Soybean yield estimation and lodging discrimination based on lightweight UAV and point cloud deep learning.

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Unmanned aerial vehicles (UAVs) are valuable for soybean breeding, but past research often missed 3D structural data.
• Existing feature fusion methods can separate spatial and spectral information, limiting analysis.

Purpose of the Study:

• To develop and evaluate novel point cloud deep learning models for soybean phenotype research.
• To integrate 3D spatial structure with RGB color and vegetation index (VI) spectral information for enhanced analysis.

Main Methods:

• Utilized cross-circling oblique (CCO) photography and Structure-from-Motion with Multi-View Stereo (SfM-MVS) for 3D soybean canopy reconstruction.
• Developed new point cloud deep learning models (SoyNet, SoyNet-Res) with data-level fusion of spatial and spectral information.
• Employed multi-task learning within the SoyNet-Res model for simultaneous yield estimation and lodging discrimination.

Main Results:

• Integrating spatial structure with RGB and VI spectral data significantly reduced RMSE for yield estimation (22.55 kg ha⁻¹) and improved F1-score for lodging discrimination (0.06).
• The SoyNet-Res model with multi-task learning outperformed H2O-AutoML in yield estimation (RMSE: 349.45 kg ha⁻¹).
• Multi-task deep learning achieved superior lodging discrimination accuracy (top-2: 0.87, top-3: 0.97) compared to single-task learning.

Conclusions:

• Point cloud deep learning effectively integrates multi-phenotype data for soybean breeding.
• The developed models and fusion techniques offer a foundation for optimizing soybean breeding programs through precision agriculture.