Challenges in Maize Root Phenotyping: Preprocessing Limits and Class Imbalance in Deep Learning

Doubled Haploid (DH) technology significantly accelerates the development of homozygous lines in maize breeding; however, its scalability is constrained by the reliable discrimination of haploid and diploid individuals. The widely used R1-nj anthocyanin marker at the seed stage is susceptible to genetic suppression and environmental variability, leading to high misclassification rates. This limitation has driven a shift toward seedling root morphology as a more robust phenotypic marker, yet it introduces major challenges, including complex image noise and severe class imbalance. In this study, we systematically evaluate the limitations of standard computer vision pipelines and baseline deep learning models for root-based classification. Automated background removal methods (HSV, Rembg) are shown to misinterpret fine root hairs as noise, resulting in significant morphological data loss. Additionally, experiments conducted under a realistic class imbalance (1:5.4) demonstrate that widely used CNN architectures (ResNet50, VGG16, EfficientNetB0, DenseNet121) exhibit strong majority class bias, with haploid recall dropping to 0.00% and 27.7%. These findings reveal a critical limitation in existing approaches and highlight the need for domain-informed datasets and imbalance-aware learning strategies for robust and scalable AI-based maize breeding systems.