Refinement of OLCI Water Vapor Retrieval Using GNSS Observations by Transformer Network in Chinese Mainland
Abstract
Atmospheric water vapor monitoring based on satellite observations has significantly contributed to achieving wide-area monitoring for global applications. However, the performance of water vapor from satellite observations is vulnerable to multiple factors under different conditions, making a significant potential for accuracy improvement. In this regard, a refined retrieving model based on the original observations of Ocean and Land Colour Imager (OLCI) and precipitable water vapor (PWV) from Global Navigation Satellite System (GNSS) was established. The empirical linear relationship between PWV and OLCI radiance observations were firstly applied for retrieval. Results showed that the root-mean-square errors (RMSE) between the model outputs and references from GNSS and ERA5 reanalysis were 3.8 mm and 4.2 mm, respectively. Afterward, three neural network architectures—LSTM, CNN, and Transformer—were implemented for better accuracy. Among them, the Transformer model achieved the best performance, with an RMSE of 2.51 mm relative to GNSS-PWV, representing a 48% improvement over the OLCI Level-2 product. Finally, atmospheric pressure and weighted mean temperature were incorporated in model refinement, which further improved the accuracy of model outputs to 1.9 mm. In general, this proposed method can greatly improve the accuracy of water vapor retrieval for OLCI, which is beneficial for meteorological and hydrological applications in Chinese mainland.