A key element of sea ice monitoring is sea ice concentration (SIC), and its temporal and spatial variability is crucial for researches in global climate change, navigational route planning and engineering projects in sea ice areas. The China-France Oceanography SATellite (CFOSAT) SCATterometer, namely CSCAT, by the character of its fan-beam rotary scanning system, can obtain multiple observing samples comprising rich information from different incidence angles and azimuth angles within a single grid, which makes it possible to retrieve SIC accurately. Considering that the quantitative relationship between the measuring elements of the scatterometer and SIC is not clear, this paper constructs a machine learning model using CSCAT backscattering coefficients and other observing elements to retrieve SIC. To generate a dataset for retrieving SIC, the microwave radiometer SIC products from the Ocean and Sea Ice Satellite Application Facility (OSI SAF) is first matched with the CSCAT backscattering coefficients. Following that, an SIC retrieval model is built using the eXtreme Gradient Boosting (XGBoost) algorithm based on CSCAT backscattering coefficients. The accuracy and real spatial distribution properties of the model outputs are then examined under various polar regions and seasons. The comparison between the Arctic and Antarctica reveals that the former has superior SIC estimations, while the result between different seasons indicates that the retrieval error is minimized during winter. There are variations in the model performance under different SICs, such as underestimation of the model results at high SICs and occasional misclassification of the grid as sea ice when it is completely covered with seawater. Overall, the findings of this study offer a new path for SIC retrieval, despite the low consistency of the results using scatterometer measurements with radiometers.