Next-Generation Plant Disease Detection: A Efficient Approach to Plant Disease Identification with HW-CNNs and Wasserstein Metrics

Timely identification and treatment of plant diseases are essential for boosting agricultural productivity and reducing economic losses. In this study, we present an innovative deep learning framework for patient-automated plant disease detection using a Hierarchical Wasserstein Convolutional Neural Network (HW-CNN). Specifically, we introduce depth-separable convolutions for computational cost savings and a new Hierarchical Wasserstein Distance (HWD) loss function which improves classification by leveraging inter-class relationships. The model was trained and validated on a large dataset containing 53200 Images across 38 different diseases in 14 different species of plants. Additionally, the proposed methodology provides a detailed description of the preprocessing steps (transformation of colour space to H, S, and V, pixel masking of green pixels), feature extraction using Hu moments, Haralick texture features, and colour histograms. The HW-CNN architecture is based on depth-separable convolutions which have been shown to yield very good performance with fewer parameters. The HWD loss function also helps build a more suitable loss landscape that enables the model to generalise across different types of diseases. The HW-CNN outperformed classical machine learning models (SVM, Random Forest, and Logistic Regression) and other deep learning architectures with an accuracy of 99.19%.The experimental results show that the HW-CNN has an accuracy of 99.19%. The experimental results showed similar improvements in performance,  while significantly reducing complexity compared to existing methods. Throwing light on the effectiveness of advanced deep learning techniques to overcome significant obstacles in plant disease detection, including serendipitous symptoms and climate differences. The novel HW-CNN architecture forms a scalable, low-power circuitry with high energy efficiency that can tremendously benefit real-world scenario applications such as agriculture, reduce potential crop losses, and improve food security in society.