Representing soil landscapes from digital soil mapping products – helping the map to speak for itself

Abstract

Since the earliest days of soil geography, it has been clear that soils occur in more-or-less clearly mappable bodies, within which soil forming factors have been either fairly homogeneous or in a regular pattern within the body, and between which there is usually a clear transition in one or more factors. This has been the basis for polygon-based soil mapping: make a concept map from landscape elements leading to a mental model of the landscape, confirm or modify it with strategically placed observations, find the transitions, delineate the soil bodies, and characterise them. By contrast, common methods of Digital Soil Mapping (DSM) predict per pixel over a regular grid, from training observations at pedon support. Accuracy assessment of DSM products has been at this “point” support, ignoring the existence of spatial soil bodies and the relations between pixels. Different approaches to DSM – datasets, model forms, analyst choices – result in maps with distinctly different patterns of predicted soil properties or types. Techniques from landscape ecology have been used to characterize spatial patterns of DSM products. The question remains as to how well these products reproduce the actual soil patterns at a given cartographic scale and categorical level of detail. Our approach is to help DSM maps to “speak for themselves” and thereby reveal spatial patterns that have been found by the DSM. We do this by grouping predictions at the individual pixel level, either (1) by aggregation based on property homogeneity using the supercells algorithm, or (2) by segmentation based on within-block property pattern similarity, using the GeoPAT suite of computer programs. Segments can be hierarchically clustered into groups of presumed soil landscape elements. Supercells and segments can be compared to existing soil maps, other land resource maps, and expert judgement. To the extent that the presumed soilscape patterns are reproduced, this is evidence that DSM has identified the soil landscape at the chosen scale. Since map users perceive patterns, and most land use decisions are for areas rather than pixels, we propose that DSM products be evaluated by their patterns as revealed by aggregation and segmentation, as well as by pointwise evaluation statistics.