Sea Ice Concentration Estimation From TechDemoSat-1 Data Using Support Vector Regression

In this paper, a framework that employs support vector regression (SVR) is proposed for estimating sea ice concentration (SIC) from TechDemoSat-l (TDS-l) delay-Doppler maps (DDMs). After a general DDM data preprocessing procedure, which includes noise floor subtraction and normalization, a simple and effective feature was extracted from DDMs. Specifically, the feature is selected as the mean value (incoherent averaging) of 128 delay bins at each Doppler pixel (20 in total). After training this proposed SVR algorithm using reference SIC data obtained from multiple passive microwave sensors, test results showed distinct improvement over those of the existing neural networks-and convolutional neural networks-based methods. Furthermore, the input size was significantly reduced (from 128×20 to 20) by using this proposed scheme.