A Convolution Neural Network-based Method for Sea Ice Remote Sensing using GNSS-R Data

Abstract

Sea ice remote sensing is of great significance to the understanding of polar climate change. At present, the global navigation satellite system reflector (GNSS-R) technology has been applied to the study of sea ice remote sensing and achieved good results. In this paper, a convolution neural network (CNN) based method for sea ice recognition (SIR) and estimation of sea ice concentration (SIC) using GNSS-R data is proposed. Specifically, a CNN model is designed to solve the classification problem of SIR and the regression problem of SIC estimation. In the stage of data set construction, first, the global GNSS-R data (TDS-l), in a certain period of time, is spatiotemporally matched with the relatively reliable sea ice data (NSIDC), and then the matched GNSS-R data is extracted to balance the amount of seawater data and sea ice data. In the stage of CNN model construction, the feature learning ability of the model is enhanced by adding convolution layer, pooling layer and full connection layer. Simulation results show that the proposed CNN -based scheme has a higher prediction accuracy of SIR and lower estimation error of SIC than other existing methods.