A hierarchical BYOL-based deep learning framework for explainable leaf disease recognition utilizing self-supervised learning and symptom preserving

Accurate leaf disease recognition remains challenging because supervised deep learning models require substantial labeled data and may struggle to distinguish diseases with overlapping color, texture, and lesion patterns. This study proposes a hierarchical self-supervised learning framework that combines Bootstrap Your Own Latent (BYOL) pretraining, an intermediate healthy-versus-diseased fine-tuning stage, and final five-class classification using a ResNet34 backbone. The methodological novelty lies in progressively specializing the learned representation before fine-grained disease classification, rather than transferring BYOL features directly to the multiclass task. In addition, the study quantitatively evaluates whether the selected augmentations preserve disease-relevant structural, color, texture, and lesion information, addressing an assumption that is commonly treated only qualitatively in plant disease studies. The framework was evaluated on a curated chili leaf dataset containing 13,342 images from healthy, cercospora, mites and trips, nutritional deficiency, and powdery mildew classes. Controlled comparisons were conducted against ResNet34 trained from scratch, ImageNet transfer learning, direct BYOL fine-tuning, and an alternative self-supervised method under a consistent experimental protocol. The proposed pipeline achieved a test accuracy of 0.9266, a macro-F1 score of 0.9190, and a macro ROC-AUC of 0.9741. Introducing the binary intermediate stage improved macro-F1 by 0.0111 compared with direct BYOL-to-multiclass fine-tuning, demonstrating a modest but consistent benefit. Post-hoc Grad-CAM analysis further indicated that predictions were generally associated with visually relevant symptomatic regions. These findings establish the significance of hierarchical representation refinement and quantitatively assessed augmentation for label-efficient leaf disease classification, while external validation remains necessary before broader cross-crop or field-level deployment.