Due to the variability between fields and across years, on-farm experimentation combined with crop response modeling are crucial aspects of decision support systems to make accurate predictions of yield and grain protein content in upcoming years for a given field. To maximize accuracy of models, models fit using environmental covariate and experimental data gathered up to the point that crop responses (yield/grain protein) are fit repeatedly over time until the model can predict future crop responses without access to future data. Besides the lack in clarity of how many years’ worth of data is required before the model can be used for prediction, when put in practical use, farmers will not have access to data past a decision point in the spring (March 30thfor MT wheat farmers). The objective of this project was to evaluate the tradeoff in ability to accurately predict future crop responses between using data up until the decision point to fit a model compared to using covariate data collected up to when crop responses are measured at harvest. It may seem obvious that training models with data constrained to a decision point will result in better predictions of crop responses at harvest when predictions must be made with data constrained to a decision point. However, this work fills a gap in the literature by explicitly testing and reporting on these results. Generalized additive models (GAMs) were trained under both data constraints, to predict crop responses at harvest using covariate data from both data constraints, creating a two-way factorial design where we compared predictive ability via RMSE. Using a model fit with covariate data up to harvest to predict data at harvest using future covariate data up until harvest provided the baseline of the hypothesized “best case” for predictive ability. We then compared the baseline predictive ability to the two realistic situations; 1) using the model fit with data up until a decision point to predict future harvest data with covariate data collected up until the decision point, and 2) using the model fit with data up until harvest to predict future harvest data with covariate data collected up until the decision point. As expected, using a model fit with data up until harvest resulted in the most accurate prediction of crop responses (yield/protein) when fed covariate data up until harvest. However, constrained to the realistic situation where only covariate data up until the decision point is available to make predictions, the models fit with data up until the decision point provided better predictions of future harvest compared to models fit with covariate data to harvest. The significance of these results is that decision support systems need to use models where training data is constrained to the data that would be available in the future for predicting crop responses.
