Deep learning for plant bioinformatics: an explainable gradient-based approach for disease detection.

Area of Science:

• Plant bioinformatics
• Computational biology
• Deep learning applications

Background:

• High-throughput omics data and hyperspectral imaging are crucial for understanding plant health.
• Analyzing complex plant data requires advanced computational and statistical methods.
• Automated plant disease detection is essential for agricultural productivity.

Purpose of the Study:

• To develop and evaluate a novel explainable gradient-based convolutional neural network (EG-CNN) model.
• To predict plant disease types using integrated omics data and hyperspectral images.
• To assess the model's accuracy, robustness, and efficiency in plant disease detection.

Main Methods:

• Collected gene expression, metabolite, and hyperspectral image data from plants with four common diseases.
• Developed an EG-CNN model integrating multi-modal plant data for disease prediction.
• Performed hyperparameter tuning, sensitivity analysis, and comparative time efficiency analysis.
• Utilized saliency maps for qualitative analysis of model's internal representations.

Main Results:

• Achieved a test set accuracy of 95.5% for plant disease prediction.
• Demonstrated high resilience to hyperparameter variations with marginal performance changes.
• Showcased faster testing time per sample compared to baseline models (SVM, Random Forest, Logistic Regression).
• Saliency maps confirmed the model's ability to capture key disease-related features.

Conclusions:

• The EG-CNN model effectively integrates omics data and hyperspectral images for accurate plant disease detection.
• The model's robustness and efficiency show significant promise for plant bioinformatics applications.
• Deep learning approaches offer powerful tools for advancing plant health monitoring and management.