вход по аккаунту



код для вставкиСкачать
A hybrid method for assessment of soil pollutants spatial distribution
D. A. Tarasov, A. N. Medvedev, A. P. Sergeev, A. V. Shichkin, and A. G. Buevich
Citation: AIP Conference Proceedings 1863, 050015 (2017);
View online:
View Table of Contents:
Published by the American Institute of Physics
Articles you may be interested in
Topsoil pollution forecasting using artificial neural networks on the example of the abnormally distributed heavy
metal at Russian subarctic
AIP Conference Proceedings 1836, 020024 (2017); 10.1063/1.4981964
Review and possible development direction of the methods for modeling of soil pollutants spatial distribution
AIP Conference Proceedings 1863, 050014 (2017); 10.1063/1.4992211
Modeling of surface dust concentrations using neural networks and kriging
AIP Conference Proceedings 1789, 020004 (2016); 10.1063/1.4968425
High variation subarctic topsoil pollutant concentration prediction using neural network residual kriging
AIP Conference Proceedings 1836, 020023 (2017); 10.1063/1.4981963
Modeling of surface dust concentration in snow cover at industrial area using neural networks and kriging
AIP Conference Proceedings 1836, 020033 (2017); 10.1063/1.4981973
A Hybrid Method for Assessment of Soil Pollutants Spatial
D.A. Tarasov1,2,a), A.N. Medvedev1,2,b), A.P. Sergeev1,2,c), A.V. Shichkin 1,2,d) and
A.G. Buevich1,e)
Institute of Industrial Ecology, Ural Branch of Russian Academy of Sciences, Ekaterinburg, RUSSIA
Ural Federal University, Ekaterinburg, RUSSIA.
Corresponding author: [email protected]
[email protected]
[email protected]
[email protected]
[email protected]
Abstract. The authors propose a hybrid method to predict the distribution of topsoil pollutants (Cu and Cr). The method
combines artificial neural networks and kriging. Corresponding computer models were built and tested on real data on
example of subarctic regions of Russia. The network structure selection was based on the minimization of the Rootmean-square error between real and predicted concentrations. The constructed models show that the prognostic accuracy
of the artificial neural network is higher than in case of the geostatistical (kriging) and deterministic methods. The
conclusion is that hybridization of models (artificial neural network and kriging) provides the improvement of the total
predictive accuracy.
The environment components such as air, snow, water, soil, biota, bottom sediment etc. might be served as
recipients of large amounts of pollutants from the multiple sources and, therefore, can be used for studying the
nature and characteristics of the pollution. The environmental monitoring data, obtained at urban territories, strongly
depend on relative position and intensity of emission sources as well as building peculiarities, climate variability,
meteorological and hydrological conditions, and other factors. These processes and factors may cause the spatial
heterogeneity and in some cases anomalies of the environmental pollution. In some studies [8, 9], the content of
insoluble form of Cr in soil was considered as abnormal, because this pollutant showed an unexpectedly high
concentration at some studied territories of Russia’s subarctic. In this context, the tasks of objective and reliable
detection and delineation of eco-geochemical anomalies and the forecasting of pollutants distribution become
Geostatistical interpolation techniques (e.g. kriging) utilize the statistical features of the measured spots together
with the spatial autocorrelation between them and account for the spatial configuration of the sample spots at the
prediction location. Thus, kriging has shown considerable advantages in the prediction of soil properties, compared
with deterministic methods [6, 7, 12]. The accuracy of kriging techniques depends on the density and size of
sampling sites, as these methods are based on interpolation, which requires some data as inputs. Therefore, a more
efficient method is required to improve the accuracy of interpolation methods for producing high-resolution
distribution maps. Nowadays, the famous suitable technique is artificial neural networks (ANNs). A brief overview
of ANNs [2] showed how ANNs can be generally applicable. They are ahead of many other methods in terms of
accuracy and speed. At present, most researchers focus primarily on the multilayer perceptron (MLP) (Fig. 1a). Lots
of studies devoted to soil research and element distribution prediction also utilize MLP networks [1, 3, 4, 5, 11, 13].
