Data and code for: Evaluating Multitask Deep Learning for Multi-Crop Plant Health Recognition under Seed Variation and Crop Transfer

<p dir="ltr">This item contains the data records and code supporting the study "Evaluating Multitask Deep Learning for Multi-Crop Plant Health Recognition under Seed Variation and Crop Transfer" (Cogent Food & Agriculture, under review).</p><p dir="ltr">The study compares four output formulations (single-task joint classification, independent multitask heads, multitask fusion, and fusion conditioned on task posteriors) on a shared frozen ResNet50 backbone, each trained with five random seeds (20 models), on a compiled benchmark of 5,235 leaf images of betel vine, brinjal, ladies finger (okra) and Malabar spinach in healthy and unhealthy states.

An ordinal severity reference was built for 96 images by independent dual assessment and pre-specified adjudication. All models were also evaluated on 3,295 images of an unseen crop (potato).</p><p dir="ltr">Contents:<br>(1) manifests_and_audits.zip – the frozen dataset manifest (5,235 images, SHA256 fingerprints), train/validation/test split manifests (seed 42, 70:15:15), provenance table, integrity and duplicate-screening audits, and the potato external-validation manifest and exclusion log;<br>(2) severity_reference.zip – the 96-image assessment batch, both assessors' rating records and event logs, agreement analysis, adjudication records and the final adjudicated severity reference;<br>(3) code_and_results.zip – Jupyter notebooks for dataset auditing, splitting, severity rating and adjudication, model training (TensorFlow 2.20.0, Keras 3.13.2, Python 3.13), the decision-policy simulation and all downstream analyses, with per-model result tables;<br>(4) README.txt – file descriptions and run order.</p><p dir="ltr">Third-party images are not redistributed.