In this study, an improved algorithm to retrieve atmospheric winds from satellite temperature soundings over the polar region is developed. The algorithm utilizes a thermal wind relationship plus a planetary boundary layer parameterization to represent the basic atmospheric dynamics over the polar regions. With given temperature profiles, a first-guess wind profile is derived from the basic dynamics. Lagrange multipliers in a variational formalism are then used to force the first-guess wind to conserve mass. To estimate the algorithm accuracy, wind retrievals based on temperature profiles from the Global Data Assimilation System (GDAS) are compared with the wind observations from 13 radiosonde stations around the Antarctic coast. The results demonstrate that the wind retrievals exhibit mean absolute errors about 1 m sec-1 for the 13 stations in the troposphere and the root-mean-square errors are greater than the GDAS assimilated winds by about 0.5 m sec-1.