Phenotypic Analysis of Diseased Plant Leaves Using Supervised and Weakly Supervised Deep Learning

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Traditional plant disease phenotyping often relies on image-level classification.
• Pixel-level analysis of disease features, like spot distribution, offers more detailed insights.
• Deep learning and computer vision are increasingly vital for automated plant disease identification.

Purpose of the Study:

• To analyze pixel-level phenotypic features of diseased plants using deep learning.
• To develop and evaluate both supervised and weakly supervised models for plant disease spot segmentation.
• To assess the generalization capabilities of different segmentation models on diverse plant datasets.

Main Methods:

• Collected and annotated a pixel-level diseased leaf dataset, including apple, grape, and strawberry samples.
• Employed supervised convolutional neural networks (CNNs) for semantic segmentation.
• Explored weakly supervised models, including ResNet-CAM and a few-shot pretrained U-Net classifier (WSLSS), trained with image-level annotations.

Main Results:

• Supervised DeepLab achieved the highest performance (IoU = 0.829) on the apple leaf dataset.
• The weakly supervised WSLSS model achieved an IoU of 0.434 on apple leaves and 0.511 on the extra testing dataset.
• WSLSS demonstrated superior generalization ability on unseen disease types compared to supervised models.

Conclusions:

• Deep learning enables precise pixel-level analysis of plant disease spots.
• Weakly supervised learning offers a promising approach for disease segmentation with reduced annotation costs and enhanced generalization.
• The study provides a valuable dataset and insights for future research in automated plant disease phenotyping.