Field phenotyping for soybean density tolerance using time-series prediction and dynamic modeling.

Area of Science:

• Agricultural Science
• Plant Breeding
• Machine Learning in Agriculture

Background:

• Increasing global food demand necessitates soybean varieties resilient to dense planting for high, stable yields.
• Traditional phenotyping lacks temporal resolution, hindering analysis of canopy development and yield stability relationships.
• Existing machine learning models often overlook temporal dependencies, limiting biological interpretability in time-series predictions.

Purpose of the Study:

• To develop an innovative approach integrating spatiotemporal deep learning and dynamic modeling for quantifying canopy parameter changes.
• To reveal key regulatory mechanisms of traits associated with soybean resistance to dense planting using UAV high-throughput phenotyping.
• To establish a high-precision, interpretable phenotypic analysis framework for screening soybean varieties resilient to dense planting.

Main Methods:

• Conducted a two-year field experiment in northeast China with high (50w plants/ha) and low (30w plants/ha) density treatments across 208 soybean varieties.
• Acquired multispectral UAV imagery (15-18 times/season) and ground-truth data to develop a time-series prediction model for Leaf Area Index (LAI) using Spatiotemporal Residual Networks (ST-ResNet).
• Extracted 15 intermediate traits from fitted time-series curves (LAI, Canopy Cover, Plant Height) using P-spline, and analyzed trait correlations with dense planting yield index (ΔYield) using mixed models and SHAP.

Main Results:

• The ST-ResNet model achieved superior LAI prediction accuracy (R² = 0.90, RMSE = 0.23 m²/m²), effectively capturing continuous canopy growth dynamics.
• The intermediate trait ΔMean_LAI-mid showed the highest correlation (r = 0.51) with the dense planting yield index (ΔYield).
• UAV-based high-throughput phenotyping enabled efficient screening of 208 varieties annually, significantly outperforming traditional methods.

Conclusions:

• The integration of spatiotemporal deep learning with dynamic trait modeling significantly improves the temporal continuity and stability of LAI estimation.
• This approach enables precise quantification of canopy development rates and systematic analysis of their influence on dense planting resistance.
• The study provides a high-precision, interpretable framework for effectively screening soybean varieties resilient to dense planting, crucial for future food security.