RST-Net: Advancing Plant Disease Prediction Using Enlightened Swin Transformer Networks

By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.Apr 24 2024

In a recent paper published in the journal Scientific Reports, researchers introduced a novel deep learning model named residual convolutional enlightened Swin transformer networks (RST-Net) for detecting and classifying plant diseases using images of plant leaves. Their technique underwent testing on a benchmark dataset comprising 14 plants with 38 disease classes and surpassed other state-of-the-art hybrid learning models. Furthermore, the research explored the potential applications and challenges of the proposed model in smart agriculture and precision farming.

Background

Plant diseases are one of the major factors that affect the growth and productivity of crops and, consequently, the food security and economic development of a nation. Early and accurate diagnosis of plant diseases is crucial for preventing the spread of infections and reducing the losses caused by them. However, traditional methods of plant disease detection, such as visual inspection, laboratory analysis, and expert consultation, are time-consuming, costly, and often inaccurate. Therefore, there is a need for developing intelligent and automated systems that can recognize plant diseases using digital images of plant leaves.

In recent years, deep learning techniques, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and deep belief networks (DBNs), have been widely used for plant disease detection and classification. These techniques can automatically learn the deep features of plant leaves from large-scale image datasets and achieve high performance in various tasks, such as segmentation, localization, and recognition. However, these techniques also face some challenges, such as high computational complexity, overfitting, and gradient vanishing or exploding problems. Moreover, these techniques may not be able to capture the global and contextual information of plant leaves, which are important for distinguishing different disease classes.

About the Research

In this paper, the authors designed RST-Net by combining residual convolutional networks and Swin transformers to extract deep features and achieve high accuracy in plant disease classification. Residual convolutional networks, a variant of CNNs, employ shortcut connections to skip some layers and avoid gradient vanishing or exploding issues. In contrast, Swin transformers utilize hierarchical structures and sliding window-based self-attention mechanisms to capture both local and global information from input patches.

The proposed method comprises three key components: data augmentation, feature extraction, and classification. Data augmentation involves increasing the number and diversity of training images through transformations like brightness adjustment, rotational adjustment, and offline transformation, enhancing the model's robustness and generalization ability.

Feature extraction serves as the core component, where integrating residual convolutional networks and Swin transformers facilitates the extraction of deep features from plant leaves. The residual convolutional networks consist of 20 layers, initially featuring 16 kernel filters in the first convolutional layer.

Subsequently, batch normalization, leaky rectified linear unit (LeakyReLU) activation, and average pooling layers are applied. Following the 16th layer, Swin transformers are incorporated and structured into four stages, each comprising a patch merging layer and a Swin transformer layer.

The patch merging layer aggregates features from adjacent patches, facilitating a hierarchical reduction in token numbers. Meanwhile, the Swin transformer layer employs sliding window-based multi-head self-attention (MHSA) and feed-forward networks to process patch features, resulting in a 1D vector representing deep features extracted from the input image.

In the classification stage, extracted deep features are inputted into a dense neural network with extreme learning machines (ELMs) for multiple classification of plant diseases. ELMs are characterized by their ability to randomly generate weights and biases for the hidden layer, determining output weights analytically using the Moore-Penrose generalized inverse, thus achieving fast learning, high accuracy, and low computational complexity.

The presented model was trained and tested on the PlantVillage dataset, an open-access repository containing 54,306 images of 14 plants with 38 disease classes. The dataset was split, with 70% for training, 20% for validation, and the remaining 10% for testing. During training, the model underwent 120 epochs using the adaptive moment estimation (ADAM) optimizer, cross-entropy loss function, and SoftMax activation function. Following this, its performance was assessed using accuracy, precision, recall, specificity, and F1-score.

Research Findings

The outcomes showed that the RST-Nets model achieved an accuracy of 99.95%, a precision of 99.95%, a recall of 99.96%, a specificity of 99.95%, and an F1-score of 99.95% on the test dataset. These results indicated that the model could effectively recognize and classify plant diseases with high accuracy and reliability.

The model also outperformed other state-of-the-art hybrid learning models, such as ResNet-50, ResNet-100, GoogleNet, AlexNet, Inception, VGG-19, CapsuleNetworks, and CAPSNETS, in terms of all the performance metrics. The model also showed a uniform performance across different plants and disease classes, demonstrating its robustness and generalization ability.

The new model has potential applications in smart agriculture and precision farming, where it can provide timely and accurate diagnosis of plant diseases using digital images of plant leaves. This can help farmers prevent the spread of infections, reduce the losses caused by diseases, and improve crop yield and quality.

Moreover, the model's integration with other technologies like the Internet of Things (IoT), drones, or mobile devices facilitates remote and automated monitoring and management of plant health. Furthermore, it can be extended to other crops or regions, by using more diverse and representative datasets and adapting the model parameters accordingly.

Conclusion

In summary, the novel model demonstrated effectiveness in plant disease recognition. It successfully managed noise and variations in leaf images, maintained consistent performance across different runs, and identified important features for disease recognition.

Moving forward, the researchers acknowledged limitations and challenges and proposed directions for future work. They suggested enhancing the model's capability to classify symptom severity or disease progression stages, optimizing the model for real-time data or resource-constrained devices, and exploring alternative deep-learning techniques or architectures for plant disease recognition.