Predicting grain yield in a novel soft winter wheat (Triticum aestivum L.) population using spike traits and machine learning approaches

Accurate yield prediction in soft winter wheat breeding requires precise quantification of spike architecture traits, which are key determinants of grain production. The study aimed to predict wheat grain yield using machine learning, taking spike attributes as input variables. The study was conducted on a novel hybrid population of winter wheat comprising 6,999 spikes, cultivated at Kaluga, Moscow, Russia, during the 2021-22 study period. Spike morphology traits (length, weight, spikelet counts) and derived metrics such as density distributions, fertility %, attributive value, and grain yield coefficient were employed with machine learning models: partial least squares (PLS), random forest (RF), and gradient boosting (GB). Python 3.11.7 was used for modelling, and the dataset was randomly split into training (70%) and testing (30%) data, fitted, trained, and predictions were made. Performance of models was evaluated using mean squared error (MSE), mean absolute error (MAE), root MSE (RMSE), R2, and ratio of performance to deviation (RPD). Grain weight was strongly correlated with density-based traits (spike weight (r = 0.594), grain number (r = 0.353), grain weight density (r = 0.617) and physiological traits such as fertility percentage (r = 0.523) and attributive value (r = 0.674). In predictive modelling, the PLS formed the best with RMSE = 0.1110, R² = 0.8957, followed by RF and GB models, respectively. RF prioritised structural traits such as spike length and spikelet number, whereas the GB model emphasised physiological traits like attributive value. The findings highlight the importance of integrating both structural and physiological trait optimisation in wheat breeding. © 2025 Elsevier B.V., All rights reserved.