Assessing performance of empirical models for forecasting crop responses to variable fertilizer rates using on-farm precision experimentation

• Published in: Precision agriculture 2023-04, Vol.24 (2), p.677-704
• Main Authors: Hegedus, Paul B., Maxwell, Bruce D., Mieno, Taro
• Format: Article
• Language: English
• Subjects: Agricultural technology; Agriculture; Artificial intelligence; Atmospheric Sciences; Bayesian analysis; Biomedical and Life Sciences; Chemistry and Earth Sciences; Computer Science; Crop yield; Crops; Decision making; Decision support systems; Decision trees; Experimentation; Farms; Fertilizers; Grain; Life Sciences; Mathematical models; Model forms; Performance assessment; Permutations; Physics; Precision farming; Regression models; Remote Sensing/Photogrammetry; Soil Science & Conservation; Statistics for Engineering; Training; Weather forecasting
• ISSN: 1385-2256; 1573-1618
• DOI: 10.1007/s11119-022-09968-2

Data-driven decision making in agriculture can be augmented by utilizing the data gathered from precision agriculture technologies to make the most informed decisions that consider spatiotemporal specificity. Decision support systems utilize underlying models of crop responses to generate management recommendations, yet there is uncertainty in the literature on the best model forms to characterize crop responses to agricultural inputs likely due for the most part to the variability in crop responses to input rates between fields and across years. Seven fields with at least three years of on-farm experimentation, in which nitrogen fertilizer rates were varied across the fields, were used to compare the ability of five different model types to forecast crop responses and net-returns in a year unseen by the model. All five model types were fit for each field using all permutations of the three years of data where two years were used for training and a third was held out to represent a “future” year. The five models tested were a frequentist based non-linear sigmoid function, a generalized additive model, a non-linear Bayesian regression model, a Bayesian multiple linear regression model and a random forest regression model. The random forest regression typically resulted in the most accurate forecasts of crop responses and net-returns across most fields. However, in some cases the model type that produced the most accurate forecast of grain yield was not the same as the model producing the most accurate forecast of grain protein concentration. Models performed best when the data used for training models was collected from years with similar weather conditions to the forecasted year. The results are important to developers of decision support tools because the underlying models used to simulate management outcomes and calculate net-returns need to be selected with consideration for the spatiotemporal specificity of the data available.