Comparative study of deep learning approaches for cucumber disease classification

Abstract

Cucumber leaf diseases, such as downy mildew and leaf miner, pose significant challenges to crop yield and quality. Accurate and timely detection is essential to efficient management. The current research assesses seven convolutional neural network (CNN) models for the classification of diseases of cucumber leaves: DenseNet121, InceptionV3, ResNet50V2, VGG16, Xception, MobileNetV2, and NASNet. The dataset includes images from the cucumber disease recognition dataset (Mendeley) and 500 real-time images captured between December 2022 and February 2023 in Karnataka, covering varied lighting conditions. After augmentation, the dataset is divided into testing, validation, and training sets and includes 804 leaf miner, 807 downy mildew, and 804 healthy images. With an overall test accuracy of 99.37% and nearly flawless precision, recall, and F1-scores in every class, ResNet50V2 showed exceptional performance. InceptionV3 and MobileNetV2 also exhibited strong performance with accuracies of 97.29% and 97.70%, respectively. DenseNet121, VGG16, Xception, and NASNet performed well but were slightly outperformed by the top models. The findings indicate ResNet50V2 as the most reliable model for cucumber leaf disease classification, providing a robust foundation for developing automated disease detection systems. This work demonstrates how precise disease detection using deep learning models can improve agricultural management.