Drones and deep learning boost soybean breeding with 3D canopy insights

By integrating spatial structure with spectral data, the approach significantly improves yield estimation accuracy and lodging detection, offering a powerful tool for accelerating the development of high-yield and resilient soybean varieties.

Soybean is a vital food, feed, and industrial crop, and its global demand continues to grow. Efficient breeding requires large-scale evaluation of hundreds of soybean lines, yet traditional field methods struggle to capture complex plant traits accurately or in real time. Remote sensing technologies, including UAVs with multispectral and LiDAR sensors, have made phenotyping more efficient. However, most research relies on 2D vegetation indices and texture-based features, which cannot fully capture canopy structure. LiDAR offers detailed 3D information but remains costly and impractical for large-scale use. Photogrammetry with UAVs provides a more economical solution, though conventional flight paths face limitations in complex environments. These challenges highlight the need for high-precision, cost-effective, and structure-aware phenotyping methods that can directly support soybean breeding decisions.

A study (DOI: 10.1016/j.plaphe.2025.100028) published in Plant Phenomics on 20 March 2025 by Yuntao Ma’s team, China Agricultural University, enhances the accuracy and efficiency of soybean phenotypic trait estimation, providing powerful tools for accelerating crop breeding.

In this study, researchers employed a lightweight UAV with cross-circling oblique (CCO) flight paths and Structure-from-Motion Multi-View Stereo (SfM-MVS) techniques to generate dense, high-quality 3D point clouds of soybean canopies. Building on these data, they developed two novel deep learning models—SoyNet and SoyNet-Res—that integrated spatial structure with RGB and vegetation index (VI) information to perform multi-task phenotype analysis, including yield estimation and lodging discrimination. Systematic experiments were conducted to optimize the strategy for soybean phenotypic estimation. First, the number of input point clouds was tested using six groups ranging from 256 to 4096 points. Accuracy improved as the number increased, reaching optimal performance at 3072 points with an RMSE of 375.37 kg ha−1 and correlation coefficient of 0.60 for yield estimation, along with a peak lodging accuracy of 0.56 in five-class discrimination. Beyond 4096 points, memory burdens reduced efficiency. Next, feature fusion strategies were examined, comparing spatial-only inputs with spatial+RGB and spatial+RGB+VI inputs. The latter achieved the best performance, lowering RMSE to 352.82 kg ha−1, raising r to 0.66, and reaching a top-1 lodging accuracy of 0.62. The models were further compared with conventional methods such as H2O-AutoML and PointNet++. Point cloud deep learning consistently outperformed traditional approaches, reducing RMSE by 6.58–10.38 kg ha−1 and enhancing lodging discrimination accuracy by 0.04–0.07. SoyNet-Res showed superior stability across repeated trials, with mean pooling favoring yield prediction (RMSE 305.71 kg ha−1) and max pooling boosting lodging detection (top-1 accuracy 0.59). Multi-task deep learning (MDL) surpassed single-task methods by exploiting correlations between traits, with dynamic weighting yielding the strongest results: yield estimation RMSE of 349.45 kg ha−1 and lodging classification accuracies of 0.87 (top-2) and 0.97 (top-3). Growth stage analysis identified S7 as the optimal period for trait estimation, while spatial error analysis confirmed the robustness of the models.

This integration of UAV photogrammetry and deep learning opens new possibilities for accelerating soybean breeding. By providing reliable, large-scale phenotypic data, the models allow breeders to more quickly identify high-yielding and lodging-resistant lines, reducing costs and improving selection efficiency. The fusion of spectral and structural data at the raw level also provides a methodological breakthrough, moving beyond traditional statistical indices toward more holistic trait assessment.

###

References

DOI

10.1016/j.plaphe.2025.100028

Original URL

https://doi.org/10.1016/j.plaphe.2025.100028

Funding information

This research was funded by the National Center of Pratacultural Technology Innovation (under preparation) Special fund for innovation platform construction (CCPTZX2023K03),Industrial Technology Innovation Program of IMAST(No. 2024RCYJ04004) and the National Natural Science Foundation of China (32271987).

About Plant Phenomics

Science Partner Journal Plant Phenomics is an online-only Open Access journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner Journal program, Plant Phenomics is editorially independent from the Science family of journals. Editorial decisions and scientific activities pursued by the journal's Editorial Board are made independently, based on scientific merit and adhering to the highest standards for accurate and ethical promotion of science. These decisions and activities are in no way influenced by the financial support of NAU, NAU administration, or any other institutions and sponsors. The Editorial Board is solely responsible for all content published in the journal. To learn more about the Science Partner Journal program, visit the SPJ program homepage.