International Conference of Numerical Analysis and Applied Mathematics (ICNAAM 2016)
AIP Conf. Proc. 1863, 050015-1–050015-4; doi: 10.1063/1.4992212
Published by AIP Publishing. 978-0-7354-1538-6/$30.00
In this work we combine geostatistical and neural methods for more accurate prediction of pollutants distributions in
soil of Tarko-Sale town located in subarctic region of Russia.
Data for the study were obtained from the results of the soil survey in the Tarko-Sale, Yamalo-Nenets
Autonomous Okrug, Russia [1]. The sampling area was approximately 6 km2. In total, 101 samples were collected.
Concentrations for two elements (Cr, Cu) were obtained by a precise chemical analysis. Copper was chosen as a
typical normally distributed pollutant to compare with chromium. The descriptive statistics of modeled elements are
shown in Table 1. The entire data set was randomly divided into two groups: 70% formed a training set for training
the ANN and building the kriging, the rest 30% formed the testing set for testing both, the network and kriging. The
ArcGIS application was used to predict the values in a test data set (31 samples). In order to accomplish this goal,
the ordinary kriging on the training data set (70 samples) was initially built. The predicted values in the test data set
were built by the function 'Prediction' in ArcGIS.
TABLE 1. Descriptive statistics of modeled elements
In order to assess the concentration of Cr in the training data set, a MLP with the Levenberg-Marquardt training
method [10] was used. The ANN was carried out in MATLAB. In our case, the input layer of MLP was compiled
with sampling points; the hidden layer included a few neurons, and the output layer represented the content of
elements in the relevant sample.
The selection of the number of neurons in the hidden layer was carried out by the lower total Root-mean-square
error (RMSE) of prediction of the pollutant (Cr, Cu) content for the training set (70 samples), test set (31 samples),
and complete set of data (101 samples). The number of neurons was varied from 2 to 20. Each network was trained
by 500 times and the best of them have been selected. Network education quality was checked by the correlation
coefficient and RMSE between the result of the network prediction and training data set. Results of the network
structure selection (number of neurons in the hidden layer) are shown in Fig. 1 (b, top) for Cr and in Fig. 1 (b,
bottom) for Cu. The following numbers of hidden neurons were selected: 10 for Cr and 9 for Cu.
FIGURE 1. a) A multilayer perceptron neural network; b) Root mean square error (RMSE) of a neural network (MPL) for test
(1), training (3) and overall (2) data under different neuron number in the hidden layer for Cr and Cu
The next step of the model was building kriging of residues, which are differences between ANN forecasts of
residues in testing points and estimates at the same points done by ordinary kriging. The final evaluation of the
pollutant content was obtained as the sum of the neural network evaluation and residues evaluation by kriging. So
that verify the method proposed in the study, a comparison with a stochastic interpolation method Universal Kriging
was carried out, then the accuracy of predictions were compared.
The accuracy assessment indices of predicted concentrations are shown in Table 2. The dependencies of
predicted concentrations vs real ones are presented in Fig. 2 (IDW - Inverse Weighted Distance).
TABLE 2. Descriptive statistics of modeled elements
FIGURE 2. . Predicted vs real concentrations for different models: a) Cr, b) Cu
A comparison of methods has shown the superiority of ANN in modeling accuracy for both elements. It was also
found that the use of a hybrid approach ANN-kriging gives an increase in the accuracy of prediction of concentration
distribution in the surface layer of soil, for chromium (about 1% relative to ANN, 42% relative to kriging, and 70%
relative to IDW), and copper (7% relative to ANN, 11% relative to kriging, and 9% relative to IDW). Contrary to
expectations, IDW technique showed higher accuracy of the model for copper, surpassing the geostatistical method.
However, the chromium-based kriging model showed a lower error than the IDW, which is likely due to the
presence of "spots" with abnormally high element content. Estimation of ANN residues by the ordinary kriging
allowed smoothing out the high and low values of concentrations of chromium and copper in the soil, which
improves the accuracy of prediction.
