A PCA-based deep feature optimization framework for explainable orange fruit disease classification

Abstract

Accurate classification of orange fruit diseases is important for precision agriculture and yield protection. This study develops and rigorously benchmarks a hybrid deep-feature framework for classifying Black Spot, Canker, Fresh, and Greening oranges (1,090 images), combining deep feature extraction, PCA-based dimensionality reduction, and classical machine-learning classification. Eight backbones (seven CNNs and a Vision Transformer, ViT) and four classifiers (32 configurations in total) were evaluated under 5 × 5 repeated stratified cross-validation, with PCA fitted exclusively on training-fold features in every iteration to eliminate data leakage. The proposed ViT + PCA+SVM configuration achieved the highest mean accuracy, 99.12%±0.71%, significantly outperforming every CNN-based backbone, including DenseNet201 + PCA + SVM (98.48%±0.81%, p < 0.001). A dedicated variance-retention sensitivity analysis justifies the 98% threshold used throughout, and ablation experiments confirm that PCA substantially reduces feature dimensionality (by ~ 55.7% for ViT and ~ 76.6% for DenseNet201) without a significant loss of accuracy for either backbone. Explainability analysis — occlusion sensitivity and SHAP for the proposed ViT model, and Grad-CAM and SHAP for the DenseNet201 comparison model — shows that both configurations base predictions on biologically relevant, disease-affected regions of the fruit rather than spurious cues. These results identify ViT + PCA+SVM as the most accurate configuration evaluated, with DenseNet201 + PCA + SVM as a closely competitive, more compact convolutional alternative for intelligent orchard disease-monitoring systems.