Bias correction of Arctic sea ice thickness products based on factor selection and machine learning methods

Abstract

With the accelerated melting of Arctic sea ice in the context of climate change, accurate sea ice thickness observations are needed for long-term climate simulation, short-term prediction, and shipping navigation. Limited by their special geographical location, in-situ data of Arctic sea ice thickness is scarce, and the quality of observation products from existing satellites and models is uneven. In this situation, we selected the key meteorological and marine environmental factors that affect the thickness of sea ice based on the thermodynamic law of sea ice, the traditional method of causal analysis, and the data mining ability of machine learning. Then, we constructed bias correction models combining these factors with geographic and temporal information, integrating random forest, extreme gradient boosting tree, and generalised regression neural network to correct the sea ice thickness data. The segmented regression models of thick and thin ice were constructed according to the characteristics of sea ice thickness products. The results show that the proposed bias correction model of sea ice thickness can effectively improve the quality of existing products and reduce the bias from the in-situ data.