Predictive modeling of glyphosate adsorption index in agricultural soils

GIANNINI KURINA, F.; HANG, S.; CÓRDOBA, M.; BALZARINI, M.

Barcelona

Conferencia; XXIXth International Biometric Conference; 2018

Environmental studies demand statistical modeling of nature variables with spatial variability. Such is the case of herbicide soil adsorption coefficients (Kds) which characterize the behavior of a phytosanitary in soil. The Kd expresses the relationship between both, the amount of herbicide retained and the amount that remains in soil solution. In this work different modeling strategies have been evaluated to generate a predictive model of glyphosate Kds from 90 soil samples distributed across the territory of Cordoba, Argentina. Each sample was characterized by 20 edaphoclimatic covariables. The Kds follow a gamma distribution (0.035, 1.129). A log transformation was used to fit predictive models. First, boosting regression trees were used to select edaphoclimatic variables with mayor contribution in the variability of the Kds. The percentage of aluminum oxides, pH and texture were the edaphic properties of greater relevance to explain the herbicide dynamic in soil. Multiple Linear Regression (REML), Random Forest Regression (RFR), Generalized Boosted Regression (GBR) and Partial Least Square Regression (PLSR), all with spatial constraint on the residual terms, were fitted with the selected variables as predictors. Quadratic terms were also included. The predictive ability of the best fitted model was evaluated by mean squared prediction error (MSPE) calculated by leave one out validation. In addition, a punctual georeferenced error, expressed as percentage of the site mean (site specific error), was performed. The lowest MSPE, relative to the mean, corresponded to spatial PLS regression. However, through a quantile regression analysis, we found differences, both in magnitude and direction of regression coefficients among quantiles. Particularly in 0.1 quantile the Kds were overestimated. The GBR model improved the site specific error evidencing the importance of evaluating both, global and site specific prediction errors to better understanding complex spatial phenomena.