Handling intra-class imbalance in part-segmentation of different wheat cultivars

Plant phenotyping is crucial for precisely measuring traits across diverse morphotypes in cereals such as wheat, where the variability in the proportion of ears to leaves poses significant challenges due to class imbalances and the granularity limitations of 3D point cloud data. Our study addresses these intra-class imbalances through two strategies using the PointNet++ network: use plant features i.e., ear ratio and ear count for weighted point cloud sampling and apply class weights in the loss function. We analyze datasets from three morphologically distinct wheat varieties: Paragon, Gladius, and Apogee. Introducing point cloud weights based on ear ratio and ear count significantly improves the network's ability to recognize underrepresented parts in datasets with uneven distributions of ear and non-ear points. We observe a differential impact in all the categories in terms of segmentation accuracy and average mIoU. In comparison to the base method, all the wheat categories display enhancement in performance after applying both techniques. The best results are obtained with point cloud weights in the loss function for the Gladius dataset, showing a substantial improvement of 10% to 12% in average ear mIoU compared to base method (0.483). Although a similar trend is observed with weighted sampling, the enhanced results in the Gladius dataset range from 0.611 to 0.626 indicate model's enhanced capability to identify different segments or parts precisely. This research demonstrates the efficacy of our methods in addressing class imbalance issues in point cloud segmentation and provides enhanced accuracy for particularly across different wheat genotypes. In conclusion, the techniques in this study can reliably handle intra class imbalance in diverse datasets, which offers a scalable solution to improve agricultural phenotyping.