Environmental correlation of three-dimensional soil spatial variability: a comparison of three adaptive techniques

• Published in: Geoderma Vol. 109; no. 1; pp. 117 - 140
• Main Authors: Park, S.J, Vlek, P.L.G
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.09.2002; Elsevier
• Subjects: Artificial neural networks; Earth sciences; Earth, ocean, space; Exact sciences and technology; Generalised linear model; Geochemistry; Regression trees; Soil and rock geochemistry; Soil spatial variability; Soils; Soil–landscape analysis; Surficial geology; England; United Kingdom; Great Britain; Somerset England; Western Europe; Soil spatial variability; Generalised linear model; Soil–landscape analysis; Regression trees; Artificial neural networks; landscapes; mobility; oxides; cation exchange capacity; Europe; neural networks; vegetation; manganese; linear models; principal components analysis; spatial distribution; lateral variations; depth; pedogenesis; soil profiles; pH; prediction; spatial variations; statistical analysis; soils
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/S0016-7061(02)00146-5

An appropriate inclusion of spatial variation of soils is becoming increasingly important for spatially distributed ecological modelling approaches. Even though soils are anisotropic vertically and laterally, most soil spatial variability studies have focused on the lateral variation of soil attributes over the landscape. This study characterizes the complexity of three-dimensional variations of individual soil attributes and examines the possibility of predicting soil property distribution using three different regression approaches: artificial neural networks (ANN), regression trees (RT) and general linear models (GLM). Thirty-two physiochemical attributes of 502 soil samples were collected from 64 soil profiles on a slope at Bicknoller Combe, Somerset, UK. After a principal component analysis, five soil attributes were selected to test for environmental correlation, assuming they reflect dominant pedological processes at the hillslope. Vegetation occurrence, soil types, terrain parameters and soil sample depth were used as predictors. Prediction using environmental variables was most successful for soil attributes whose spatial distribution is strongly influenced by lateral hydrological and slope processes with relatively simple depth functions (e.g. total exchangeable bases, Mn oxides and soil pH). These soil attributes also showed a high mobility, which implies that their spatial distribution quickly reaches an equilibrium with current slope processes. Soil taxonomic information only marginally improved the performance of models constructed from surface information such as vegetation and terrain parameters. On the other hand, soil attributes whose vertical distribution is strongly governed by vertical pedogenesis or unknown factors were poorly modelled by environmental variables due to stronger nonlinearity in their vertical distribution. Soil taxonomic information becomes more important for predicting these soil attributes. As an empirical modelling tool, GLM with interaction terms outperformed the other two methods tested, ANN and RT, in terms of both the simplicity of the model structure and the performance of derived empirical functions.