Maize-IRNet: a deep-learning “field scorer” for tassel anthers and ear seed set

Instead of relying on time-consuming and subjective visual scoring, the team designed Maize-IRNet, a deep-learning model that grades tassel anther emergence and ear seed setting directly from images. Integrated into a smartphone application, the system enables rapid, standardized, and reproducible field assessments of haploid fertility.

In conventional breeding, developing inbred lines typically requires 6–8 generations of selfing, whereas double haploid (DH) breeding can shorten this process to just 2–3 generations. Despite this advantage, large-scale DH programs still rely on accurate evaluation of haploid genome doubling and spontaneous haploid genome doubling (SHGD), which vary widely across genetic backgrounds and environments. Many SHGD-related quantitative trait loci have been identified, but their genetic bases remain poorly resolved, partly due to inconsistent phenotyping. Faster and more standardized assessment of tassel anther emergence and ear seed setting is therefore critical for improving genetic mapping accuracy, optimizing doubling strategies, and strengthening DH breeding pipelines.

A study (DOI: 10.1016/j.plaphe.2025.100140) published in Plant Phenomics on 7 November 2025 by Jianxiao Liu’s & Yanzhi Qu’s team, Huazhong Agricultural University, reduces labor demands, minimizes observer bias, and supports high-throughput phenotyping needed for efficient DH breeding and genetic analysis.

Using a deep-learning evaluation framework, the researchers first assessed the grading performance of Maize-IRNet for haploid ear seed setting and tassel anther emergence by randomly splitting the dataset into training (90%) and testing (10%) sets and benchmarking it against six representative architectures (VGG11_bn, ResNet50, ResNet101, ViT-Base-16, gMLP, and MLP-Mixer). To further examine robustness and generalization, they conducted systematic data augmentation (Gaussian noise, brightness adjustment, and blur), extensive ablation experiments to isolate the contributions of Inception–ResNet modules, Reduction modules, and attention mechanisms, comparisons of different parameter initialization strategies, tests on balanced versus imbalanced datasets, multiple data partitioning strategies (ten-fold cross-validation and independent test sets), resolution-sensitivity analyses, Grad-CAM–based interpretability evaluations, computational efficiency comparisons, and finally deployed the optimized model in a cross-platform mobile application. Correspondingly, the results showed that Maize-IRNet consistently delivered the best overall performance, achieving 84.2% accuracy for ear seed setting and 84.0% accuracy for tassel anther emergence, clearly outperforming classical CNNs and newer transformer-style models. Confusion matrix analyses demonstrated that Maize-IRNet was particularly effective at discriminating higher fertility grades and reducing misclassification in intermediate levels. Under data augmentation, Maize-IRNet maintained top accuracy, indicating strong resistance to noise and environmental interference. Ablation studies revealed that removing Inception–ResNet modules or replacing Reduction modules caused notable performance drops, while integrating the proposed Global Attention Mechanism (GAM) improved accuracy from 82.1% to 84.2%, outperforming CBAM, SE, and EMA attention alternatives. He initialization proved optimal among four initialization strategies. Performance remained stable after class balancing and across different data partitioning schemes, confirming strong robustness. Importantly, reducing image resolution from 2048×2048 to 299×299 preserved accuracy while drastically lowering computational cost, enabling efficient deployment. Grad-CAM visualizations further showed that Maize-IRNet focused more precisely on biologically relevant regions such as seed-dense ear areas and exposed anthers. Finally, the model achieved a favorable balance between parameter size (137.15 MB) and inference speed (0.00663 s), supporting its successful integration into the Haploid-Fertility mobile application for real-time field phenotyping.

In conclusion, this study demonstrates that integrating the Maize-IRNet model into the Haploid-Fertility mobile application enables rapid, standardized, and high-throughput evaluation of maize haploid fertility restoration. By improving the accuracy and consistency of phenotyping, the approach supports more reliable SHGD genetic analyses, optimizes genome doubling strategies, and reduces production costs, offering a practical and scalable solution for accelerating double haploid breeding programs.

###

References

DOI

10.1016/j.plaphe.2025.100140

Original Source URl

https://doi.org/10.1016/j.plaphe.2025.100140

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.