Machine learning for unmanned aerial vehicles-based soybean phenotyping: limits of cross-environment transfer and opportunities to reduce field measurements

Abstract

High-throughput phenotyping using unmanned aerial vehicles (UAVs) and spectral vegetation indices has been proposed to overcome the cost and logistical constraints of manual measurements in multi-environment breeding trials. However, the reliability of models trained on spectral data to predict structural traits across genotypes and environments remains unclear. This study aimed to develop an approach for predicting soybean plant height (PH) and first pod insertion height (FPIH) using UAV-based vegetation indices acquired at the flowering stage, as well as to compare extreme gradient boosting (XGBoost), multilayer perceptron (MLP), random forest (RF), and multiple linear regression (MLR) models under realistic cross-validation scenarios. Trials were conducted across multiple seasons using UAV multispectral imagery, with PH and FPIH manually measured. The models were evaluated under five phenotyping scenarios: baseline calibration using all data; prediction in a completely unmeasured future season; estimation of missing genotypes within a partially sampled season; calibration using a small fraction of data from a new season; and prediction under absence of field records for specific genotypes across environments. When all data were used for calibration, non-linear models showed a high apparent accuracy. However, prediction in unseen seasons failed for all models, reflecting strong genotype × environment interactions. Under reduced phenotyping within the same environment network, the models maintained a robust accuracy for PH, whereas FPIH predictions declined to moderate levels. UAV-based models are reliable for interpolation, but limited for extrapolation without local calibration, enabling reductions of up to 80 % in manual measurements for PH and 20-30 % for FPIH.

KEYWORDS: Glycine max, multispectral remote sensing, canopy architecture.