Low-annotation apple flower counting gets a semi-supervised boost
en-GBde-DEes-ESfr-FR

Low-annotation apple flower counting gets a semi-supervised boost

16/11/2026 TranSpread

Deep learning has made automated flower counting more efficient, but high accuracy often depends on large, costly annotated datasets. In orchards, annotation is scarce, and unmanned aerial vehicle (UAV) images add complex backgrounds, tiny targets, dense occlusion, and large scale variation. Semi-supervised learning (SSL) can reduce labeling needs, yet conventional pseudo-labeling is vulnerable to noisy agricultural scenes. Density estimation helps handle dense and overlapping targets, but existing preprocessing or multi-task strategies may increase model complexity or rely on handcrafted color features with limited robustness. Based on these challenges, there is a need for a low-annotation framework that purifies inputs, suppresses unreliable pseudo-labels, and adapts to stage-specific flower morphology.

Researchers at Shandong Agricultural University in China published (DOI: 10.1016/j.plaph.2026.100190) the study in Plant Phenomics (2026). The work presents combines a training-free color-SAM flower extractor (CSAM-FE), a training-free preprocessing module that combines color thresholding with the Segment Anything Model (SAM), and USCount-Net, an uncertainty-guided semi-supervised network for counting bud, anthesis, and petal-fall apple flowers from UAV imagery. The framework was tested on a self-built apple flower dataset and two public datasets. It reports lower mean absolute error (MAE) and root mean square error (RMSE) than several state-of-the-art supervised and semi-supervised counting models.

The proposed pipeline first applies CSAM-FE to suppress orchard background noise. A strict red-green-blue (RGB) color threshold roughly isolates flower pixels, and the centroids of connected regions become point prompts for SAM. SAM then refines flower-cluster masks, which are cropped using minimum bounding rectangles to preserve local context. This step reduces processed pixels by at least 80% and by 98.76% on average, yielding 9,042 flower-cluster images. USCount-Net then uses a mean-teacher architecture with a modified VGG-19 backbone, an adaptive gated fusion (AGF) module, a convolutional block attention module (CBAM), and a density map regressor. The AGF module uses spatial gating to fuse multi-scale features while suppressing redundant or noisy information. For unlabeled data, frequent forward uncertainty estimation (FFUE) performs multiple forward passes and measures prediction variance. A pseudo-label filtering (PLF) mechanism uses channel-specific adaptive thresholds to mask high-uncertainty pixels in the consistency loss. On the self-built dataset, USCount-Net achieved lower MAE and RMSE than state-of-the-art methods at 10%, 30%, and 50% labeling ratios; at 10%, average MAE and RMSE dropped by 3.93 and 5.64, respectively, compared with the best baseline. At 30% labeling, performance approached supervised levels.

The authors said the work targets a practical orchard bottleneck: bud and anthesis counts are most useful for thinning decisions, but petal-fall flowers are small and easily confused with background. They said CSAM-FE removes much of the background before counting, while the uncertainty-guided filter prevents noisy pseudo-labels from dominating training. They added that the framework is not intended to replace field scouting, but to convert UAV imagery into canopy-level flower density maps that can make orchard assessment faster and more consistent.

The framework could support intelligent orchard systems by providing canopy-level, stage-specific flower density maps for chemical or mechanical thinning, yield prediction, and bloom monitoring. Its low-annotation design may also lower the cost of deploying deep learning in specialty crops. However, the color-threshold prompts in CSAM-FE remain sensitive to lighting and imaging conditions, and the method was validated mainly in flat orchards with fixed UAV settings. Future work may add adaptive thresholding, multi-spectral or thermal imagery, and more diverse field data. The authors also point to lightweight deployment and cross-species generalization as next steps.

###

References

DOI

10.1016/j.plaphe.2026.100190

Original Source URL

https://doi.org/10.1016/j.plaphe.2026.100190

Funding information

This work was supported by the National Natural Science Foundation of China under grant 32401702, Shandong Provincial Natural Science Foundation of China under grant ZR2023QF016, Shandong Provincial Natural Science Foundation of China under grant ZR2025QC206, and Program for Scientific Research Innovation Team of Young Scholar in Colleges and Universities of Shandong Province under grant 2024KJH091. Thanks to all the help of the teachers and students.

About Plant Phenomics

Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting‑edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real‑world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high‑throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.

Paper title: Low-annotation apple flower counting: A color-SAM enhanced and uncertainty-guided semi-supervised framework
Fichiers joints
  • Apple flower counting dataset construction pipeline.
16/11/2026 TranSpread
Regions: North America, United States, Asia, China
Keywords: Applied science, Engineering, Technology

Disclaimer: AlphaGalileo is not responsible for the accuracy of content posted to AlphaGalileo by contributing institutions or for the use of any information through the AlphaGalileo system.

Témoignages

We have used AlphaGalileo since its foundation but frankly we need it more than ever now to ensure our research news is heard across Europe, Asia and North America. As one of the UK’s leading research universities we want to continue to work with other outstanding researchers in Europe. AlphaGalileo helps us to continue to bring our research story to them and the rest of the world.
Peter Dunn, Director of Press and Media Relations at the University of Warwick
AlphaGalileo has helped us more than double our reach at SciDev.Net. The service has enabled our journalists around the world to reach the mainstream media with articles about the impact of science on people in low- and middle-income countries, leading to big increases in the number of SciDev.Net articles that have been republished.
Ben Deighton, SciDevNet
AlphaGalileo is a great source of global research news. I use it regularly.
Robert Lee Hotz, LA Times

Nous travaillons en étroite collaboration avec...


  • The Research Council of Norway
  • SciDevNet
  • Swiss National Science Foundation
  • iesResearch
Copyright 2026 by DNN Corp Terms Of Use Privacy Statement