Tsai Shih-chiao
Department of Environmental Information and Engineering, College of Science, NDU
Taiwan launched its first indigenous meteorological satellite, TRITON, in October 2023. It carries the domestically developed Global Navigation Satellite System-Reflectometry (GNSS-R) instrument, which provides data on sea surface wind speed, Mean Square Slope (MSS), and Significant Wave Height (SWH). This data offers crucial support for global weather observation and forecasting.
GNSS-R technology has the advantage of high timeliness, and its signals are less affected by severe weather. However, its wind speed retrieval accuracy decreases significantly under high wind speed conditions. The primary reason is that this technology relies on analyzing the power distribution of reflected signals across different delays and Doppler frequencies to create a Delay-Doppler Map (DDM). The average value within the specular point region of the DDM (DDMA) is then used as the basis for retrieval. A Geophysical Model Function (GMF) is then established between the DDMA and sea surface wind speed. When wind speed increases beyond a certain threshold, the DDMA’s sensitivity to wind speed decreases, leading to higher retrieval uncertainty.
This study used TRITON Level 1b data from November 2023 to September 2024 to design five combinations of feature parameters, including physical, geometric, spatiotemporal, instrument characteristics, and external inputs (ERA5’s MSS and SWH). We trained a Feedforward Neural Network (FNN) and a Long Short-Term Memory (LSTM) model, using CCMP 10-meter wind speed as the validation reference.
The results showed that the best performance was achieved when the input features included both ERA5’s MSS and SWH, with an RMSE of only 1.23 m/s. This was a clear improvement over combinations that used only physical, geometric, spatiotemporal, or instrument characteristic parameters. Overall, the FNN performed better than the LSTM.
Additionally, considering the difficulty of obtaining real-time MSS data, a method was tested where the model was trained using MSS converted from ERA5 wind speed and then tested with MSS converted from WRF-simulated wind speed, along with all other available TRITON parameters. This approach yielded an RMSE of 1.89 m/s for wind speed retrieval. Under high wind speed conditions, this method showed a near 1:1 consistency with CCMP, outperforming the TRITON’s original GMF method (RMSE = 2.820 m/s) and WRF simulation results (RMSE = 1.984 m/s). This highlights MSS as a key parameter for improving wind speed retrieval capabilities.
The MSS conversion method provides a practical and feasible alternative. As more TRITON data becomes available, the training dataset will be expanded, and input features and neural network structures will be optimized. This will ultimately enhance the reliability and application value of TRITON’s retrieval products under high wind speed and complex sea state conditions.Keywords: TRITON, GNSS-R, MSS, ERA5, WRF, FNN, LSTM, CCMP
