High-throughput UAV phenotyping for plot-level harvest index estimation in wheat fields

Accurate estimation of the harvest index (HI), the ratio of grain yield to total aboveground biomass (AGB), is crucial for evaluating crop productivity and resource-use efficiency in wheat breeding programs. While traditional HI measurement methods use destructive field sampling, which is labour-intensive and impractical for large-scale breeding trials, recent advances in UAV-based remote sensing now offer non-destructive alternatives capable of delivering high-throughput, plot-level HI estimation. In this study, we present a high-throughput phenotyping framework that combines UAV-based multispectral imaging and ensemble machine learning to estimate HI under field environments. Multispectral data were collected at two key growth stages, anthesis and maturity, using a DJI M300 RTK drone equipped with a RedEdge-P sensor. Vegetation indices (VIs), including the normalized difference vegetation index (NDVI), normalized difference red edge index (NDRE), and green NDVI (G-NDVI), were extracted using data from sensors and ground truth monitoring and used as predictors to estimate grain yield and AGB for calculating HI. An ensemble learning model, based on a stacking architecture comprising five regressors and a ridge regression meta-learner, was employed to enhance prediction accuracy. Results showed strong correlations between UAV-derived and ground-truth VIs (R2 > 0.94, RMSE < 0.023). The ensemble model demonstrated high accuracy and strong generalization for HI estimation across both experimental sites and growing seasons. At the anthesis stage, the NDVI-based ensemble model achieved the best performance. For the Indian Head site, it yielded a testing R2 of 0.87, RMSE of 4.18 g/p, and NRMSE of 2.73 %, based on a training R2 of 0.83. At the Swift Current site, the model produced a testing R2 of 0.84, RMSE of 8.67 g/p, and NRMSE of 5.67 %. Similarly, at the maturity stage, the NDRE-based ensemble model was the top performer. It recorded a testing R2 of 0.86, RMSE of 7.10 g/p, and NRMSE of 4.64 % at Indian Head, and a testing R2 of 0.83 with an RMSE of 8.06 g/p, and NRMSE of 5.27 % at Swift Current. Across all indices and stages, the ensemble model consistently outperformed individual models, achieving high testing R2 values and low RMSE, which confirms its robustness and predictive power on unseen data. The proposed UAV machine learning framework demonstrates a reliable and non-destructive approach for field-level HI estimation, thereby improving germplasm selection efficiency for yield improvement. It offers a valuable tool for accelerating trait-based wheat breeding and precision agriculture applications.