Rose leaf disease classification and severity estimation using an interpretable vision transformer-based multi-task framework

This study proposes RoViT-KAN, a multi-task deep learning framework for plant disease classification, severity estimation, and uncertainty quantification. The architecture integrates a DeiT-Tiny Vision Transformer backbone with task-specific heads for disease classification, ordinal severity prediction, and heteroscedastic uncertainty estimation. To enhance interpretability, a Kolmogorov–Arnold Network (KAN) module is introduced to model continuous disease severity through learnable spline-based transformations. A four-stage curriculum learning strategy is employed to stabilize multi-task optimization by progressively activating prediction objectives. The model is evaluated on a rose leaf disease dataset comprising 3,113 original images and 10,000 augmented samples across four classes: healthy leaf, leaf holes, black spot disease, and dry leaf condition. Experimental results demonstrate a classification accuracy of 99.70%, with calibrated uncertainty estimates (Brier score = 0.0914) and reliable severity prediction. Ablation studies validate the contribution of each architectural component. The model highlights the potential of combining transformer-based architectures with uncertainty-aware learning and interpretable neural representations for robust plant disease analysis.