Accurately estimate soybean growth stages from UAV imagery by accounting for spatial heterogeneity and climate factors across multiple environments

Multi-environment trials (METs) are widely used in soybean breeding to evaluate soybean cultivars’ adaptability and performance in specific geographic regions. However, METs’ reliability is affected by spatial and temporal variation in testing environments, requiring further knowledge to correct such changes. To improve METs’ accuracy, the growth of 1303 soybean cultivars was accurately estimated by accounting for climatic effects and spatial heterogeneity using a linear mixed-effect model and a field spatial-correction model, respectively. The METs across 10 sites varied in climate and planting dates, spanning N16°41′52″ in latitude. A soybean growth and development monitoring algorithm was proposed based on the photothermal accumulation area (AUC_pt) rather than using calendar dates to reduce the impact of planting dates variability and climate factors. The AUC_pt correlates strongly with latitude of the above trial sites (r > 0.77). The proposed merit-based integrated filter decreases the influence of noise on photosynthetic vegetation (f_PV) and non-photosynthetic vegetation (f_NPV) more effectively than S-G filter and locally estimated scatterplot smoothing. The field spatial-correction model helped account for spatial heterogeneity with a better estimation accuracy (R² ≥ 0.62, RMSE≤0.17). Broad-sense heritability (H²) with the field spatial-correction model outperformed the models without the model by an average of 52 % across the entire aerial surveys. Model transferability was evaluated across Sanya and Nanchang. Rescaled shape models in Sanya (R² = 0.97) were consistent with the growth curve in Nanchang (R² = 0.89). Finally, the methodology's precision estimations of crop genotypes’ growth dynamics under differing environments displayed potential applications in precision agriculture and selecting high-yielding and stable soybean germplasm resources in METs.