Decision-Oriented Environmental Mapping with Radial Basis Function Neural Networks

The work deals with development and application of Radial basis functions neural networks (RBFNN) for spatial predictions. Geostatistical tools for spatial correlation analysis (variography) are used to qualify and quantify the estimation results. Geostatistical analysis is performed on the residuals obtained at the training and test sample locations. Variogram of residuals explores spatial correlation remaining after RBF modelling and is an independent criteria of the results quality. Networks with different number of hidden units and different RBF functions were considered. Extended RBF networks allow to obtain not only the estimate itself by also the estimate variance, which qualifies the prediction map. Two real case studies were considered: radioactive soil contamination after the Chernobyl fallout and heavy metal sediments in Lake Geneva. Decision-oriented prediction maps accompanied by maps of estimation errors and “thick contours” are presented as the outputs for decision making.