Process-based models are increasingly used as tools for studying complex agroecosystem interactions and N2O emissions from agricultural fields. The widespread use of these models to conduct research and inform policy benefits from periodic model comparisons that assess the state of agroecosystem modeling and indicate areas for model improvement. The increasingly broad application of models requires an assessment of model performance using datasets that span multiple biogeophysical contexts. While limited in the capacity to identify specific areas for model improvement, general evaluations provide a critical perspective on the use of model estimates to inform policy and also identify necessary model improvements that require further evaluation.
What did we do?
The objectives of this model comparison were to i) calibrate and validate three process-based models using a large dataset; ii) evaluate the performance of a multi-model ensemble to estimate observed data; and iii) construct a linear model to identify and quantify possible model bias in the estimation of soil N2O flux from agricultural fields. We selected three models that have been used to evaluate N2O emissions from agricultural fields: DayCent, DNDC, and EPIC. Using data from two field experiments over five years, we calibrated and validated each model using observations of soil temperature (n = 887), volumetric soil water content (VSWC) (n = 880), crop yield (n = 67), and soil N2O flux (n = 896). Our model validations and comparisons consisted of commonly conducted statistical evaluations of root mean squared error, correlation, and model efficiency. Additionally, the large sample sizes used here allowed for more robust linear regression models that offered additional insight into relationships between model estimations and observations of N2O flux. We hypothesized that such a linear model would indicate if there was model bias in estimations of soil N2O flux. Ensemble modeling can reduce the error associated with climate projections and has recently been applied to the estimation of N2O flux from agroecosystems. Thus, we also constructed a multi-model ensemble to evaluate the use of multiple models to improve estimates of soil N2O flux.
What have we learned?
In a comparison of three process-based models, calibration to a large dataset produced favorable estimations of soil temperature, VSWC, average yield, and N2O flux when the models were evaluated using RMSE, R2, and the Nash-Sutcliffe E-statistic. However, an evaluation of linear regression models revealed a consistent bias towards underestimating high-magnitude daily N2O flux and cumulative N2O flux. Observations of soil temperature and VSWC were unable to significantly explain model bias. Calibration to available data did not result in consistent model estimation of additional system properties that contribute to N2O flux, which suggests a need for additional model comparisons that make use of a wide variety of data types. The major contribution of this work has been to identify a potential model bias and future steps required to evaluate its source and improve the simulation of nitrogen cycling in agroecosystems. Process-based models are powerful tools, and it is not our objective to undermine their past and future application. However, more work is left to be done in understanding the biogeophysical system that produces soil N2O and in harmonizing the process-based models that simulate that system and which are used to evaluate management and generate policy.
Future work should test our findings in additional agroecological contexts to determine the extent to which a bias towards underestimating peak N2O flux persists. A meta-analysis of published data may be the most direct method for doing so. New datasets will need to be collected that represent simultaneous observations of multiple system properties (e.g. soil NO3-, soil NH4+, and heterotrophic respiration) from different soil layers and at increased temporal frequencies. Model developers should use these rich datasets to identify the source of N2O estimation bias and improve the structure and function of process-based models.
Corresponding author, title, and affiliation
Richard K. Gaillard, Graduate Student, University of Wisconsin-Madison
Corresponding author email
Curtis D. Jones, Assistant Research Professor, University of Maryland; Pete Ingraham, Research Scientist, Applied Geosolutions;
Sarah Collier, Research Associate, University of Wisconsin-Madison;
Roberto Cesar Izaurralde, Research Professor, University of Maryland;
William Jokela, Research Scientist (retired), USDA-ARS;
William Osterholz, Research Associate, University of Wisconsin-Madison;
William Salas, President and Chief Scientist, Applied Geosolutions;
Peter Vadas, Research Scientist, USDA-ARS;
Matthew Ruark; Associate Professor; University of Wisconsin-Madison
The authors are solely responsible for the content of these proceedings. The technical information does not necessarily reflect the official position of the sponsoring agencies or institutions represented by planning committee members, and inclusion and distribution herein does not constitute an endorsement of views expressed by the same. Printed materials included herein are not refereed publications. Citations should appear as follows. EXAMPLE: Authors. 2017. Title of presentation. Waste to Worth: Spreading Science and Solutions. Cary, NC. April 18-21, 2017. URL of this page. Accessed on: today’s date.