Near Real-time GNSS-IR Sea Level Retrieval and Prediction Using Deep Learning
Abstract
Global Navigation Satellite System interferometric reflectometry (GNSS-IR) retrieves sea level from signal-to-noise ratio (SNR) data recorded by coastal GNSS stations, providing valuable data support for marine environmental monitoring. Although GNSS-IR has demonstrated centimeter-level accuracy in sea level retrieval, most existing approaches rely on postprocessing procedures for height variation error correction and outlier elimination, restricting their applicability in near real-time scenarios. Furthermore, accurate retrieval requires sufficiently long SNR sequences, which means that the real-time sea level retrievals at the current moment cannot be obtained. Consequently, it can only rely on sea level predictions to estimate real-time sea level. However, current prediction methods often fail to simultaneously achieve high accuracy and high-temporal resolution. To address these limitations, this study proposes a method that integrates Transformer and LSTM deep neural networks to perform real-time height variation error correction, outlier elimination, and sea level prediction, which enables near real-time sea level retrieval at a 6-min interval and further provides sea level prediction for up to 72 h. Four years of data from two stations are utilized for model training and two months for testing. At station SC02, the near real-time sea level retrieval achieves an RMSE of 3.78 cm, compared to 13.48 cm for the original observations. The RMSE of 2, 6, 24, and 72-h prediction reaches 5.04, 6.74, 7.19, and 12.75 cm, respectively. In addition, the real-time outlier elimination improves data usability to 98.63%, and the height variation error correction enhances inversion precision by 23.79%. The experimental results demonstrate that the proposed method effectively enables near real-time data processing and yields reliable sea level retrieval and prediction results.