Modelling site-specific yield response to agronomic inputs in on-farm precision experimentation

GIANNINI KURINA, F.; BALZARINI, M.; CÓRDOBA, M.; PACCIORETTI, P.; BULLOCK, D. S.

Congreso; 31 st International Biometric Conference; 2022

With the availability of precision agriculture technology, there is an increasing number of on-farm experiments (OFE) conducted cooperatively by farmers and researchers, in which rates of agronomic inputs (treatments) are highly repeated on plots spread across the crop field. Usually, OFE follow a classical randomized designs with many plots and hundreds of yield data collected per plot. The objective of this work was to assess hierarchical Bayesian regression (BR) models that can be used by researchers analysing OFE data to estimate a field-specific productivity function and describe its spatial within-field variability. We implemented, in several OFE datasets from the Data Intensive Farm Management (DIFM) project, a hierarchical Bayesian regression model with fixed treatment effects, random plot effects, and treatment-by-site interactions, with and without including site information such as soil and topographic covariates. The sites corresponding to yield observations (in average, 350 plots and 120 sites per plot) and the soil variables used as covariate in the regression models were aligned in a joint spatial model of yield, covariate and site effects estimated by Integrated Nested Laplace Approximations (INLA). The results showed that inclusion of site covariates allows to decompose and partially explain site effects. Further, site covariate effects on yield were spatialized within the field. The selected BR model was able to deliver a complete probability distribution of predicted yields at each plot within the field. We assessed predictive capacity of the fitted models through prediction error measures calculated with spatial cross-validation. We conclude that the BR for spatial data is a useful tool to process OFE data, allowing direct derivation of prediction uncertainty measures related to site-specific yield response to inputs. The spatialized expected response to a given input rate offers a key information to prescribe variable-rate inputs across the field for future crops. Thus, modelling yield response to variable rate of inputs and site effects within a field contribute to better farm management decisions.