Work

PointBALNet: A Deep Learning Framework for Soybean 3D Point Cloud Branching Point Segmentation Combining Label-Balanced Sampling and Asymmetric Loss

Year: 2026
Type: article
Abstract: Point-wise segmentation of soybean branching regions is challenged by class imbalance and complex local geometry. We present PointBALNet, a PointNet++-based framework combining label-balanced training inputs, a composite asymmetric loss, and multi-scale spatial–channel attention. The dataset comprised point clouds of 168 soybean plants after manual leaf and pod removal, with 118, 16, and 34 plants assigned to training, validation, and testing, respectively. Under a label-aware, class-rebalanced test protocol with 2048 records per plant, the reported full-model scores were 98.815% accuracy, 97.961% macro-averaged recall, 98.143% macro-averaged F1, and 96.411% plant-averaged mIoU. Matched-input comparisons across three paired training seeds did not detect statistically significant mIoU differences between target training-input positive ratios of 0.8 and 0.9. Shared-input comparisons showed uneven changes across module combinations, while comparisons of the full model with other configurations also differed in test-input composition. These results describe segmentation of the prepared individual-plant point clouds. Future work will prioritize field experiments on naturally growing plants and validation of downstream phenotyping applications.
Source: Agriculture
Language: English
Subfield: Plant Science
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7236563595
FWCI:
Cites:
Cited by:
Related to: 10