Evaluating differences among crop models in simulating soybean in-season growth

Kritika Kothari, Rafael Battisti, Kenneth J. Boote, Sotirios V. Archontoulis, Adriana Confalone, Julie Constantin, Santiago V. Cuadra, Philippe Debaeke, Babacar Faye, Brian Grant, Gerrit Hoogenboom, Qi Jing, Michael van der Laan, Fernando Antônio Macena da Silva, Fabio R. Marin, Alireza Nehbandani, Claas Nendel, Larry C. Purcell, Budong Qian, Alex C. RuaneCéline Schoving, Evandro H.F.M. Silva, Ward Smith, Afshin Soltani, Amit Srivastava, Nilson A. Vieira, Montserrat Salmerón

Research output: Contribution to journalArticlepeer-review


Crop models are useful tools for simulating agricultural systems that require continued model development and testing to increase their robustness and improve how they describe our current understanding of processes. Coordinated and “blind” evaluation of multiple models using same protocols and experimental datasets provides unique opportunities to further improve models and enhance their reliability. For soybean [Glycine max (L.) Merr.], there has been limited coordinated multi-model evaluations for the simulation of in-season plant growth dynamics. We evaluated ten dynamic soybean crop models for their simulation of in-season plant growth using data from five experiments conducted in Argentina, Brazil, France, and USA. We evaluated models after a Blind (using only phenology data) and a Full calibration (with in-season and end-of-season variables). Calibration reduced model uncertainty by reducing standard bias for the simulation of in-season variables (biomass, leaf, pod, and stem weights, and leaf area index, LAI). However, we found that most models had difficulty in reproducing leaf growth dynamics, with normalized root mean squared error (nRMSE) of 56% for leaf weight and 43% for LAI (across locations and models after Full calibration). Models with different levels of complexity and experience were capable of simulating final seed yield at maturity with reasonable accuracy (nRMSE of 8–31% after Full calibration). However, the nRMSE for pod weight (of 17–64% after Full calibration) was two-fold larger than that of seed yield. Moreover, the models differed in how they simulated timing from sowing to beginning seed growth (47–93 days) and effective seed filling period (18–54 days), owing to model structural differences in defining the reproductive developmental stages. Overall, we identified the following processes that can benefit from further model improvement: leaf expansion and senescence, reproductive phenology, and partitioning to reproductive growth. Simulation of pod wall tissue and individual seed cohorts is another aspect that many models currently lack. Model improvement can benefit from high-temporal resolution experimental datasets that concurrently account for phenology, plant growth, and partitioning. Further, we recommend collecting reproductive phenology in the field consistent with actual dry matter allocation to organs in the models and collecting multiple observations of seed and pod weight to aid model improvement for simulation of seed growth and yield formation.

Original languageEnglish
Article number109306
JournalField Crops Research
StatePublished - Apr 1 2024

Bibliographical note

Publisher Copyright:
© 2024


  • Agricultural Model Intercomparison and Improvement Project (AgMIP)
  • leaf area index
  • multi-model evaluation
  • pod growth
  • reproductive partitioning

ASJC Scopus subject areas

  • Agronomy and Crop Science
  • Soil Science


Dive into the research topics of 'Evaluating differences among crop models in simulating soybean in-season growth'. Together they form a unique fingerprint.

Cite this