LightGBM-Driven Correction of Integer-Cycle Phase Biases in GNSS-IR for Robust Sea Level Retrieval

Sea level change is becoming increasingly complex in the context of global climate change. Accurate and reliable methods for monitoring water levels are essential for advancing the research of oceanic variations. The GNSS Interferometric Reflectometry (GNSS-IR) technique emerges as a complementary approach, leveraging signal-to-noise ratio (SNR) oscillations from reflected GNSS signals to estimate sea surface level. In GNSS-IR sea level retrieval, various factors including dynamic sea surface variations, surface roughness, and observation noise can introduce biases in reflector height (RH). These biases subsequently lead to phase errors in the SNR fitting process. A linear model is often employed to correct such phase deviations. The core challenge lies in the linear phase correction model’s inability to resolve phase deviations exceeding ±π, which introduces integer-cycle ambiguities. These errors, driven by the combined effects of multiple sources of uncertainty, result in RH residuals clustered around ±30–50 cm. To mitigate this issue, a two-stage correction method is developed: (1) a LightGBM (Light Gradient Boosting Machine) classifier identifies and corrects ±2π phase biases by analyzing SNR quality metrics (e.g., peak-to-noise ratio, full width at half maximum), environmental parameters (e.g., sea surface height change rate), and fitting residuals; (2) a sliding-window robust estimation refines RH values by dynamically compensating for residual outliers and tidal fluctuations. Validation across three coastal GNSS stations (SC02, CALC, FLCK) demonstrates significant improvements. The LightGBM model achieved 97.9~99.3% classification accuracy, effectively isolating and correcting 72–93% of ±2π deviations. Combined with robust estimation, the method reduced root mean square error (RMSE) by up to 48.6% compared to classical Lomb-Scargle Periodogram (LSP) results. Residual distributions transitioned from bimodal clusters to centralized peaks near zero, confirming the elimination of stratification artifacts.