Estimating Spatially Variable Crop Response Functions Using On-Farm Precision Experimentation Trials

ALESSO, C.A.; BOLLERO, G.A.; CIPRIOTTI, P.A.; MARTIN, N.F.; GONCALVES-TRIVSAN, R.

Minneapolis

Conferencia; 15th International Conference on Precision Agriculture; 2020

ISPA

On-farm precision experimentation (OFPE) enables farmers and agronomists to gaininsights from their fields to guide site-specific decisions. Developing of site-specific prescriptions requires the approximation of crop response function to controllable inputs across the field. The geographically weighted regression (GWR) is one of the spatially varying coefficient (SVC) models proposed for the estimation of these functions. This method deals with the lack ofstationarity allowing the estimation of the regression coefficients locally. Despite its potential, this technique is sensitive to the kernel density function and bandwidth used in the selection of neighbors, which may result in the detection of misleading relationships. The effect of the spatiallayout of the experiment (treatment levels, plot sizes, randomization, etc.) and spatial structure on the estimation of these models has not been addressed. Detailed information about these effects are needed to improve OFPE methods. A simulation study was conducted using 216,000 yield data sets simulated on a 4,608 pixels grid of 9 m resolution. These yield maps were simulated using spatial variable maps of coefficientes drawn from 10 random fields byunconditional Gaussian geostatistical simulation technique and assuming linear response no 4 levels of nitrogen assignad by systematic and randomized chessboard desings. The simulated yield data were modeled using the GWR technique and the distribution of measures of agreement (correlation coefficient, mean absolute error) between the true and estimated mapsof regression coefficients. All designs tested in this study resulted in similar ability to capture or approximate the true spatial pattern of the response function. However, some differences related to the plot geometry (length and width) were observed among systematic designs. The ability of the GWR model to capture the true spatial variability of the response function wasmore related to the degree of spatial variability of those coefficients than the spatial arrangement of the treatments or the choices of model parameters (bandwidth, kernel). As the true spatial pattern of crop response function is always unkown the design tested showed similar average performance over a wide range of conditions.