Strategies for Disaggregating Soil Patterns in a Soil-Landscape

Digital soil modeling has been extensively used to predict soil properties over large regions with many different quantitative and knowledge-based methods. This raises numerous questions: Which model is appropriate to model soil properties in a given soil-landscape? Is there an universal equation or model that fits all soil-landscapes or do we need region-specific models to describe soil properties within a restricted domain? Commonly, soil properties show varying degrees of spatial variability and distribution patterns, which are often anisotropic and complex. Some of this soil variability across soil-landscapes can be explained by ancillary environmental properties (or CLORPT factors) using environmental correlation (EC-model) or by spatial models that quantify spatial autocorrelations of soil properties (S-model). The objective of this study was to compare EC- and S-based soil predictions under various scenarios. Fields were generated in a geographic information system that showed short, long, linear, random patterns and combinations of them representing different soil patterns resembling real soil-landscapes. Then various combinations of EC and S-models were compared using cross-validation statistics to identify their ability to describe soil patterns accurately. Findings suggest that the accuracy of S-model predictions declined in the order of linear > long > long/linear > short/linear > long/short > long/random > short > random and short/random soil patterns. This illustrates the importance to map soil properties at a scale appropriate to unravel the underlying spatial variability of a given property. Depending on the underlying spatial structure (variability) and distribution patterns of soil properties the S-model, EC-model or combinations of them are most appropriate to generate soil grids. Thus, domain boundaries should be given special attention. We exemplify the approach disaggregating patterns of soil total phosphorus, nitrogen and calcium and incorporating remote sensing imagery into the modeling process in a subtropical wetland in Florida.