Revealing Hierarchical Structure of Leaf Venations in Plant Science via Label-Efficient Segmentation: Dataset and Method

Weizhen Liu, Ao Li, Ze Wu, Yue Li, Baobin Ge, Guangyu Lan, Shilin Chen, Minghe Li, Yunfei Liu, Xiaohui Yuan, Nanqing Dong

Abstract

The architecture of leaf veins serves as a vital indicator of plant physiological health, yet hierarchical analysis has been limited by the extreme labor required for manual annotation. This paper presents HALVS, the first public dataset for hierarchical leaf vein segmentation, comprising over 5,000 high-resolution scans across three species. To address the "annotation tax," we introduce a Partially Supervised Semantic Segmentation (PSSS) module. This label-efficient approach leverages a novel class-specific exclusion loss to learn intricate tertiary structures from partially labeled data. Our method achieves state-of-the-art performance, significantly improving tertiary vein segmentation accuracy while reducing human effort.

Introduction

In plant science, leaf venation networks—the "plumbing" of the plant—are essential for understanding photosynthesis, nutrient transport, and drought resistance. Botanists categorize these networks into a hierarchy: primary (1°), secondary (2°), and tertiary (3°) veins. While deep learning has advanced general image segmentation, hierarchical leaf vein analysis remains a bottleneck due to data scarcity and the immense effort required to label dense 3° networks. For instance, fully annotating a single London planetree leaf can take an expert nearly eight hours.

Hierarchical vein orders (1°, 2°, 3°) across three species: (a) Soybean, (b) Sweet cherry, and (c) London planetree.

The HALVS Dataset

To facilitate research in this area, we introduce the HierArchical Leaf Vein Segmentation (HALVS) dataset. It contains 5,057 high-resolution scans of soybean, sweet cherry, and London planetree leaves. Images were captured using a flat-bed scanner in transmission mode to maximize contrast. The dataset includes 150 fully annotated images (50 per species), representing a total labeling effort of 83.8 person-days. The remaining images provide a vast pool for semi-supervised and partially supervised learning.

Methods: Label-Efficient Segmentation

We propose the Partially Supervised Semantic Segmentation (PSSS) module, a pluggable framework compatible with existing semi-supervised architectures. PSSS is specifically designed to handle data where only 1° and 2° veins are labeled.

The module employs three primary loss functions:

  1. Supervised Loss ($L_P^s$): Applied to known 1° and 2° vein pixels.
  2. Unsupervised Loss ($L_P^u$): Utilizes high-confidence pseudo-labels for background and 3° veins generated by a teacher model.
  3. Class-Specific Exclusion Loss ($L_P^c$): This novel loss exploits the mutual exclusivity of classes. If the model predicts a pixel is neither background nor a 1°/2° vein, it logically infers the pixel belongs to the 3° vein class. This allows the model to learn complex tertiary patterns even without explicit labels.

Results and Discussion

The PSSS module demonstrates significant improvements over standard supervised and semi-supervised baselines. When integrated with the FixMatch architecture, PSSS improved the Intersection-over-Union (IoU) for 3° veins by 14.37% and the overall mean IoU (mIoU) by 8.55%.

Qualitative results showing original leaf crops (top) and their corresponding hierarchical segmentation masks (bottom) for (a) Soybean, (b) Sweet cherry, and (c) London planetree. Red: 1°, Yellow: 2°, White: 3°.

Qualitative analysis reveals that PSSS-enhanced models produce cleaner, more continuous vein networks. However, challenges remain, particularly regarding "dataset shift" between species; models trained on one species often struggle to generalize to others due to unique geometric patterns.

Conclusion

This work provides the first comprehensive dataset and a label-efficient methodology for hierarchical leaf vein segmentation. By reducing the annotation burden through logical deduction and partial supervision, we enable more scalable plant physiological studies. Future research will focus on cross-species domain adaptation to create a universal leaf vein segmenter.

Summary updated

BibTeX
@inproceedings{liu2024revealing,
  author = {Weizhen Liu and Ao Li and Ze Wu and Yue Li and Baobin Ge and Guangyu Lan and Shilin Chen and Minghe Li and Yunfei Liu and Xiaohui Yuan and Nanqing Dong},
  title = {{Revealing Hierarchical Structure of Leaf Venations in Plant Science via Label-Efficient Segmentation: Dataset and Method}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure