Plant disease identification using explainable 3D deep learning on hyperspectral images.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Hyperspectral imaging offers a promising avenue for plant disease identification.
• Deep learning models are well-suited for analyzing complex hyperspectral data cubes.
• Soybean charcoal rot is a significant soil-borne fungal disease impacting crop yields globally.

Purpose of the Study:

• To develop and evaluate a novel 3D deep convolutional neural network (DCNN) for hyperspectral-based plant disease identification.
• To interrogate the DCNN model for physiologically meaningful explanations of its predictions.
• To focus on the identification of charcoal rot in soybean crops.

Main Methods:

• Deployment of a novel 3D deep convolutional neural network (DCNN) for direct assimilation of hyperspectral data.
• Utilizing hyperspectral imaging of soybean stems (inoculated and mock-inoculated).
• Employing saliency maps to visualize sensitive pixel locations and wavelengths.

Main Results:

• The 3D DCNN achieved a classification accuracy of 95.73% and an F1 score of 0.87 for the infected class.
• Saliency maps confirmed that the model focused on spatial regions exhibiting visible disease symptoms.
• The model identified near-infrared (NIR) wavelengths as most sensitive for classification, correlating with plant vegetative health.

Conclusions:

• Explainable deep learning models provide high accuracy and physiological insights for plant disease identification.
• The model's predictions generate confidence, enabling applications in precision agriculture.
• Automated phenotyping platforms can leverage these explained predictions for research and practical applications.