Unassailable citrus disease classification via multi-stage deep ensemble learning with vision transformers

To reduce losses from agriculture as well as enhance food security, we propose a three-stage deep ensemble for early citrus disease diagnosis from actual-field images of oranges (n = 2,240) as well as lemons (n = 208). To prevent leakage, augmentation is strictly enforced following splitting (70:30 stratified) following curation, normalisation, resizing by 224 × 224. Our contribution comes from combining state-of-the-art deep features by adding explicit texture priors. Namely, we add Local Binary Patterns (LBP) as well as Grey-Level Co-occurrence Matrix (GLCM) descriptors for micro-textures of lesions (e.g., stippling, scab rims around them, chlorosis encircled by veins) as well as second statistics (e.g., contrast, homogeneity, entropy) that CNNs/ViTs tend to discount by virtue of their small size coupled with variable-field data. These hand-crafted signals are z-score normalised as well as PCA-compressed for overfit protection as well as removal of collinearity then combined by deep embeddings. InceptionV3 (90% lemon) as well as DenseNet121 (93% orange) are the best of the five pretraining CNNs (ResNet50, DenseNet121, VGG16, InceptionV3, EfficientNetB0) that we test at Stage-1. The best CNNs are enlisted with a Vision Transformer (ViT) at Stage-2 for capture of long-range contextual capture improving upon Stage-1 by 98% (lemon) as well as 97% (orange). t-SNE confirms class separation while Stage-3 employs a multiclass SVM over the combined description that achieves 99% (lemon) while holding at 97% (orange) at another curation. The pipeline outperforms single-backbone variants, minimises variance while remaining lightweight enough for deployment, thus showing that LBP + GLCM texture priors compressed by PCA but combined by CNN/ViT features substantially enhance robustness plus generalisation for in-orchard citrus disease testing.