Comparison of artificial neural network and decision tree methods for mapping soil units and soil salinity in Ardakan region.

In response to the demand for soil spatial information and in order to improve natural resource management outcomes through the development of soil suitability maps, the acquisition of digital auxiliary data and matching it to field soil observation is increasing. With the harmonization of these data sets, through computer based methods, so-called Digital soil Maps are increasingly being found to be as reliable as traditional soil mapping practices but without the prohibitive costs. Therefore, at present research, we have attempted to develop decision tree (DTA) and artificial neural network (ANN) models for spatial prediction of soil taxonomic classes and soil salinity in an area covering 720 km2 located in arid region of central Iran where traditional soil survey methods are very difficult to undertake. In this area, using the conditioned Latin hypercube sampling method, location of 187 soil profiles were selected, which then described, sampled, analysed and allocated in taxonomic classes according to soil taxonomy of America. Auxiliary data used in this study to represent predictive soil forming factors were terrain attributes (derived from a digital elevation model), Landsat 7 ETM+ data, apparent electrical conductivity (ECa)—measured using an electromagnetic induction instrument (EMI)—and a geomorphologic surfaces map. Results showed that the DTA had the higher accuracy than ANN for prediction of soil classes (great group). DTA had 67.6% and ANN had 48% overall accuracy. Similar results found in soil salinity modeling up to 1m. Regression tree model could find the strong relationship between soil salinity data derived from depth function and ancillary data. For DTA and ANN models, root mean square error and R2 values (i.e. 0-15 cm) were 33.99, 30.61, 0.80 and 0.85 respectively. Our results showed some auxiliary variables had more influence on predictive soil class (great group) model which included: wetness index, geomorphology map and multi-resolution index of valley bottom flatness. The same auxiliary data and apparent electrical conductivity were selected as the most powerful predictors in regression tree model. In general, results showed that decision tree models had higher accuracy than ANN models and also their results are more convenient for interpretation. With application of these rules, soil class (sub-great group level) and soil salinity (up to 1m) maps produced. Therefore, it is suggested using of decision tree models for spatial prediction of soil properties (category and continuous soil data) in future studies.