EC-LDM: An Enhanced Class-Conditioned Latent Diffusion Framework with Self-attention for Explainable Potato Leaf Disease Classification

Thakre, Pankajkumar, Chakole, Jagdish
Potato Research · Springer Nature · 2026-10-24 · DOI: 10.1007/s11540-026-10168-y

Аннотация

Reliable potato leaf disease classification depends not only on classifier design but also on how disease-specific visual information is represented. Conventional deep learning classifiers mainly learn discriminative features from labelled images, while diffusion models in agricultural applications have largely been used for image synthesis and data augmentation. To address this gap, this study proposes a 1024-D Enhanced Class-Conditioned Latent Diffusion Model (EC-LDM) that employs latent diffusion directly for disease-aware representation learning and potato leaf disease classification. The framework integrates residual feature extraction, self-attention, and class conditioning with latent diffusion to learn structured disease-discriminative representations. Reconstruction, diffusion, classification, and KL-regularisation objectives are jointly optimised to refine the latent representation. Ablation studies compare frozen and fine-tuned 512-D configurations with the proposed 1024-D configuration to examine the effects of latent dimensionality and encoder optimisation. Experiments were conducted on the publicly available potato leaf disease (PLD) dataset, which contains early blight, late blight, and healthy classes. Five-fold stratified cross-validation was employed to evaluate the proposed model against eight Bayesian-optimised deep learning classifiers. Across five-fold stratified cross-validation, the proposed model achieved a mean accuracy of 98.69%, with weighted precision, recall, and F1-score of 98.70%, 98.69%, and 98.69%, respectively, and an AUC of 99.87%. Statistical testing supported the observed performance differences, while t-SNE, confusion-matrix analysis, and expert assessment of Grad-CAM and SHAP visualisations provided further evidence on latent-space structure, classification behaviour, and explanation relevance. These findings indicate that class-conditioned latent diffusion can serve as a disease-aware representation-learning mechanism for direct classification of potato leaf disease, beyond its conventional use in image synthesis and data augmentation.

Читать у издателя →

Похожие статьи