Spatial filtering of a legacy dataset to characterize relationships between soil organic carbon and soil texture

• Filtered kriging decomposed variability of soil properties by spatial scales.
• The filtering highlighted a probable bias in the legacy database.
• Correlations between soil properties/depth layers were increased by the filtering.

The spatial distribution of soil properties often displays complex and multiscale patterns of variation. It results from multiple soil processes acting simultaneously but at different scales. Hence, characterizing the influence of a given controlling factor on the soil property is made more difficult by the variation due to other controlling factors. In this context, separating the variation of the soil properties by spatial scales could allow disentangling the combined effect of controlling factors and would provide a qualitative and quantitative characterization of controlling factors separately. In this paper, geostatistical tools have been used to separate the scales of variation of two soil properties (i.e. SOC and texture) coming from a legacy dataset in the Belgian Loess Belt. Scale components were predicted separately and the relationships between soil properties were analyzed at different scales. Results illustrated that the contents of a given soil property in different depth layers were typically more correlated when only the long range components were compared. Similarly, the link between SOC and texture components was also clearer for the long range components. This means that soil processes acting at local or landscape scale influence soil properties differently according to their nature or to the depth considered. Eliminating the variation at this scale allows to better characterize the relationships between depth layers and soil properties. The study gives insights for further spatial mapping of SOC by focusing on more appropriate variables at specific spatial scales. Furthermore, we raise the interest of spatial filtering for detecting inconsistencies inside composite datasets.