A resource-efficient framework for plant disease classification: integrating reduced-order modeling with...

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Automated plant disease detection faces challenges like subtle symptoms and high visual similarity.
• Deep learning models offer accuracy but require substantial computational resources, limiting field deployment.

Purpose of the Study:

• To develop a computationally efficient framework for plant disease diagnosis on resource-constrained hardware.
• To integrate Reduced Order Modelling (ROM) with deep learning and classical classification for improved detection.

Main Methods:

• A novel ROM framework was proposed, combining a YOLOv8m backbone for feature extraction and PCA for compression.
• A treatment-based label engineering approach consolidated the PlantWildV2 dataset into 11 classes.
• Support Vector Classifier (SVC) was tuned for classification on the compressed feature space.

Main Results:

• Accuracy peaked at 100 principal components, indicating a natural regularization effect of feature compression.
• The ROM framework achieved 87.52% test accuracy and a 0.882 macro F1-score.
• The framework demonstrated a 670-fold reduction in training time compared to EfficientNet-B0 (30.8s CPU vs. 5.8h GPU).

Conclusions:

• Reduced Order Modelling (ROM) is a viable approach for accurate and efficient plant disease detection on low-resource hardware.
• The proposed framework surpasses existing models like EfficientNet-B0 in both accuracy and computational efficiency.
• This work facilitates the deployment of advanced plant disease diagnostic tools in field settings.