Unraveling uncertainty drivers of the maize yield response to nitrogen: A Bayesian and machine learning approach

Adrian A. Correndo, Nicolas Tremblay, Jeffrey A. Coulter, Dorivar Ruiz-Diaz, David Franzen, Emerson Nafziger, Vara Prasad, Luiz H.Moro Rosso, Kurt Steinke, Juan Du, Carlos D. Messina, Ignacio A. Ciampitti

Research output: Contribution to journalArticlepeer-review

20 Scopus citations


Development of predictive algorithms accounting for uncertainty in processes underpinning the maize (Zea Mays L.) yield response to nitrogen (N) are needed in order to provide new N fertilization guidelines. The aims of this study were to unravel the relative importance of crop management, soil, and weather factors on both the estimate and the size of uncertainty (as a risk magnitude assessment) of the main components of the maize yield response to N: i) yield without N fertilizer (B0); ii) yield at economic optimum N rate (YEONR); iii) EONR; and iv) the N fertilizer efficiency (NFE) at the EONR. Combining Bayesian statistics to fit the N response curves and a machine learning algorithm (extreme gradient boosting) to assess features importance on the predictability of the process, we analyzed data of 730 response curves from 481 site-years (4297 observations) in maize N rate fertilization studies conducted between 1999 and 2020 in the United States and Canada. The EONR was the most difficult attribute to predict, with an average uncertainty of 50 kg N ha−1, increasing towards low (<100 kg N ha−1) and high (>200 kg N ha−1) EONR expected values. Crop management factors such as previous crop and irrigation contributed substantially (∼50%) to the estimation of B0, but minorly to other components of the maize yield response to N. Weather contributed about two-thirds of explained variance of the estimated values of YEONR, EONR, and NFE. Additionally, weather factors governed the uncertainty (72% to 81%) of all components of the N response process. Soil factors provided a consistent but limited (10% to 23%) contribution to explain both expected N response as well as its associated uncertainties. Efforts to improve N decision support tools should consider the uncertainty of models as a type of risk, potential in-season weather scenarios, and develop probabilistic frameworks for improving this data-driven decision-making process of N fertilization in maize crop.

Original languageEnglish (US)
Article number108668
JournalAgricultural and Forest Meteorology
StatePublished - Dec 15 2021

Bibliographical note

Publisher Copyright:
© 2021 Elsevier B.V.


  • Bayesian models
  • Fertilizer N response
  • Maize
  • Response curves
  • Uncertainty


Dive into the research topics of 'Unraveling uncertainty drivers of the maize yield response to nitrogen: A Bayesian and machine learning approach'. Together they form a unique fingerprint.

Cite this