Bean Leaf Lesions Image Classification: A Robust Ensemble Deep Learning Approach

Abstract

Growing beans is important since they are a staple meal for so many people throughout the world. Bean rust and angular leaf spot are just two of the many diseases that threaten the well-being of bean crops and, in turn, cause considerable output losses. In this research, an ensemble deep learning strategy named EnDeel, is proposed to solve the problem of reliably identifying bean leaf lesions as healthy, angular leaf spots, or bean rust. Five different deep convolutional neural network architectures (MobileNetV2, ResNet50, EfficientNetB2, DenseNet121, and VGG16) are trained and have their parameters initialized via transfer learning. Images of bean leaf lesions are fed into these models to extract relevant features, and the fully connected layer was classified using softmax. By using majority voting, the predictions from the top three deep learning architectures are combined to construct the EnDeeL ensemble classifier. To gauge how well each deep learning classifier did, it is compared to the ensemble classifier EnDeeL. The findings show that EnDeeL outperformed the examined single deep-learning classifiers with an astounding 92.12% test accuracy. This performance improvement demonstrates the usefulness of the ensemble strategy, which increases classification accuracy when compared to that of individual classifiers.