Envirotyping-Based Optimization of the Pan-African Soybean Trial Network Through Redundant Site Pruning

The expansion of soybean into new agricultural frontiers across Africa has intensified efforts to develop adapted cultivars and harnessing G×E interaction under high environmental heterogeneity. Establishing and maintaining multi-environment trials (METs) across such diverse landscapes is logistically complex and economically demanding. Therefore, identifying a smaller set of representative environments can substantially improve breeding efficiency by reducing experimental trials costs while maintaining the predictive power of experimental networks. In this context, this study aimed to: (i) define target populations of environments (TPEs) across the Southern African soybean trial network using envirotyping-based characterization; (ii) evaluate a data-driven strategy for pruning redundant trials to optimize experimental representativeness and maintain trial quality. Soybean yield trials were part of the Pan-African network coordinated by the Soybean Innovation Lab. The dataset comprised 97 yield trials conducted between 2017 and 2023/24 across 58 sites in Malawi, Mozambique, Zambia, and Zimbabwe, involving 180 genotypes. Environmental characterization integrated multi-source data from SoilGrids, NASA POWER, Open-Elevation API, and Sentinel-2 processed via Google Earth Engine to build an environmental matrix. TPEs were delineated using a PCA followed by k-means clustering of environmental covariates. Within each TPE, a factor-analytic (FA)  mixed model was fitted to estimate genotype performance and G×E patterns, and the top 20 genotypes were ranked using the FA Selection Tool index. To optimize network efficiency, a similarity-based pruning analysis was performed within each TPE to identify redundant trial sites. For each pair of environments, three concordance metrics were computed: (i) environmental similarity, based on Euclidean Ward distances of standardized PCA scores; (ii) genetic similarity, derived from the FA covariance matrix to obtain mean genetic correlations between sites; and (iii) rank concordance, estimated by Spearman correlations of genotype eBLUP rankings. An iterative threshold search adjusted all metric limits stepwise until reaching a maximum of 15% of sites considered redundant. Among each redundant pair, the site with higher CV (coefficient of variation) values was removed, ensuring that the remaining sites represented the most precise and informative environments in the network. Four distinct TPEs were identified, representing contrasting agroclimatic conditions across Southern Africa. TPE1 included cooler and drier highland environments with low precipitation. TPE2 comprised warm and humid regions with the highest rainfall and yield. TPE3 represented intermediate temperature and rainfall. TPE4 encompassed warm low-altitude areas with moderate-to-high rainfall and greater temperature variability. The environmental distance thresholds for site removal increased from 1.2 in TPE1 to 4.0 in TPE4, while mean genetic and rank correlations decreased from 0.88 and 0.93 in TPE1 to 0.60 and 0.65 in TPE4, indicating variable sensitivity to pruning among TPEs. Heritability ranged from 0.42 to 0.50 before pruning and from 0.39 to 0.51 afterward, showing subtle change in genetic precision. The overlap of the top 20 genotypes selected by the FAST index before and after pruning ranged from 60% to 75%, demonstrating that, even in datasets with limited connectivity, the pruning approach maintained consistent selection outcomes and improved environmental representativeness without compromising trial quality.