In the agricultural sector, the pursuit of heightened crop production remains a pivotal response to the ongoing challenge of global food insecurity. However, this pursuit is entwined with the omnipresent threat of plant diseases and pests. This underscores the urgency for proactive measures such as early detection and effective treatment strategies.
Study: Revolutionizing Plant Disease Detection with Optimized Deep Learning: GJ-GSO-based DbneAlexNet. Image credit: Living Thoughts/Shutterstock
In a recent publication featured in the Journal of Biotechnology, researchers introduced a novel deep learning (DL) model for precise plant disease detection and classification. The technique focuses on tomato leaf disease detection using deep batch-normalized eLu Alex Net (DbneAlexnet).
Background
The agriculture sector is currently witnessing an increase in crop production, accompanied by heightened concerns about food security due to the global issue of plant diseases. Modern agricultural practices prioritize improving both crop yield and quality while considering environmental factors. As instances of crop occurrences increase, the complexity of diseases also increases, necessitating expert observation for disease detection. However, manual identification based on visual inspection is influenced by environmental conditions and subjectivity, thus requiring specialized skills. Automated disease detection remains a challenge, emphasizing the need for advanced diagnostic and preventive techniques.
Plant diseases hold the potential to impact economies, societies, and ecosystems, posing a threat to agriculture if untreated. Effective disease management relies on early detection and prompt treatment. Notable examples include tomato blight caused by fungi and diseases such as the tomato mosaic virus affecting leaves in the early growth phase. Traditional computer vision methods, including various segmentation techniques, are used for identifying leaf lesions through noise reduction, feature processing, and lesion selection. In recent times, deep learning, especially convolutional neural networks (CNNs), has revolutionized disease detection by automating feature extraction, boosting accuracy, and finding applications beyond agriculture.
Related work
Agriculture plays a critical role in the economic prosperity of numerous nations. Yet, the onslaught of plant diseases presents formidable obstacles for farmers, impacting plant quality, productivity, and quantity. The efforts focus on developing a disease identification technique to amplify productivity. Noteworthy attempts in this realm include the utilization of machine learning for tomato disease image classification, which grappled with complexity in predictions.
The use of CNN for precise early-phase disease detection lacked subsequent accuracy enhancements. Employing DL for disease detection fell short of integrating segmentation and localization. Struggles with varied image sizes emerged while using residual net-50 (ResNet50) for leaf classification. Extending DL for disease prediction was clouded by transparency issues. Applying EfficientNet for disease classification lacked real-time efficiency.
A CNN baseline for leaf classification faced accuracy hurdles. Employing CNN for early disease detection encountered overfitting due to limited data.
Incorporating deep models for plant disease detection is fraught with challenges encompassing constrained database access, initial segmentation complexities, parameter intricacies, unclear visualization, and the rare disease classification dilemma owing to dataset scarcity.
Proposed method
The current study introduces an innovative technique, gradient Jaya-golden search optimization (GJ-GSO)-based DbneAlexNet, for the classification of tomato leaf diseases. In agricultural economies, early disease detection is critical for preventing economic losses and ensuring agricultural sustainability. The primary objective is to identify diseases in tomato plant leaves using the GJ-GSO approach. Initially, tomato leaf images from a specific dataset undergo preprocessing through anisotropic filtering to correct distortions. The segmentation is performed using U-net, trained with the Gradient-GSO Algorithm, which combines GSO and Gradient concepts.
The segmented images undergo image augmentation, involving adjustments in both position and color. Position augmentation includes padding, rotation, affine transformations, cropping, and scaling, while color augmentation encompasses modifications in saturation, contrast, and hue. Finally, the DbneAlexnet classifies multiclass plant leaf diseases, trained using the proposed GJ-GSO method, which integrates Gradient, Jaya, and GSO algorithms to enhance performance.
Study Results
The current study evaluated GJ-GSO-based DbneAlexNet's efficacy using metrics such as true positive rate (TPR), false positive rate (FPR), true negative rate (TNR), and accuracy. Assessments manipulate learning data and use K-fold validation. The experimental setup implements GJ-GSO-based DbneAlexNet in Python. The dataset includes various tomato leaf diseases for detection. Algorithm comparisons include particle swarm optimization (PSO+DbneAlexNet), competitive swarm optimizer (CSO+DbneAlexNet), Jaya+DbneAlexNet, golden search optimization (GSO+DbneAlexNet), and the proposed GJ_GSO+DbneAlexNet.
Results indicate GJ_GSO+DbneAlexNet's superiority, with accuracy improvements of 19.450%, 18.131%, 4.395%, and 12.307% over existing methods. Similar trends occur in TPR, TNR, and FPR. Learning rate and K-fold analyses also affirm GJ_GSO+DbneAlexNet's consistently enhanced performance. The proposed approach demonstrates improved accuracy, TPR, TNR, and FPR, highlighting its overall effectiveness.
Conclusion
In summary, the researchers introduced a novel, optimized deep model for accurate plant leaf disease identification. Utilizing augmentation filters reduces image noise, while U-Net combined with Gradient-GSO ensures precise segmentation of tomato leaf images. The proposed technique effectively employs position and color augmentation, thus reducing data dimensionality.
Multiclass disease classification involves training DbneAlexNet with GJ-GSO, blending Gradient, Jaya, and GSO algorithms. This approach achieves superior results, including 92.4% accuracy, 91.9% TPR, 92.2% TNR, and a minimal 0.078 FPR. Additionally, the unified segmentation and classification strategy proves effective for disease detection, backed by empirical validation, with the potential for further development using different databases.
Drought Prediction with Deep Learning