Optimized ensemble learning frameworks for early grain yield prediction in winter wheat using static and multitemporal multispectral UAV data

A stacked ensemble learning framework was developed leveraging the large-scale spatial and temporal diversity in a winter wheat breeding program. Multispectral data were collected from 1098 plots comprising 586 genetically diverse advanced breeding lines at tillering, stem elongation, reproductive, early grain filling, mid grain filling, and late dough stages. Raw spectral information, 28 vegetation indices and 16 grey-level co-occurrence matrices were used– both individually and in combination –to train five machine learning algorithms– elastic net , support vector regression , k-nearest neighbours , random forest , and extreme gradient boosting . Static and temporal models were trained for spatial interpolation and genetic extrapolation, followed by internal and external validation. Multispectral features showed stage-specific discriminatory power for separating yield groups, peaking between the reproductive and mid grain filling stages. Classification was limited by overlap between medium- and high-yielding groups. Models integrating multispectral data across growth stages outperformed all other strategies. Elastic net performed best overall, while support vector regression showed the most consistent ranking. Base learners achieved average prediction accuracy of 55.85 % and 46.00 %, with AUC values of 0.73 and 0.66 for internal and external validation, respectively. Selected base-learners were used for stacked elastic net ensemble learning, which increased overall prediction (∼18 %) and classification (∼7 %) accuracies, while reducing errors (∼10 %). These improvements were primarily attributed to reduced bias. Overall, this study presents a scalable machine learning framework trained and validated on an operational breeding program for data-driven UAV-assisted early generation selection, providing a methodological reference for future research and deployment.