Using morphometric attributes to improve the prediction of compositional data: an application to soil texture in a forested catchment

Soil texture is a key property influencing most soil physical, chemical, and biological processes and its catchment-scale spatial variation may yields insights for soil management. Texture data are generally available only at a few locations, since sampling and laboratory analyses are time consuming. Therefore, it is essential: (i) to optimize the sampling design for improving the quality and effectiveness of predicted soil variability and (ii) to select an appropriate method for texture spatial prediction. In fact, the sum of the three soil textural fractions (sand, silt and clay) returns to 100 and traditional kriging algorithms do not meet such requirement when the three fractions are interpolated independently. Moreover, soil texture distribution across a catchment is influenced by erosion and sedimentation processes controlled by hillslope morphology. The study was developed within the project AlForlab PON03PE_00024_1, and was aimed to map soil texture variability within a forested catchment in southern Italy (Calabria), combining a sampling design technique with additive-logratio cokriging. The study focused on a 139 ha catchment on granitic parent material and subordinately alluvial deposits, where soils classified as Typic Xerumbrepts and Ultic Haploxeralf crop out. Soils samples were collected at 136 locations and the sample design was developed using a spatial simulated annealing algorithm with two steps. First, one half of the samples were located over the whole basin onto a 5 m grid, optimizing for minimal distance between observations. Then, since the variation in topography represents one of the target variable accounting for soil variability, the remaining samples were located using the Weighted Means of Shortest Distances algorithm, with slope gradient as a weigh function. In the field, soils were sampled up to a depth of 0.20 m and the geographical coordinates of each point were recorded with a differential Trimble receiver at 1-m accuracy. In the laboratory, after ovendrying soils, through sieve and hydrometer methods, the percentage of sand silt and clay was evaluated for each soil sample. A geostatistical approach was used to map the texture at the 0.20 m depth over the whole catchment, after randomly splitting data into a calibration and a validation set. To ensure constant sum of interpolated values at a given location, additive-logratio cokriging was used to remove the non-negativity constraint on compositional variables. To improve the estimation, primary and secondary terrain attributes, derived from a DEM of 1 m scale resolution, were used as auxiliary information within cokriging with external drift. The results of validation allowed assessing the improvement of auxiliary information for the goodness of the three textural fractions maps and the understanding of slope processes.