Abstract:
The paper is devoted to the problem of modeling and forecasting the areas of thermokarst lakes, which are one of the main sources of methane emissions into the atmosphere, leading to global warming. Most approaches to modeling the area of lakes are based on linear models of the main climatic parameters available for objective measurement. However, this approach does not allow taking into account the dependence of the lake area on its previous values, that is, to take into account the dynamic characteristics of the lake evolution process. In this paper, an attempt is made to construct a dynamic model of the area with its subsequent validation on real data within the framework of a repeatable and reproducible computational experiment, which increases the generalizing ability of the constructed predictive model. The model is trained using the randomized machine learning method based on the maximum entropy estimation of the distributions of parameters and model noise, followed by their sampling and construction of an ensemble of predictive trajectories. The constructed model is trained and tested on real data from the Kola and North Yakut tundra. The results of the experiments show differences for these regions both in the average increase in the area of lakes and in the degree of dependence on climatic indicators and, due to their remoteness from each other, allows us to draw a conclusion about the difference in the processes of lake evolution and, consequently, about the processes that determine them.
Keywords:thermokarst lakes, Arctic, maximum entropy method, randomized forecasting.
Fulltext:
PDF file (816 kB)
First page:
PDF file