Assessing spatially heterogeneous scale representation with applied digital soil mapping

Abstract

Topographic data are increasingly important to environmental models as fine-scale resolution, wide coverage data sets become available. Scale is an important consideration for predictive model quality. Recent advances in multiscale terrain analysis led to scaling techniques that allow the scale at which a topographic parameter is represented to vary spatially. This research compared predictive soil model performance across feature sets generated with different scaling strategies; including multiple heterogeneous strategies, common feature selection algorithms applied to homogeneously scaled data, and unscaled data. Model performance was assessed for accuracy and uncertainty. The results showed that unscaled data performed worse in all circumstances compared to multiscale feature sets. Overall, heterogeneous and homogeneous feature sets did not differ substantially in accuracy, prediction uncertainty, or error. However, one scaling strategy exploited the flexibility of heterogeneous scaling to consistently perform better than other feature sets for most soil properties in terms of accuracy, and consistently ranked among the least uncertain and least error prone (up to a 0.080 increase in accuracy with a corresponding 0.017 decrease in prediction uncertainty and 0.011 decrease in error relative to the second best method, in the case of the proportion of clay modelled at 5–15 cm depth). This was achieved by decoupling the definition of process scales from analytical parameterization, allowing the optimization to occur within broadly defined process scales. This research demonstrates how to exploit heterogeneous scaling of topographic attributes to improve model performance.