Spatially restricted partial least square regression to explain within field grain yield variability

CÓRDOBA, M.; BALZARINI, M.; VEGA, A.; PACCIORETTI, P.

Montpellier

Conferencia; 12th European Conference on Precision Agriculture; 2019

The new information technologies that allow us to capture different types of data associated with spatial localization have been promoted in the last decades. Grain yield and several site variables are intensively measured within a crop field. Therefore, the challenging is to use all available data to better understand the agronomical process underlying within field yield variability. Because of the site variables are usually correlated between them (multicollinearity) and present spatially auto-correlated, the regression models to evaluate de relative contribution of each site variable should account for both types of correlations. Here we propose an extended version of partial least square regression (PLS) (Abdi, 2010) as designed algorithms to treat the multicollinearity and spatial autocorrelation. The algorithm combines PLS and ordinary kriging (OK) and is named as spatial PLS (sPLS). Initially, a PLS regression technique of yield using predictive ancillary variables was carried out in order to model the trend component. In the second step, OK is applied to the residuals of PLS and a spatial prediction of the residuals was created. The final prediction was an additive combination of both models.