Improvement of Coastal Sea-Level Altimetry Derived From GNSS SNR Measurements Using the SNR Forward Network and T-LSTM Anomaly Detection

Abstract

In the Global Navigation Satellite System reflectometry (GNSS-R), the spectral analysis approach is widely used to derive sea level height from signal-to-noise ratio (SNR) data because of its simplicity and ease of implementation. However, it requires many corrections to improve accuracy such as data quality control, outlier removal, and SNR bias correction. Moreover, the correction methods are normally post-processing, which are not suitable for real-time estimation. In this paper, we propose a novel method using a combination of two neural networks, including the SNR Forward Network and Time-aware Long Short-Term Memory Model Anomaly Detection (T-LSTM-AD), to improve the spectral analysis approach to be more accurate and closer to real-time estimation. SNR Forward Network is the special neural network designed based on the SNR physical model with consideration of the surface roughness term. It learns SNR biases from historical SNR data and is applied to new SNR data for SNR biases correction without using other external information. T-LSTM-AD is a new approach for outlier detection and dynamic surface modeling based on the temporal dependency of data. The trained T-LSTM-AD is used to classify the outlier and correct the dynamic surface for the new estimated sea level. To verify the performance, 1-year data from GTGU is separated to the training and testing data. The results of testing data with 6.3 cm RMSE and 0.939 correlation coefficient show that the proposed method has good performance. Furthermore, when applied to 1-year data, the proposed method provides accurate results compared to the existing methods.