Self-organizing map clustering technique for ANN-based spatiotemporal modeling of groundwater quality parameters

Abstract

The present study integrates co-kriging as spatial estimator and self-organizing map (SOM) as clustering technique to identify spatially homogeneous clusters of groundwater quality data and to choose the most effective input data for feed-forward neural network (FFNN) model to simulate electrical conductivity (EC) and total dissolved solids (TDS) of groundwater. The methodology is presented in three stages. In the first stage, a geostatistics approach of co-kriging is used to estimate groundwater quality parameters at locations where the groundwater levels are measured. In stage two, a SOM clustering technique is used to identify spatially homogeneous clusters of groundwater quality data. The dominant input data, selected by spatial clustering and mutual information are then imposed into the FFNN model for one-step-ahead predictions of groundwater quality parameters at stage three. The performance of the newly proposed model is compared to a conventional linear forecasting method of multiple linear regression (MLR). The results suggest that the proposed model decreases dimensionality of the input layer and consequently the complexity of the FFNN model with acceptable efficiency in spatiotemporal simulation of groundwater quality parameters. The application of FFNN for modeling EC and TDS parameters increases the accuracy of predictions respectively up to 84.5% and 17% on average with regard to the MLR model.