In plant science, understanding the hierarchical structure of leaf venations is crucial for insights into plant physiology, evolution, and ecology. However, data-driven segmentation methods are hampered by the lack of specialized datasets for hierarchical leaf vein analysis. To address this, we introduce the HierArchical Leaf Vein Segmentation (HALVS) dataset, the first of its kind, containing 5057 high-definition scanned leaf images from three species with 83.8 person-days of human annotations across three vein levels. We propose a novel label-efficient hierarchical segmentation framework combining Partially Supervised Semantic Segmentation (PSSS) and Denoising Diffusion Label Refinement (DDLR). PSSS classifies leaf pixels using primary and secondary vein annotations to generate high-confidence pseudo-labels for the background and tertiary veins, reducing omission errors. DDLR then refines these pseudo-labels, propagating structural priors from sparse veins to accurately recover tertiary veins. This framework significantly improves the integrity and connectivity of tertiary veins at low annotation costs. We also pioneer cross-species learning, training models on easily-annotated species and applying them to difficult ones. Despite challenges, DDLR remarkably enhances segmentation performance across all vein levels, providing an effective solution for complex hierarchical patterns and advancing agricultural research.

Wheat stripe rust poses a marked threat to global wheat production. Accurate and effective disease severity assessments are crucial for disease resistance breeding and timely management of field diseases. In this study, we propose a practical solution using mobile-based deep learning and model-assisted labeling. StripeRust-Pocket, a user-friendly mobile application developed based on deep learning models, accurately quantifies disease severity in wheat stripe rust leaf images, even under complex backgrounds. Additionally, StripeRust-Pocket facilitates image acquisition, result storage, organization, and sharing. The underlying model employed by StripeRust-Pocket, called StripeRustNet, is a balanced lightweight 2-stage model. The first stage utilizes MobileNetV2-DeepLabV3+ for leaf segmentation, followed by ResNet50-DeepLabV3+ in the second stage for lesion segmentation. Disease severity is estimated by calculating the ratio of the lesion pixel area to the leaf pixel area. StripeRustNet achieves 98.65% mean intersection over union (MIoU) for leaf segmentation and 86.08% MIoU for lesion segmentation. Validation using an additional 100 field images demonstrated a mean correlation of over 0.964 with 3 expert visual scores. To address the challenges in manual labeling, we introduce a 2-stage labeling pipeline that combines model-assisted labeling, manual correction, and spatial complementarity. We apply this pipeline to our self-collected dataset, reducing the annotation time from 20 min to 3 min per image. Our method provides an efficient and practical solution for wheat stripe rust severity assessments, empowering wheat breeders and pathologists to implement timely disease management. It also demonstrates how to address the “last mile” challenge of applying computer vision technology to plant phenomics.