CLDA-Net: A Novel Citrus Leaf Disease Attention Network for Early Identification of Leaf Diseases

Abstract

Efficient and precise identification of plant disease is crucial for disease prevention. Deep learning models have been the main stream methods for plant disease identification. However, the performance has been compromised in extracting the tiny lesion feature of a plant leaf, resulting in low accuracy. Moreover, the leaves of the citrus crop are flimsy and highly susceptible to various diseases, such as canker, blackspot, and greening. To mitigate the same, this paper presents a novel citrus leaf disease attention (CLDA)-Net. To enhance the learning ability of tiny lesion features, the network embeds the convolutional block attention module (CBAM) into the passage layer for extracting the channel and spatial features, which results in avoiding feature redundancy. The performance of the proposed CLDA-Net has been compared on four citrus plant diseases against seven state-of-the-art deep learning models, namely, XceptionNet, DenseNet-121, ResNet-50, VGGNet16, AlexNet, EfficientNet B2, and SoyNet. Eight performance parameters have been considered for performance validation i.e. accuracy, error, precision, recall, sensitivity, specificity, F1-score, and MCC (Matthews correlation coefficient). From experimental results, the proposed model outperforms the compared models with a classification accuracy of 94.74% without augmenting the dataset, and 97.91% on augmenting the dataset.