Dmitry Tarasov

Multilayer perceptron, generalized regression neural network, and hybrid model in predicting the spatial distribution of impurity in the topsoil of urbanized area

The study is based on the data obtained as a result of soil screening in the city of Noyabrsk, Russia. A comparison of two types of neural networks most commonly used in this type of research was carried out: multi-layer perceptron (MLP), generalized regression neural network (GRNN), and a combined MLP and ordinary kriging approach (MLPRK) for predicting the spatial distribution of the chemical element Chromium (Cr) in the surface layer of the urbanized territory. The model structures were developed using computer modeling, based on minimizing of a root mean squared error (RMSE). As input parameters, the spatial coordinates were used, and the concentration of Cr - as the output. The hybrid MLPRK approach showed the best prognostic accuracy.

Forecasting of chromium distribution in subarctic noyabrsk using generalized regression neural networks and multilayer perceptron

The paper deals with application of the artificial neural network (ANN) to the spatial prediction of soil pollution by Chromium (Cr). In the work, we examined and compared two neural networks: Generalized Regression Neural Network (GRNN) and Multilayer Perceptron (MLP) as classes of neural networks widely used for continuous function interpolation. For comparison, a geostatistical method co-kriging was also used. The case study is based on the survey on surface contamination by Cr at the subarctic Noyabrsk, Russia. The proposed models have been built, implemented and validated using ArcGIS and MATLAB software. The models frameworks have been developed using a computer simulation based on a minimization of the root mean squared error (RMSE). Each network has its own benefits and drawbacks; however both demonstrated fast training and good prediction possibilities. MLP showed the best predictive accuracy comparing to co-kriging and even to GRNN.

Chromium distribution forecasting using multilayer perceptron neural network and multilayer perceptron residual kriging

It is known that combination of geostatistical interpolation techniques (e.g. kriging) and machine learning (e.g. neural networks) leads to better prediction accuracy and productivity. The paper deals with application of the artificial neural network residual kriging (ANNRK) to the spatial prediction of soil pollution by Chromium (Cr). In the work, we examined and compared two neural networks: Multilayer Perceptron (MLP) and Multilayer Perceptron Residual Kriging (MLPRK). The case study is based on the survey on surface contamination by Cr at the subarctic Noyabrsk, Russia. The proposed models have been built, implemented and validated using ArcGIS and MATLAB software. The models frameworks have been developed using a computer simulation based on a minimization of the root mean squared error (RMSE). Both models showed almost identical results.

Artificial neural networks selection for soil chemical elements distribution prediction

In this work we suggest the way for selection of the type and structure of an artificial neural network (ANN) for restoring the topsoil spatial distribution of chemical elements. To restore the surface distribution using computer modeling, various types and structures of ANN were chosen. For each chemical element, its own ANN was selected and its own estimation of the prediction accuracy was used. Comparison of the concentrations distributions of chemical elements were made by different ANNs with preliminary known values of concentrations. It is known that trained ANNs provide high prediction accuracy. The proposed approach allows choosing type and structure of a neural network for an arbitrary site, which is one of the main difficulties in modeling the distribution of chemical elements by the ANN method.

Ink-Jet Printer’s Characterization by 3D Gradation Trajectories on an Equidistant Color Difference Basis

We suggest using 3D gradation curves of CIE Lab space, which we call “gradation trajectories”, as further development of common gradation curves. The trajectories are considered in terms of 3D curves of differential geometry. We offer the gradation trajectories, as well as their calculating method, as a powerful tool for ink-jet system characterization and further profile-making. In the work, we develop our method and apply it to ink-jet printer’s characterization on a basis of equidistant color difference CIE Lab ΔE. We discuss the information that might be derived from the trajectories’ analysis and show how they might me generally applicable.

Gradation Surfaces as a Method for Multi-color Ink-Jet Printers Color Specifications Management

We offer 3D gradation surfaces as a further development of 3D gradation curves approach. The surfaces reflect the interaction between selected pair of colorants in CIE Lab space. The surfaces are strained on the gradation trajectories of selected colorants. They are considered in terms of 3D surfaces of differential geometry. Application of the method leads to increase the smoothness of the gradient in the pair. Geodesic lines of the surfaces are suggested to be gradation trajectories of binary colors impositions. In the work, we develop the approach and discuss mathematical methods of the surfaces’ description as well as geodesic lines’ estimation.

Modeling of surface dust concentrations using neural networks and kriging

Creating models which are able to accurately predict the distribution of pollutants based on a limited set of input data is an important task in environmental studies. In the paper two neural approaches: (multilayer perceptron (MLP)) and generalized regression neural network (GRNN)), and two geostatistical approaches: (kriging and cokriging), are using for modeling and forecasting of dust concentrations in snow cover. The area of study is under the influence of dust emissions from a copper quarry and a several industrial companies. The comparison of two mentioned approaches is conducted. Three indices are used as the indicators of the models accuracy: the mean absolute error (MAE), root mean square error (RMSE) and relative root mean square error (RRMSE). Models based on artificial neural networks (ANN) have shown better accuracy. When considering all indices, the most precision model was the GRNN, which uses as input parameters for modeling the coordinates of sampling points and the distance to the proba...