Spatial Estimation of Saturated Hydraulic Conductivity from Terrain Attributes Using Regression, Kriging, and Artificial Neural Networks

Several methods, including stepwise regression, ordinary kriging, cokriging, kriging with external drift, kriging with varying local means, regression-kriging, ordinary artificial neural networks, and kriging combined with artificial neural networks, were compared to predict spatial variation of saturated hydraulic conductivity from environmental covariates. All methods except ordinary kriging allow for inclusion of secondary variables. The secondary spatial information used was terrain attributes including elevation, slope gradient, slope aspect, profile curvature and contour curvature. A multiple jackknifing procedure was used as a validation method. Root mean square error (RMSE) and mean absolute error (MAE) were used as the validation indices, with the mean RMSE and mean MAE used to judge the prediction quality. Prediction performance by ordinary kriging was poor, indicating that prediction of saturated hydraulic conductivity can be improved by incorporating ancillary data such as terrain variables. Kriging combined with artificial neural networks performed best. These prediction models made better use of ancillary information in predicting saturated hydraulic conductivity compared with the competing models. The combination of geostatistical predictors with neural computing techniques offers more capability for incorporating ancillary information in predictive soil mapping. There is great potential for further research and development of hybrid methods for digital soil mapping.