Plant disease diagnosis in field settings is challenged by subtle symptomology, high inter-class visual similarity, and class imbalance, making automated detection particularly difficult. While deep learning models achieve high accuracy, traditional architectures impose prohibitive computational costs that hinder deployment on resource-constrained hardware. This paper proposes a novel Reduced Order Modelling (ROM) framework integrating a YOLOv8m backbone for spatially sensitive feature extraction, PCA-based compression to isolate the most discriminative features, and classical classification. A treatment-based label engineering approach was applied to consolidate the PlantWildV2 dataset from 115 to 11 agronomically relevant classes. Experimental results showed that a highly compressed feature space acts as a natural regularizer, with accuracy peaking at 100 principal components and declining beyond that threshold. The tuned SVC classifier achieved a test accuracy of 87.52% and a macro F1-score of 0.882, outperforming all other classifiers evaluated. The proposed ROM framework surpassed EfficientNet-B0 in accuracy (87.52% vs. 82.50%) while reducing training time from 5.8 hours on GPU to 30.8 seconds on CPU, a 670-fold efficiency gain, demonstrating the viability of Reduced Order Modelling for plant disease detection on low-resource hardware.
Natij et al. (Mon,) studied this question.