StomataSeg: Semi-Supervised Instance Segmentation for Sorghum Stomatal Components

Sorghum is a globally important cereal grown widely in water-limited and stress-prone regions. Its strong drought tolerance makes it a priority crop for climate-resilient agriculture. Improving water-use efficiency in sorghum requires precise characterisation of stomatal traits, as stomata control of gas exchange, transpiration and photosynthesis have a major influence on crop performance. Automated analysis of sorghum stomata is difficult because the stomata are small (often less than 40 $μ$m in length in grasses such as sorghum) and vary in shape across genotypes and leaf surfaces. Automated segmentation contributes to high-throughput stomatal phenotyping, yet current methods still face challenges related to nested small structures and annotation bottlenecks. In this paper, we propose a semi-supervised instance segmentation framework tailored for analysis of sorghum stomatal components. We collect and annotate a sorghum leaf imagery dataset containing 11,060 human-annotated patches, covering the three stomatal components (pore, guard cell and complex area) across multiple genotypes and leaf surfaces. To improve the detection of tiny structures, we split high-resolution microscopy images into overlapping small patches. We then apply a pseudo-labelling strategy to unannotated images, producing an additional 56,428 pseudo-labelled patches. Benchmarking across semantic and instance segmentation models shows substantial performance gains: for semantic models the top mIoU increases from 65.93% to 70.35%, whereas for instance models the top AP rises from 28.30% to 46.10%. These results demonstrate that combining patch-based preprocessing with semi-supervised learning significantly improves the segmentation of fine stomatal structures. The proposed framework supports scalable extraction of stomatal traits and facilitates broader adoption of AI-driven phenotyping in crop science.

💡 Research Summary

The paper presents StomataSeg, a semi‑supervised instance segmentation framework specifically designed for the fine‑grained analysis of sorghum stomatal components. Sorghum (Sorghum bicolor) is a drought‑tolerant cereal of high importance for climate‑resilient agriculture, and precise characterization of stomatal traits (density, size, pore openness, spatial distribution) is essential for breeding programs aimed at improving water‑use efficiency. However, automated phenotyping of sorghum stomata is challenging because the structures are tiny (often < 40 µm), exhibit considerable shape variation across genotypes and leaf surfaces, and consist of nested sub‑structures (complex area, guard cells, pore). Existing deep‑learning approaches largely focus on semantic segmentation or bounding‑box detection and struggle with these nested, low‑contrast objects.

To address these gaps, the authors built a comprehensive dataset named StomataSeg. High‑resolution digital microscope images (2592 × 1944 px, JPEG) were captured from fully expanded leaves of five sorghum genotypes grown in a solar‑weave greenhouse. For each genotype, images were taken from both adaxial and abaxial surfaces and from three longitudinal leaf regions (base, middle, tip), yielding a biologically diverse set. After rigorous quality control, 318 original images were selected for annotation. Using the V7 cloud‑based platform, three trained annotators manually delineated three classes per stomatal instance: (1) stomatal complex area (including guard cells and pore), (2) guard cell area, and (3) pore area. The annotation protocol emphasized precise polygon masks, non‑overlapping instances, and inter‑annotator consensus, resulting in 11,060 human‑annotated patches (512 × 512 px, overlapping) and a review pass rate of 90 %.

Recognizing that dense mask annotation is a major bottleneck, the authors introduced a two‑stage semi‑supervised learning pipeline. First, a strong instance segmentation backbone (Mask R‑CNN) is trained on the human‑labeled patches. This model then generates high‑confidence pseudo‑labels on the remaining unlabeled images; only predictions exceeding a confidence threshold (e.g., 0.9) and meeting class‑balance criteria are retained. This process yields an additional 56,428 pseudo‑labeled patches, expanding the training set more than sixfold without extra manual effort. In the second stage, the model is re‑trained on the combined set (human + pseudo labels), allowing it to refine its representation of tiny structures.

The authors benchmarked both semantic segmentation models (U‑Net, DeepLabV3+) and instance segmentation models (Mask R‑CNN, Cascade Mask R‑CNN) under three conditions: (i) baseline with only human labels, (ii) baseline with patch‑based preprocessing only, and (iii) the full semi‑supervised pipeline. Performance was measured using mean Intersection‑over‑Union (mIoU) for semantic tasks and Average Precision (AP) for instance tasks. The semi‑supervised approach achieved a top mIoU increase from 65.93 % to 70.35 % (+4.42 pp) and a top AP increase from 28.30 % to 46.10 % (+17.80 pp). Notably, the AP for the pore class—historically the most difficult due to its minute size and low contrast—improved by over 20 pp, demonstrating the effectiveness of the patch‑based magnification and pseudo‑label expansion.

The paper discusses several strengths: (1) patch‑based preprocessing mitigates scale imbalance, allowing the network to focus on fine details; (2) semi‑supervised pseudo‑labeling dramatically reduces annotation cost while preserving or enhancing accuracy; (3) the dataset captures extensive biological variability (multiple genotypes, leaf surfaces, developmental stages), enhancing model generalizability. Limitations include potential error propagation from inaccurate pseudo‑labels, computational overhead from overlapping patches, and the current focus on static images rather than time‑series or in‑situ video data. Future work is suggested to explore multi‑scale pyramid networks, self‑supervised pretraining, and real‑time video analysis for dynamic stomatal monitoring.

In conclusion, StomataSeg delivers a high‑quality, multi‑class instance segmentation dataset and a practical semi‑supervised training pipeline that together enable reliable, high‑throughput extraction of sorghum stomatal traits. By overcoming the twin challenges of tiny nested structures and annotation bottlenecks, the framework paves the way for broader adoption of AI‑driven phenotyping in crop science and supports breeding efforts aimed at improving drought resilience and water‑use efficiency in sorghum and potentially other cereal crops.