Accurate 3D maize ear phenotyping using voxel grids derived from RGB machine vision

Maize ear phenotyping is receiving increasing attention, driven by the demand of plant scientists for large-scale trait data and enabled by the rapid development of AI-based machine vision. While manual measurement of ear traits is labour-intensive and often destructive, existing digital solutions typically rely on a limited number of 2D images without capturing the entire 3D structure of an ear. Therefore, a novel 3D ear scanning pipeline was developed that computes voxel grids based on RGB images automatically captured from 100 different viewing angles. For kernel segmentation, open-source deep learning models for object detection and segmentation were evaluated and combined, achieving a dice score of 0.94 on 2D ear images. A novel strategy for mapping 2D segmentation masks to 3D-oriented voxels resulted in high coefficients of determination (R ² ) for predicting kernel number (R ²  = 0.98), kernel row number (R ²  = 0.77) and maximum kernel row length (R ²  = 0.87). Similarly, ear width (R ²  = 0.82), ear length (R ²  = 0.96) and ear volume (R ²  = 0.97) were accurately predicted when compared to manual reference measurements. The applicability of the pipeline was demonstrated on ears grown under two levels of phosphorus supply, which revealed significant effects on ear width, ear length, ear volume and kernel size. Data acquisition required 30 s per ear, while computation of all traits took an average of 47 s. This study establishes an accurate and fast 3D phenotyping pipeline while bridging high-throughput ear phenotyping with plant nutrition research, demonstrating its potential to link structural traits with nutrient-driven responses in crops.