network Architectures for Apple Leaf Disease Classification

Abstract

Leaf diseases can significantly reduce both the productivity and quality of crops; therefore, early and accurate diagnosis is required for sustainable farming. Reliance on visual examination by experts is time-consuming and expensive and may result in inconsistent or inaccurate diagnoses. The present study investigated a deep learning approach based on convolutional neural networks (CNNs) for the automatic classification of apple leaves into three classes, namely, Alternaria blotch, Marssonina blotch and healthy leaves. A field dataset of 9,000 images was acquired with a mobile phone in orchards of Himachal Pradesh under natural lighting, from different shooting angles, and with natural backgrounds. Ten pre-trained CNN models, five lightweight and five heavyweights were evaluated and compared using transfer learning, and each model was trained three times with different random seeds to check the statistical reliability of the results. EfficientNetV2B0 and DenseNet201 were identified as the best lightweight and heavyweight models, achieving mean accuracies of 94.30% and 92.96%, respectively. Both models were subsequently fine-tuned using hyperparameters selected through Bayesian optimization, which raised their accuracies to 98.07% and 96.00%, respectively. The fine-tuned EfficientNetV2B0 performed significantly better than the fine-tuned DenseNet201 while using about one-third of the parameters, and running nearly five times faster on a CPU. A well-optimized lightweight network can, therefore, outperform a much larger one and remain practical for mobile and Internet of Things (IoT) use. This study offers a practical and efficient approach with strong potential for adaptation to other crops, and integration into mobile and IoT-based disease diagnosis systems.