Using neural networks as an alternative to statistical modeling in kriging interpolation procedures: an investigation

Abstract

Spatial interpolation techniques provide a means of predicting values of a variable of interest (e.g., an attribute such as elevation, rainfall, contaminant levels, soil type, etc.) at locations where, due to practical constraints, the variable cannot be measured. Kriging is one of the most widely used interpolation procedures, and it has been shown to produce accurate estimates in many cases. However, there are documented cases in which kriging does not produce accurate results. In these cases, it is possible that kriging procedures could be improved. This study explored the ability of a commonly used artificial neural network (ANN) to correct the limitations of the regression-based approach to semivariogram analysis commonly employed in conventional kriging. A hybrid ANN-based kriging model (KrigANN) was developed, in which a multilayer feedforward neural network trained using the MEKA algorithm replaced the traditional kriging's regression-based semivariogram development approach, while retaining all the other properties of traditional kriging. The accuracy and precision of the predictions produced by this hybrid model were compared to those obtained from the conventional regression-based kriging procedure by applying both models to 2250 artificially generated datasets created for this purpose. The study demonstrated that the KrigANN model produced more accurate interpolation results than regression-based kriging at low to medium degrees of spatial autocorrelation. These findings lead to the conclusion that the ANN-based procedures proposed in this study are appropriate alternatives to the conventional regression-based approaches, when spatial data exhibit low to medium degrees of autocorrelation.