Forward prediction of grain yield across early and advanced generations in winter wheat ( Triticum aestivum L.) using NIRS‐based models

Near‐infrared spectroscopy (NIRS) is a rapid, non‐destructive, and low‐cost phenotyping tool with strong potential to accelerate plant breeding. We evaluated NIRS‐based models for forward prediction of grain yield in winter wheat ( Triticum aestivum L.), using historical phenotypic and spectral data across multiple years and breeding cycles. Two forward prediction scenarios were tested: a fixed‐size training set and a cumulative training set spanning up to 5 years. NIRS‐based predictions showed moderate to high predictive ability in F 5 and advanced lines across all models and scenarios (0.28–0.58). Machine learning methods such as support vector machine and random forest performed competitively, while ridge regression was stable and effective. For F 4 generation lines, grown in unreplicated small plots without direct grain yield measurements, we predicted grain yield from their NIRS profile. Training sets consisted of F 5 –advanced lines (e.g., 2015–2017) to predict the performance of F 4 nursery lines without direct grain yield measurements (e.g., 2018). The predictions were subsequently validated with the F 4 ‐derived F 5 generation lines, field‐tested in multi‐location trials in the subsequent year (e.g., 2019), where still moderate predictive abilities were observed (0.22–0.31). Therefore, NIRS‐based predictions can be practically applied to identify promising genotypes for subsequent selection stages (such as multi‐location yield trials) at early breeding phases, even in the absence of grain yield data. This highlights the potential of this approach to support more efficient and data‐informed selection decisions, particularly in breeding programs with limited genotyping resources.