Fig. 2.

An example of data-driven integration, based on ref. 58. Two approaches are being compared: decentralized data-driven (3D) breeding and a benchmark. The arrows indicate for two approaches how data are linked into a single analytical model. Each model seeks to maximize both complexity and representativeness of the information that guides plant selection. 3D breeding covers more complexity because tricot on-farm testing achieves better predictions of farmers’ overall appreciation of varieties (a complex “trait” involving many variables apart from yield). Also, it attains higher representativeness because more environments could be sampled (locations and planting dates). Target production environment (TPE) characterization increases representativeness in both cases. Biased sampling of environments (locations and planting dates) in the multienvironmental trial limit representativeness, however, which could not be entirely corrected by linking with TPE characterization data.