Integrating an SW-LSTM Model with GNSS-IR to Enhance Sea-Level Measurements during Storm Surges

Conventional tide gauges have limitations in extreme weather due to their susceptibility to damage from high waves and storm surges, as well as potential inaccuracies caused by rapid changes in wind speed and atmospheric pressure. To overcome these potential issues, global navigation satellite system-interferometric reflectometry (GNSS-IR) is adopted for detecting storm surges as a supplementary part of observation systems. The current GNSS-IR technique has insufficient accuracy and temporal resolution to capture high storm surge. In this study, we develop a sliding-window-based long short-term memory (SW-LSTM) framework to postprocess GNSS-IR retrieval results. Three scenarios of the SW-LSTM model are proposed in terms of the prior information involved as input features into long-short term memory (LSTM) networks. The models are validated in the inversion of sea levels in Quarry Bay of Hong Kong during typhoons Hato, Khanun, Mangkhut, Wipha, and Kompasu. The root-mean-square errors (RMSEs) in the GNSS-IR retrieval results are between 15.0 and 18.8 cm, and the skill scores vary between 0.963 and 0.979. The data during Mangkhut, Wipha, and Kompasu are adopted to train the SW-LSTM models. After postprocessing with the SW-LSTM models, the RMSE in the obtained sea levels is reduced to 6.8 cm for the super typhoon Hato and 6.5 cm for the severe typhoon Khanun. The skill scores are all above 0.990. Moreover, a comparison analysis shows that the SW-LSTM models outperform the least squares method (LSMs) and the cubic spline interpolation for GNSS-IR retrieval improvement. The SW-LSTM models show potential to extend GNSS-IR altimetry to real-time monitoring and synchronous predictions.