A deep learning approach to retrieve cold-season snow depth over Arctic sea ice from AMSR2 measurements

Abstract

Snowpack on sea ice can adjust changes in sea ice conditions and plays a vital role in the Earth's climate system. Snow depth, an important parameter of snowpack, is a necessary variable for retrieving sea ice thicknesses based on satellite altimeter data. Here, regression analysis (RA) is used to determine the best gradient ratio (GR) combination of brightness temperatures for estimating snow depths, and the RA model is proposed. Based on the RA model, one additional deep learning model is built, namely, the 5-variable long short-term memory (5VLSTM) model (or the RA-5VLSTM model). Meanwhile, an additional neural network model is built for comparisons, namely, the 5-variable genetic wavelet neural network (5VGWNN) model (or the RA-5VGWNN model). Using Operation IceBridge (OIB) and ice mass balance buoy (IMB) data, these three models, plus three existing algorithms, are compared to assess their performances in estimating snow depth. The results show that the RA-5VLSTM model performed pretty well among the six algorithms, with an RMSE of 7.16 cm. The RA-5VLSTM model, a robust approach, was less influenced by the uncertainty in the input data. From January to April during 2012–2020, the average monthly snow depth in the Beaufort Sea and the Chukchi Sea mainly showed a downward trend, while an upward trend was observed in the Central Arctic in most months. Variations in snow depth in the Central Arctic were mainly affected by the autumn 2-m air temperature (T2m) and the sea surface temperature (SST). Variations in snow depth in the Chukchi Sea were mainly affected by the autumn sea ice velocity.