Ecological Systems and Devices Annotation << Back A HYBRID APPROACH BASED ON KRIGING AND ARTIFICIAL NEURAL NETWORKS FOR THE HEAVY METAL DISTRIBUTION PREDICTION IN THE SURFACE SOIL LAYER OF AN ARCTIC URBANIZED TERRITORY A.G. Buevich, D.A. Tarasov, A.P. Sergeev, A.N. Medvedev, E.M. Baglaeva, I.E. Subbotina, M.V. Sergeeva The increase in an anthropogenic pressure, particularly at urban areas, enlarges the threat of deterioration in the habitat quality and the potential negative impact on the lives and health of citizens, as well as the possible negative effect on the environment as a whole. Soil is one of the major storage medium capable to provide objective information about the contamination. Detection of abnormally distributed elements at some areas (especially heavy metals), which distribution prediction by conventional methods is diffi cult or impossible is the substantial complexity during the monitoring of soil. Availability of models, able to predict the distribution of pollutants in heterogeneous environments, would greatly facilitate the monitoring and management of environmental risks in the study areas. The paper proposes for such a prediction a hybrid model, combining the traditional geostatistical techniques (kriging) and modern approaches (artifi cial neural networks). The model is used to predict an abnormal distribution of chromium and, for comparison, a conventional distribution of copper at the site in Tarko-Sale, Yamalo-Nenets Autonomous District, using the data of soil chemical analysis. The proposed model is built and tested using ArcGIS and Matlab. The confi gurations of the neural networks (multilayer perceptron) were adjusted individually for each item using a computer simulation: (2-10-1) for chromium, (2-9-1) for copper. The algorithm of application of the model is as follows: a comparison of predictions by kriging and neural network, the construction of residues as a difference between the predicted values, residual kriging and its combination with the estimates of the neural network. The predicted concentration distribution model is compared with the traditional method of inverse distance weighted (IDW). The results confi rmed that the trained neural network (multilayer perceptron) is suitable for modeling the spatial distribution of both regular and abnormally distributed elements. The predictive accuracy of the neural network is higher than in the geostatistical (kriging) and deterministic (IDW) methods. The kriging-ANN hybrid model can improve the predictive accuracy for the studied elements. Keywords: Artifi cial neural network; chromium; cuprum; kriging; contaminant. Contacts: E-mail: bagalex3@gmail.com Pp. 18-29.