DLCA-CapsNet: Dual-Lane CDH Atrous CapsNet for the Detection of Plant Diseases

Humanity's survival, development, and existence are deeply intertwined with agriculture, the source of most of our food. Plant disease detection helps in securing food, but manual plant disease detection is error-prone and labor-intensive. Convolutional Neural Networks (CNNs) are highly effective for automated plant disease classification, but their difficulty in recognizing differently oriented images means they need large datasets with many variations to work best. Capsule Networks (CapsNets) were developed to overcome the shortcomings of CNNs and can function effectively with smaller datasets. However, CapsNets process every part of an input image, so their performance can suffer when dealing with complex visuals. To tackle this challenge, DLCA-CapsNet was introduced. DLCA-CapsNet integrates a Color Difference Histogram (CDH) layer for key feature extraction, atrous convolution layers to enlarge receptive fields while maintaining spatial details, along with max-pooling, standard convolutional layers, and a dropout layer. The proposed DLCA-CapsNet method was evaluated on datasets including apple, banana, grape, maize, mango, pepper, potato, rice, tomato, as well as CIFAR-10 and Fashion-MNIST. The model demonstrated strong performance with high test accuracies in plant disease detection and on CIFAR-10 and Fashion-MNIST. It improved test accuracies by 6.78%, 14.82%, 6.14%, 5.07%, 21.12%, 40.32%, 4.64%, 0.76%, 10.23%, 13.73%, and 2.03%, while also reducing the number of parameters in millions by 6.16M, 6.16M, 6.16M, 6.16M, 7.14M, 5.68M, 5.92M, 7.62M, 7.62M, and 6.54M respectively when compared with the original CapsNet. On sensitivity, F1-Score, precision, specificity, Receiver Operating Characteristics, Precision-Recall values, accuracy, disk size, and parameters generated, etc., the DLCA-CapsNet achieved better performance compared to the original CapsNet and other advanced CapsNets reported in the literature. The findings suggest that this efficient and computationally less demanding method can significantly enhance plant disease classification and contribute incrementally to efforts aligned with the SDG 2 goal by offering a lightweight, scalable solution that can be adapted for field use in resource-constrained settings.