Winter wheat yield prediction using temporal fusion of multispectral imagery and LiDAR point clouds acquired from a ground-based phenotyping platform
- 1Agricultural Information Institute, Chinese Academy of Agricultural Sciences
- 2Institute of Agricultural Information and Economics, Hebei Academy of Agriculture and Forestry Sciences
Abstract
Winter wheat yield reflects the combined development of canopy structure and physiological status, making multitemporal structural and spectral observations potentially valuable for pre-maturity prediction. We used a ground-based phenotyping platform to acquire five-band multispectral imagery at 14 growth stages and LiDAR point clouds at 10 growth stages. Both data streams were aligned to a common phenological axis and integrated in a temporal graph convolutional network (Temporal-GCN) that encoded canopy spectral responses, three-dimensional structural states, and signed changes between adjacent stages. A stagewise cutoff analysis evaluated yield prediction from erecting through maturity. Multitemporal multispectral trajectories captured SPAD dynamics, whereas LiDAR directly provided plant height and achieved high accuracy in LAI estimation, indicating complementary sensing responses. In three-fold cross-validation grouped by complete biological replicate, the fused model comprising VARI, EVI, NDRE, and seven point-cloud structural variables reached R ² = 0.883 at heading and its highest full-season cutoff point estimate at early grain filling (R ² = 0.903, RMSE = 25.52 kg mu ⁻¹ , MAE = 20.84 kg mu ⁻¹ ). The corresponding current-stage-only model yielded R ² = 0.760. Paired bootstrap analysis indicated a cumulative-versus-current-stage improvement of ΔR ² = 0.143 (95% CI, 0.059–0.299) and an RMSE reduction of 14.69 kg mu ⁻¹ (95% CI, 7.28–22.02 kg mu ⁻¹ ). Performance remained similar at mid grain filling (R ² = 0.900), and the early-versus-mid difference was small (ΔR ² = 0.003; 95% CI, −0.015 to 0.025). A nested grouped sensitivity analysis for endpoint-pooling and regularization yielded R ² = 0.896. Early grain filling occurred approximately 27 d before harvest. These results show that multitemporal structural–spectral information can improve pre-maturity winter wheat yield prediction, while the identified prediction window should be interpreted within the present one-site, one-season experiment.
Related
- Multi-temporal UAV multispectral phenotyping and machine learning for modeling cotton defoliation dynamics across diverse accessions
- Rice multi-variety yield prediction using multi-temporal fusion of thermal infrared and multispectral UAV data based on random Forest feature selection
- Multispectral UAV dataset for multi-temporal mangrove monitoring in the Cirebon Coastal Area, Indonesia (2025-2026).
- Democratizing three-dimensional surface phenotyping: an open structured-light platform reveals and removes the projection bias in biological imaging
- Aerial Phenotyping of Coffee Flowering Using Deep Learning for Early Yield Prediction
BibTeX
@article{li2026winter,
title = {{Winter wheat yield prediction using temporal fusion of multispectral imagery and LiDAR point clouds acquired from a ground-based phenotyping platform}},
author = {Haodong Li and Yeping Zhu and Shengping Liu and Fenghuan Du and Yu Zhang and Yihan Jiang and Run Li and Cunchao Li and Shijuan Li and Huili Yue},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1953299},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1953299/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Winter wheat yield prediction using temporal fusion of multispectral imagery and LiDAR point clouds acquired from a ground-based phenotyping platform
AU - Haodong Li
AU - Yeping Zhu
AU - Shengping Liu
AU - Fenghuan Du
AU - Yu Zhang
AU - Yihan Jiang
AU - Run Li
AU - Cunchao Li
AU - Shijuan Li
AU - Huili Yue
PY - 2026
DA - 2026/10/14
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1953299
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1953299/full
AB - Winter wheat yield reflects the combined development of canopy structure and physiological status, making multitemporal structural and spectral observations potentially valuable for pre-maturity prediction. We used a ground-based phenotyping platform to acquire five-band multispectral imagery at 14 growth stages and LiDAR point clouds at 10 growth stages. Both data streams were aligned to a common phenological axis and integrated in a temporal graph convolutional network (Temporal-GCN) that encoded canopy spectral responses, three-dimensional structural states, and signed changes between adjacent stages. A stagewise cutoff analysis evaluated yield prediction from erecting through maturity. Multitemporal multispectral trajectories captured SPAD dynamics, whereas LiDAR directly provided plant height and achieved high accuracy in LAI estimation, indicating complementary sensing responses. In three-fold cross-validation grouped by complete biological replicate, the fused model comprising VARI, EVI, NDRE, and seven point-cloud structural variables reached R ² = 0.883 at heading and its highest full-season cutoff point estimate at early grain filling (R ² = 0.903, RMSE = 25.52 kg mu ⁻¹ , MAE = 20.84 kg mu ⁻¹ ). The corresponding current-stage-only model yielded R ² = 0.760. Paired bootstrap analysis indicated a cumulative-versus-current-stage improvement of ΔR ² = 0.143 (95% CI, 0.059–0.299) and an RMSE reduction of 14.69 kg mu ⁻¹ (95% CI, 7.28–22.02 kg mu ⁻¹ ). Performance remained similar at mid grain filling (R ² = 0.900), and the early-versus-mid difference was small (ΔR ² = 0.003; 95% CI, −0.015 to 0.025). A nested grouped sensitivity analysis for endpoint-pooling and regularization yielded R ² = 0.896. Early grain filling occurred approximately 27 d before harvest. These results show that multitemporal structural–spectral information can improve pre-maturity winter wheat yield prediction, while the identified prediction window should be interpreted within the present one-site, one-season experiment.
SN - 1664-462X
ER -