From seed centres to tray cells: a geometry-aware machine vision framework for full-tray sowing quality assessment in maize plug trays
- 1Faculty of Software Technologies, Shanxi Agricultural University
Abstract
Accurate sowing-quality assessment of maize seedling trays requires both reliable seed localisation and correct correspondence between detected seeds and physical tray cells. This study presents a semi-automated machine-vision framework combining heatmap-based seed-centre localisation, sliding-window inference, global non-maximum suppression (NMS), and four-corner-calibrated adaptive grid assignment. Each assessable cell was classified as empty, single-seeded, or multi-seeded according to the number of assigned seed centres. On source-image-independent subsets, the framework achieved an accuracy of 0.9821 and a macro F1-score of 0.9040 for 27 complete-tray images (3,456 cells), and 0.9798 and 0.9344, respectively, for 9 visible-cell images (1,140 assessable cells). Separately, the larger evaluation sets yielded accuracy/macro F1 values of 0.9783/0.9145 for 44 complete trays and 0.9782/0.9368 for 30 visible-cell images. These results support cell-level sowing-quality assessment under the evaluated acquisition conditions and within manually assessable regions. The source-image-independent subsets control direct source overlap but remain within the same acquisition domain; external generalisation across production sites, imaging devices, tray specifications, and acquisition conditions has not yet been established.
Related
- Application of spectroscopic and multispectral imaging technologies on the assessment of ready-to-eat pineapple quality: A performance evaluation study of machine learning models generated from two commercial data analytics tools
- Ensemble-based support vector machine classifiers as an efficient tool for quality assessment of beef fillets from electronic nose data
- Advanced machine learning framework for hole quality prediction and optimization in polycarbonate AWJD
- MMM: A deterministic Mendelian mismatch matrix framework for germplasm collection quality assessment, demonstrated in apple and sweet cherry
- Empirical Assessment of Non-Invasive Population Genetics in Bats: Comparison of DNA Quality from Faecal and Tissue Samples
BibTeX
@article{zan2026from,
title = {{From seed centres to tray cells: a geometry-aware machine vision framework for full-tray sowing quality assessment in maize plug trays}},
author = {Jiangchen Zan and Jinfeng Peng and Ruijun Gao and Xiangyu Ma and Bowen Xue and Xiaoying Zhang},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1981493},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1981493/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - From seed centres to tray cells: a geometry-aware machine vision framework for full-tray sowing quality assessment in maize plug trays
AU - Jiangchen Zan
AU - Jinfeng Peng
AU - Ruijun Gao
AU - Xiangyu Ma
AU - Bowen Xue
AU - Xiaoying Zhang
PY - 2026
DA - 2026/10/23
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1981493
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1981493/full
AB - Accurate sowing-quality assessment of maize seedling trays requires both reliable seed localisation and correct correspondence between detected seeds and physical tray cells. This study presents a semi-automated machine-vision framework combining heatmap-based seed-centre localisation, sliding-window inference, global non-maximum suppression (NMS), and four-corner-calibrated adaptive grid assignment. Each assessable cell was classified as empty, single-seeded, or multi-seeded according to the number of assigned seed centres. On source-image-independent subsets, the framework achieved an accuracy of 0.9821 and a macro F1-score of 0.9040 for 27 complete-tray images (3,456 cells), and 0.9798 and 0.9344, respectively, for 9 visible-cell images (1,140 assessable cells). Separately, the larger evaluation sets yielded accuracy/macro F1 values of 0.9783/0.9145 for 44 complete trays and 0.9782/0.9368 for 30 visible-cell images. These results support cell-level sowing-quality assessment under the evaluated acquisition conditions and within manually assessable regions. The source-image-independent subsets control direct source overlap but remain within the same acquisition domain; external generalisation across production sites, imaging devices, tray specifications, and acquisition conditions has not yet been established.
SN - 1664-462X
ER -