Improving yield estimates in on-farm experiments using vegetation indices as covariates

Purpose On-farm experiments often involve large plots because of the size of the farm’s equipment used and because plot size is chosen to achieve realistic conditions in the comparison of the treatments under test. We here consider on-farm experiments where plots are too large to fully harvest whole plots for assessing the primary response (e.g., yield). However, drone flights are used to assess the normalized difference vegetation index (NDVI) on the whole plot. Methods This paper considers methods to extrapolate the primary response for the whole plot using NDVI as a covariate. A bivariate mixed model is proposed as a plausible model for the data generation mechanism. Fitting the bivariate model implicitly achieves the extrapolation and provides a fully efficient analysis with respect to treatment comparisons for the primary response. The bivariate model may also be used to derive a regression approach by conditioning on the covariate¾akin to analysis of covariance (ANCOVA)¾that yields results very similar to bivariate analysis. We use data from two on-farm experiments, one with winter barley and one with oilseed rape, to illustrate and compare the methods. Results Our analyses show that bivariate analyses and ANCOVA provide improved yield estimates compared to univariate analyses. Conclusion Our recommendation is to use the ANCOVA approach for routine analysis because of its simplicity compared to fitting the bivariate model.