Detecting Sea Ice From TechDemoSat-1 Data Using Support Vector Machines With Feature Selection

Abstract

In this paper, a framework that employs support vector machines (SVMs) is proposed for the classification of TechDemoSat-1 delay-Doppler maps (DDMs), specifically, for separating DDMs of sea ice from those of seawater. DDM data were first operated with a general data preprocessing procedure, which included noise floor subtraction and normalization. In addition, a simple and effective feature selection (FS) was devised so that the input size was significantly reduced (from 128 × 20 to 20) while the classification accuracy was enhanced. To be specific, the feature was selected as the mean value along the delay-axis (128 in length) at each Doppler bin (20 in all). Here, expected classification labels (sea ice/seawater) were obtained based on reference sea ice concentration data collected by multiple passive microwave sensors. In practice, this trained SVMs-FS algorithm showed improved accuracy but less data storage demands than the existing neural networks (NNs)- and convolutional neural networks-based methods. In addition, the designed FS was also proven to be effective for both SVMs and NNs.