Study of variability of the Arctic sea ice thickness (1982–2003)

During the past decade, the scientific community has been concerned about a considerable change in the age structure, a decrease in the summer areal minimum, and thinning of the Arctic sea ice, which represents the most significant component of Arctic ecosystems and climate. Individual aspects of this problem have been considered in a series of papers [1‐6] and different models have been elaborated. However, results of the simulation of variations in the sea ice cover, particularly, its thinning, are often contradictory [7]. We investigated the long-term variability in the Arctic sea ice thickness based on artificial neural networks (ANN) [3], tracking of the prehistory of each sea ice pixel in the satellite data, and assessment of geophysical parameters from reanalysis data (NCEP/NCAR Reanalysis, http://www.cdc.noaa.gov), which allow empirical simulation of long-term variability of the sea ice thickness. To assess the sea ice thickness, we used the following databases (i) sea ice concentrations and motion vectors based on satellite measurements [3]; (ii) surface air temperature based on the buoy network [8] and estimates of radiation flux (NCEP/NCAR). The ANN training was based on sea ice thickness measured by submarines [9] and polar sea ice survey data [10] reduced to the scanning format with a resolution of 25 × 25 km in the polar stereographic projection (without correction for hummocking). Assessment of the sea ice thickness was based on the assumption that the monthly thickness of a sea ice pixel for the current year depends on the cumulative effect of basic geophysical parameters that govern the energy exchange between sea ice, ocean, and atmosphere during the whole pixel prehistory [4]. The investigation showed that seven out of twenty-two parameters make up a set of ANN inputs sufficient for approximation: monthly mean estimates of clear sky downward solar and longwave fluxes; monthly mean estimates of net shortwave and longwave radiation fluxes; sum of monthly mean air temperatures below ‐2 ° C (or 0 if the temperature is above ‐2 ° C); the sea ice drift rate; and the index of sea ice divergence‐convergence. The index was determined based on sea ice drift vectors [4]. The optimal topology of the neural network turned out to be 7-12-1. The mean discrepancy between ANN-based annual average sea ice thickness and training data made up +0.04 m (RMS = 0.19 m, 1982‐1999). The standard deviation based on the ANN data was 29 ± 19% lower relative to the training data. The monthly mean thickness based on the ANN data differed from training data by +0.02 m (RMS = 0.16 m, February‐October). The