Evaluating multitask deep learning for multi-crop plant health recognition under seed variation and crop transfer

Architectural claims in image-based plant-disease recognition usually rest on a single training run per configuration, and disease severity is rarely measured independently of the classifier. We compared four output formulations for joint crop–health recognition on a shared frozen ResNet50 backbone: single-task joint classification, multitask learning with independent heads, multitask fusion, and multitask fusion conditioned on crop and health posteriors. Each was trained with five seeds (twenty models) on 5,235 leaf images of betel vine, brinjal, ladies finger and Malabar spinach. An ordinal severity reference for 96 images was built by independent dual assessment and adjudication, and all models were evaluated on 3,295 unseen potato images. Single-task and both fusion formulations were statistically indistinguishable (joint accuracy 89.36–89.77%); posterior conditioning had no detectable effect (+0.18 points, p=0.475), whereas independent heads underperformed in health-state accuracy (−4.73 points, p=0.019). Seed variation was comparable to formulation differences. Binary labels diverged from adjudicated severity on 18.75% of reference images. On potato, all models discriminated well (ROC-AUC 0.919–0.956) but needed higher thresholds, and formulation rankings reversed between field and repository images. Model outputs, though trained on binary labels, correlated with adjudicated severity (Spearman ρ 0.605–0.715).