Sea ice extent forecasting using statistical and deep learning models

Abstract

The subject matter of the article is the forecasting of time series of sea ice extent using statistical and deep learning methods. Sea ice extent is one of the most important indicators of climate change. Today, there are trends towards melting glaciers, which leads to a rise in sea level and, in turn, creates a significant threat of flooding of coastal regions around the globe. In addition, melting glaciers affect the flora and fauna of the Arctic and Antarctic regions, as well as economic stability in the world, covering economic development and food security. The spheres of agriculture, tourism, logistics are directly dependent on climate change, therefore, forecasting future changes is critically important for stability and sustainable development. The article analyzes the main trends in the change in sea ice extent. The goal of the study is to increase the reliability of long-term forecasting by designing a framework that covers the full forecasting cycle from data analysis to the use of predictive statistical methods and deep learning techniques. The tasks of the article are to conduct a comparative analysis of statistical methods and deep learning methods and their evaluation for the task of forecasting the area of sea ice distribution. The study used forecasting methods based on statistical models and deep learning. A study was conducted on the use of different approaches to forecasting future changes in a time series based on statistical methods, deep learning methods and ensemble models. The results obtained allow to evaluate the performance of models in the short term and an approach to long-term forecasting was formed. The use of autoregressors and deep learning methods is proposed to create a reliable long-term forecast. The comparison of the performance of the methods was carried out for the Northern and Southern Hemispheres. Conclusions. The scientific novelty of the results obtained is as follows: the method of forecasting time series of sea ice distribution using statistical methods and deep learning methods has been further developed. It was propose a generalizable forecasting framework that links time-series characteristics to model class selection and ensemble construction. The use of ensemble approaches allows us to ensure both the consideration of the main trends and the recognition of hidden patterns. The results obtained allow for a comprehensive assessment of time series for the Northern and Southern Hemispheres and indicate the feasibility of using both statistical forecasting methods for data with clearly defined patterns, such as the Arctic region, and deep learning methods to recognize hidden patterns observed in time series data for the Antarctic region.