Knowledge discovery from models of soil properties developed through data mining

We modelled the distribution of soil properties across the agricultural zone on the Australian continent using data mining and knowledge discovery from databases (DM&KDD) tools. Piecewise linear tree models were built choosing from 19 climate variables, digital elevation model (DEM) and derived terrain attributes, four Landsat multi-spectral scanner (MSS) bands, land use and lithology maps as predictors of topsoil and subsoil pH, organic C, % clay, and total N and P. The actual geographic location of the sampled soil data points was not used as a predictor. Classification trees were used to estimate topsoil and subsoil horizon thickness and texture class using similar predictors. That maps could be made using the decision tree models attests to the occurrence of knowledge discovery from the soil point databases used as training data. The decision tree models were evaluated and interpreted in a spatial context by: (1) tabulation of variables selected by the tree models; (2) mapping of the extent over which individual predictors were used and their thresholds; and (3) mapping of the extent over which combinations of predictors were used. The evaluation and interpretation process indicates that the models are consistent with general principles of soil genesis and that detailed investigation of the models’ structure in a spatial context may have other uses in biogeographical studies and geo-ecological process modelling.