Sequential Soybean Phenotypes for Grain Yield Prediction in Reciprocal Two-Season Validation

Whether sequential crop phenotypes improve grain yield prediction after transfer to another growing season remains poorly quantified. We evaluated soybean (Glycine max (L.) Merr.) at one experimental site over two seasons in a split-plot experiment with five sowing periods, two cultivars, four seeding rates, and four blocks (160 subplots per season). Information gates were cumulative predictor sets available through V4, R2, R5, and R6. They combined subplot means of leaf area, total dry matter, branching, main-stem nodes, and plant height from distinct destructive samples. Grain yield at R8 was the outcome. Elastic-net models estimated phenotypic corrections to development-season means for matching management combinations. Reciprocal whole-season evaluation used development-only grouped tuning and conditional within-season whole-block bootstrap intervals. The management comparator produced root mean squared errors (RMSEs) of 1398 and 1394 kg ha−1 and predictive R2 values of 0.282 and −0.182 for 2021/22 → 2022/23 and the reciprocal direction, respectively. Paired RMSE gains from cumulative phenotypes ranged from −0.419% to 0.038% and from −0.712% to −0.104%, respectively, where positive values denote lower independent-season error. Alternative management comparators yielded similarly small changes. Predictor distributions and calibration differed between transfer directions. These evaluations identify little incremental benefit from the measured phenotypes under the tested conditions, while demonstrating the importance of distinguishing yield ranking, calibration, and additional predictive information. Broader environmental validation is needed to determine when phenotyping improves forecasting beyond management expectations.