Intercontinental prediction of soybean phenology via hybrid ensemble of knowledge-based and data-driven models

Abstract

The timing of crop development has significant impacts on management decisions and subsequent yield formation. A large intercontinental dataset recording the timing of soybean developmental stages was used to establish ensembling approaches that leverage both discrete-time dynamical system models of soybean phenology and data-driven, machine-learned models to achieve accurate and interpretable predictions. We demonstrate that the knowledge-based, dynamical models can improve machine learning by generating expert-engineered features. Combining the predictions of the diverse component models via super learning resulted in a mean absolute error of 4.12 and 4.55 days to flowering (R1) and physiological maturity (R7), providing an improvement relative to the best benchmark model error of 6.90 and 15.47 days, respectively. The hybrid intercontinental model applies to a much wider range of management and temperature conditions than previous mechanistic models, enabling improved decision support as alternative cropping systems arise, farm sizes increase, and changes in the global climate continue to accelerate.

Can we harness digital technologies and physiology to hasten genetic gain in U.S. maize breeding?

“…This is a super wicked problem because the information needed to train data-driven models is only routinely available for few genotypes and creating training sets for many genotypes could be prohibitively expensive (Fig 1). The integration of physiology-based and data-driven approaches has been proposed as a workable solution whereby scientific understanding can effectively deal with model underdetermination (Messina et al, 2018; Hammer et al, 2019; Messina et al, 2020b; McCormick et al, 2020).…”

Interdisciplinary strategies to enable data-driven plant breeding in a changing climate

“…Combining these approaches offers the potential to improve prediction accuracies for decision-making in breeding while providing insights into underlying biological processes. 53,54 Historical models for yield prediction Prior to the advances in computing power that have enabled large-scale application of ML and CGMs, trait prediction and breeding methods relied heavily on quantitative genetics theory. The pioneering work of Charles Henderson on the use of pedigrees and mixed linear models for predictions of genetic merit in animal breeding was widely adopted in both animal and plant breeding.…”