Sea ice remote sensing using spaceborne global navigation satellite system reflectometry

In this research, the application of spaceborne Global Navigation Satellite System- Reflectometry (GNSS-R) delay-Doppler maps (DDMs) for sea ice remote sensing is investigated. Firstly, a scheme is presented for detecting sea ice from TechDemoSat-1 (TDS-1) DDMs. Less spreading along delay and Doppler axes is observed in the DDMs of sea ice relative to those of seawater. This enables us to distinguish sea ice from seawater through studying the values of various DDM observables, which describe the extent of DDM spreading. Secondly, three machine learning-based methods, specifically neural networks (NNs), convolutional neural networks (CNNs) and support vector machine (SVM), are developed for detecting sea ice and retrieving sea ice concentration (SIC) from TDS-1 data. For these three methods, the architectures with different outputs (i.e. category labels and SIC values) are separately devised for sea ice detection (classification problem) and SIC retrieval (regression problem) purposes. In the training phase, different designs of input that include the cropped DDM (40-by-20), the full-size DDM (128-by-20), and the feature selection (FS) (1-by-20) are tested. The SIC data obtained by Nimbus-7 SMMR and DMSP SSM/I-SSMIS sensors are used as the target data, which are also regarded as ground-truth data in this work. In the experimental stage, CNN output resulted from inputting full-size DDM data shows better accuracy than that of the NN-based method. Besides, performance of both CNNs and NNs is enhanced with the cropped DDMs. It is found that when DDM data are adequately preprocessed CNNs and NNs share similar accuracy. Further comparison is made between NN and SVM with FS. The SVM algorithm demonstrates improved accuracy compared with the NN method. In addition, the designed FS is proven to be effective for both SVM- and NN-based approaches. Lastly, a reflectivity