Section: Methodology and research methods. Models and forecasts
Snow cover is an informative indicator of atmospheric air pollution since it accumulates and preserves pollutants coming from the atmosphere as part of precipitation. Modeling the surface distribution of pollutant content in snow cover involves selecting predictors that maximize the predictive ability of the models. The purpose of this study was to compare multiple regression and multilayer perceptron models for mapping the dust surface distribution in the snow cover in the southern part of Novy Urengoy (Yamalo-Nenets Autonomous Okrug, Russia). The authors proposed applying to the selection of spatial variables a Land Use Regression (LUR) approach, which uses data on the relative positions of potential pollution sources and sampling sites to build a multiple linear regression model. To consider the nonlinear relationships between model predictors and dust concentration, a neural network model, a multilayer perceptron (MLP), was used. A total of four different models were tested: two LUR- and two MLP-based models. The MLP model with selected for the standard LUR model predictors and added coordinates of the sampling sites shows the best performance. The selected predictors contain spatial information about dust distribution. The added geographical coordinates made it possible to supplement the model with geostatistical information and improve its predictive ability. Finally, surface dust distribution maps were restored using four models and kriging. LUR and MLP models with spatial variables, which considered the location of potential pollution sources, produced dust distribution maps demonstrating the influence of these sources on surface distribution of dust accumulated in the snow cover.
Keywords: snow cover, dust content, surface distribution, mapping, land use regression, multilayer perceptron
Article published in number 4 for 2025 DOI: 10.25750/1995-4301-2025-4-037-043
ISSN 1995-4301
Select viewing options
