MODELLING THE DISTRIBUTION OF SOILS AND SOIL PROPERTIES ACROSS THE AUSTRALIAN CONTINENT

We have modelled the distribution of soils and soil properties across the agricultural zone on the Australian continent using data mining and knowledge discovery from databases (DM&KDD) tools. Piecewise linear decision trees were built using 19 climate variables, DEM and derived terrain attributes, 4 Landsat MSS bands, and lithology maps as predictors of soil pH, organic C... The climatic variables were derived from daily climate records for the 20-year period from 1980 to 1999. From these daily records 20-year monthly average surfaces were interpolated for 19 climatic variables: annual mean temperature, mean diurnal range, isothermality, temperature seasonality, maximum temperature of warmest month, minimum temperature of coolest month, temperature annual range, annual precipitation, precipitation of the wettest and driest months, precipitation seasonality, annual mean radiation, highest and lowest monthly radiation, radiation seasonality, annual mean moisture index, highest and lowest month moisture index, and moisture index seasonality. Predictor variables can feature as splitting criteria in sub-setting the dataset and in the linear regressions at each leaf of the decision trees. The details of the method have been submitted for publication (Henderson et al. 2001, http://www.clw.csiro.au/publications/technical2001) so here I will focus on whether there is any knowledge discovery from data mining of the soil-landscape databases in Australia.

The climatic variables feature prominently in all models but different climatic variables appear in different models. For the topsoil pH model, the precipitation in the driest month and the annual mean soil moisture feature most prominently in the condition (sub-setting) rules and in the regression rules. For the subsoil pH model, the annual mean soil moisture and the lowest monthly solar radiation are most important. For the topsoil organic C model, the annual mean soil moisture is most important. The most used terrain attributes are generally elevation, relative elevation, and relief. The four bands from the Landsat MSS are also important predictors in all models. What and where variables are being used in each of the predictions will be presented and discussed.