Using digital terrain modelling for estimation of soil properties

Sustainable land management involves assessment and prediction of the spatial distribution of soil properties. Among spatial interpolation techniques geostatistics is generally preferred because predicting attribute values at unsampled locations it allows to take into account spatial correlation between neighbouring observations. A way of increasing the accuracy of predictions of sparsely observations of the primary attribute (time-consuming, labour-intensive and costly) is using ancillary information from more intensively recorded data such as digital elevation models. Kriging also allow that sparsely observations of the primary attribute can be complemented by secondary attributes more densely sampled.Objectives of the study were estimating a DEM using geostatistics and providing a realistic distribution of the errors and to demonstrate if using widely available secondary data provides more accurate estimates of soil properties than the ones obtained by using standard popular interpolation procedures such as ordinary kriging.The study area is a doline of approximately 1.5 ha at 1800 m above sea level and it belongs to the plateau of Andossi located in the northern part of Valchiavenna (the Alps, Italy).Elevation was measured at 467 randomly distributed points and they were converted into a regular DEM using ordinary kriging, while soil samples were collected at 110 pedological profiles and their locations were defined using Spatial Simulated Annealing (SSA) method.We used ordinary kriging (OK), which uses only soil data collected, considered as a reference and algorithms, which combine soil data with a digital terrain model (linear regression, simple kriging with varying local means (SKlm), kriging with external drift and multi-collocated ordinary cokriging). A cross-validation test was used to assess the prediction performances of the different algorithms and then evaluate which methods performed best.Results did show no clear differences among the methods and it is difficult to assign performance scores which have general validity. The advantage to incorporate secondary variable is real only when correlation coefficient with primary variable is greater than 0.75. Otherwise, the benefit of multivariate techniques is only marginal and the easier univariate OK could be used with an insignificant loss of estimate precision. When primary variable is not spatially stationary and a clear deterministic component is present in the trend related to topography, IRF-k kriging with external drift is to be preferred to SKlm. Multi-collocated cokriging assumes spatial stationarity of primary variable, therefore when there are more types of non-stationarity present, the greater complexity of this approach, because three semivariograms must be inferred and jointly modelled, may not pay off. Choosing the appropriate interpolation method to account for redundant auxiliary information is then an extremely flexible process, strongly conditioned by the particular contingent situations.