Spatial prediction of soil texture in Region Centre (France) from summary data

Land use and agricultural practices, in interaction with soil texture, influence soil fauna and soil microbial community, water holding capacity, and C and nutrient cycling among other agroecosystem properties. Detailed knowledge on the spatial distribution of soil texture can improve land-use planning and crop management. Our objective was to predict soil texture in agricultural land for the Region Centre (France), combining regression models and area-to-point kriging. The French soil-test database (BDAT) is largely populated with topsoil analysis requested by farmers mainly interested in soil fertility. To protect the anonymity of the farms, their coordinates are unknown and texture is aggregated by municipality. The nature of the data requires novel disaggregation techniques (i.e., area-to-point kriging) to develop high-resolution maps on point support. We applied an additive log-ratio transformation (alr-transform) on texture data to remove the closure effect and achieve normality. Average values of environmental covariates by municipality were used to fit predictive models with multiple linear regression, Cubist, and boosted regression trees (BRT). Data from 104 plots from the systematic soil quality monitoring network (RMQS) were used for independent validation. Only BRT models provided better predictions (clay-alr R2 = 0.54, sand-alr R2 = 0.76) than reference BDAT texture values averaged by commune (clay-alr R2 = 0.33, sand-alr R2 = 0.64). In a second step, BRT predictions were used as auxiliary variables for area-to-point kriging following the summary statistics approach developed by Orton et al. (2012). To deal with the dependence between clay- and sand-alr transforms we applied a linear model of coregionalization. This approach allowed to include the relationships between soil forming factors and soil texture, and to account for the uncertainty on areal means in the area-to-point kriging step. We are currently testing whether incorporating remote sensing data (e.g., Landsat 8) in the regression models further improves soil texture predictions despite the loss of information when averaging by municipality. The combination of regression and area-to-point kriging is a promising method to produce high-resolution maps from soil-test data missing the exact coordinates.