Early Discrimination of Maternal Haploid and Diploid Maize ( Zea mays L.) Seedlings Using Morphological Traits and Random Forest Classifier

ABSTRACT Doubled haploid (DH) technology has been widely adopted in maize ( Zea mays L.) breeding programs due to its ability to reduce breeding cycle time and optimize costs. Early ploidy identification is essential for maximizing the efficiency of DH production, particularly prior to chromosome doubling. This study aimed to evaluate the use of maize seedling traits as a strategy for discriminating haploids and diploids derived from in vivo haploid induction. Three commercial maize hybrids were crossed with 10 haploid inducers, and seeds were initially classified using the R1‐navajo ( R1‐nj ) marker. Seedlings were germinated on paper rolls and, after 96 h under controlled conditions, were evaluated for coleoptile length (CL), radicle length (RL), coleoptile diameter (CD), radicle diameter (RD) and number of lateral seminal roots (NLSR). Subsequently, seedlings were transplanted to the field, and true ploidy was determined at flowering based on gold standard phenotypic evaluation of vigour and tassel fertility. All seedling traits presented significant effects of hybrid and ploidy level, with diploid plants exhibiting greater values than haploids for the morphological traits. A random forest classifier integrating all seedling traits achieved a high area under the ROC curve (AUC) (0.892), accuracy (0.812), sensitivity (0.788) and specificity (0.858) for the overall model, englobing all three hybrids in a repeated stratified cross‐validation scheme. The error rates for the overall model were 8.4% for false discovery rate (FDR) and 21.2% for false negative rate (FNR), showing that the model was more effective at identifying true diploid seedlings. These results demonstrate that seedling phenotyping, when used integrated with machine learning, provides a complementary tool for early ploidy identification, improving efficiency of resource use in maize DH breeding programs.