Prediction of topsoil texture for Region Centre (France) applying model ensemble methods

Abstract

With the rapid development of digital soil mapping it is not unusual to find several maps for the same soil property in an area of interest. We applied two standard methods of model averaging for combining two regional maps and a European map of topsoil texture in agricultural land for the Region Centre (France). The two methods for model ensemble were the Granger-Ramanathan (G-R) and the Bates-Granger (B-G). A calibration dataset was used for fitting the coefficients of the G-R model, and for calculating a global variance: prediction error ratio which was then used to re-scale the weights of the B-G model. The prediction performance of the three primary maps and the two ensemble maps was compared with an independent validation dataset consisting on 100 observations from the French soil monitoring network. The prediction accuracy of the ensemble models improved only for clay in comparison to the primary maps (∆R2=0.02–0.06, ∆RMSE=−1.56–−4.97gkg−1). Overall, the G-R models obtained smaller RMSE and greater bias than B-G, and G-R estimated better the prediction uncertainty. The dissimilarities between the methods for estimating the prediction variance and non-optimal estimated uncertainties were important limitations for the B-G models despite applying a global correction factor for the prediction variances. The results suggested that both the calibration and validation datasets should represent the patterns of spatial variation and range of values of the soil property for the prediction space. Nonetheless, model ensemble methods proved to be useful for merging maps with different types of datasets, spatial coverage, and methodological approaches.