A study on the distribution of chromium and copper concentrations in the surface layer of soil at the urbanized
terrain of the Tarko-Sale town, Yamalo-Nenets Autonomous Okrug, Russia, was conducted. To simulate the
concentrations distribution, it was proposed to use a hybrid method that included simulation by the artificial neural
network and evaluation of ANN errors by ordinary kriging. A comparison of different approaches to the prediction
of the contaminants distribution in the surface layer of soil was carried out. A neural network type was MLP
(Levenberg-Marquardt training algorithm) with two input layer, one hidden layer and one output layer. 10 neurons
in the hidden layer have been selected for modeling the distribution of Cr, 9 neurons – for Cu. The results showed
that the ANN-based model was more accurate than the model based on the IDW and kriging.
Estimation of ANN prediction residues by ordinary kriging reduced ANN prediction errors, which increased the
accuracy of the chromium and copper concentrations distribution model in the surface layer of soil. In comparison
with other methods, the most significant improvement in RMSE index (40-70%) was observed in ANN-based
models when predicting chromium distribution, which concentration in the soil samples of study area forms spots
with abnormally high levels.
The obtained results confirm vast capabilities of hybrid ANN-kriging methods that can be utilized to improve the
accuracy of modeling the spatial distribution of the contaminants concentrations in the topsoil of urban areas, which
characterized by high heterogeneity.
1. I. Anagu, J. Ingwersen, J. Utermann, T. Streck, "Estimation of heavy metal sorption in German soils using artificial neural
networks", Geoderma 152, 104–112 (2009).
2. C. Bishop, Neural networks for pattern recognition (Clarendon, Oxford, 1995).
3. A. Falamaki, "Artificial neural network application for predicting soil distribution coefficient of nickel", Journal of
Environmental Radioactivity 115, 6–12 (2013).
4. O. S. Hilko, S. P. Kundas, and I. A. Gishkeluk, "Radionuclides migration modeling using artificial neural networks and
parallel computing", European water 39, 3–13 (2012).
5. Y. Li, C. Li, J.-J. Tao, L.-D. Wang, "Study on Spatial Distribution of Soil Heavy Metals in Huizhou City Based on BP-ANN
Modeling and GIS", Procedia Environmental Sciences 10, 1953–1960 (2011).
6. Z. H. Liu, Y. Chang, H. W. Chen, "Estimation of forest volume in Huzhong forest area based on RS, GIS and ANN (in
Chinese)", Chin. J. Appl. Ecol. 19, 1891–1896 (2008).
7. C. A. Schloeder, N. E. Zimmerman, M. J. Jacobs, "Comparison of methods for interpolating soil properties using limited
data", Soil Sci. Soc. Am. J. 65, 470-479 (2001).
8. A. P. Sergeev, E. M. Baglaeva, A. V. Shichkin, "Case of soil surface chromium anomaly of a northern urban territory –
preliminary results", Atmospheric Pollution Research 1, 44–49 (2010).
9. A. P. Sergeev, A. G. Buevich, A. N. Medvedev, I. E. Subbotina, M. V. Sergeeva, "Artificial neural network and kriging
interpolation for the chemical elements contents in the surface layer of soil on a background area", in 15th International
Multidisciplinary Scientific Geoconference SGEM 2015. Water Resources. Forest, Marine and Ocean Ecosystems,
Conference Proceedings, Book 3 Vol. 2 (STEF92 Technology Ltd., Albena, 2015), pp. 49–56.
10. A. J. Shepherd, Second-Order Methods for Neural Networks: Fast and Reliable Training Methods for Multi-Layer
Perceptrons (Springer-Verlag, 1997).
11. J.-B. Sirven, B. Bousquet, L. Canioni, L. Sarger, S. Tellier, M. Potin-Gautier, I. Le Hecho, "Qualitative and quantitative
investigation of chromium-polluted soils by laser-induced breakdown spectroscopy combined with neural networks analysis",
Anal Bioanal Chem. 385, 256–262 (2006).
12. L. Worsham, D. Markewitz, & N. Nibbelink, "Incorporating spatial dependence into estimates of soil carbon contents under
different land covers", Soil Sci. Am. J. 74, 635–646 (2010).
13. K.-O. Zeissler,
T. Hertwig,
Без категории
Размер файла
344 Кб
Пожаловаться на содержимое документа