Label-efficient structure-aware learning framework for wheat tiller density estimation from UAV remote sensing
Computers and Electronics in Agriculture · 2026-11 · 254 · 112284
DOI: 10.1016/j.compag.2026.112284 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computers and Electronics in Agriculture, 2026-11. Verfasst von Mengqing Zhang und 7 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Tiller density is an important agronomic trait related to wheat growth diagnosis and yield formation. UAV remote sensing enables field-scale monitoring, but reliable estimation is challenged by changing canopy appearance, soil and shadow interference, and costly ground-label collection. We propose a label-efficient, structure-aware framework for wheat tiller-density estimation from UAV RGB and multispectral imagery. The framework integrates 3B-SSCL for self-supervised crop-background segmentation, MVDINO for RGB canopy-context representation learning, and GIDVF for guided fusion with multispectral vegetation indices. Using 377 source-field samples from Field 1 collected across four dates, the model achieved R 2 = 0.8811 ± 0.0249 under repeated random validation and R 2 = 0.8753 ± 0.0141 under leave-location-out validation. Under cross-field adaptation to an independent Field 2, it achieved R 2 = 0.8145 ± 0.0959 with 50 target-field labels under a locked target-test protocol. At this budget, its mean R 2 exceeded the corresponding values of the evaluated baseline methods when they used the complete 70-label adaptation pool. Across the tested budgets, the proposed framework maintained this advantage, with its learning curve approaching a stable performance level at approximately 50 labels. Together, these results show that the framework can achieve effective cross-field transfer with a practical target-field label budget while reducing reliance on extensive new labels.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Zhang, M., Zhang, Y., Gao, T., Tong, F., Ye, Z., Cen, L., Zhang, M., & Li, M. (2026). Label-efficient structure-aware learning framework for wheat tiller density estimation from UAV remote sensing. Computers and Electronics in Agriculture, 254, 112284. https://doi.org/10.1016/j.compag.2026.112284
Vancouver
Zhang M, Zhang Y, Gao T, Tong F, Ye Z, Cen L, et al. Label-efficient structure-aware learning framework for wheat tiller density estimation from UAV remote sensing. Computers and Electronics in Agriculture. 2026;254:112284. doi:10.1016/j.compag.2026.112284
BibTeX
@article{Zhang2026,
author = {Zhang, Mengqing and Zhang, Yao and Gao, Tingyao and Tong, Fanghui and Ye, Ziqing and Cen, Lunyu and Zhang, Man and Li, Minzan},
title = {Label-efficient structure-aware learning framework for wheat tiller density estimation from UAV remote sensing},
journal = {Computers and Electronics in Agriculture},
year = {2026},
volume = {254},
pages = {112284},
doi = {10.1016/j.compag.2026.112284}
}
EUVIMED