Improving the Wet Refractivity Estimation Using the Extremely Learning Machine (ELM) Technique

Ehsan Forootan, Masood Dehvari, Saeed Farzaneh, Sedigheh Karimi

Publikation: Bidrag til tidsskriftTidsskriftartikelForskningpeer review

5 Downloads (Pure)


Constructing accurate models that provide information about water vapor content in the troposphere improves the reliability of numerical weather forecasts and the position accuracy of low-cost Global Navigation Satellite System (GNSS) receivers. However, developing models with high spatial-temporal resolution demands compact observational datasets in the regions of interest. Empirical models, such as the Global Pressure and Temperature 3 (GPT3w), have been constructed based on the monthly averaged outputs of numerical weather models. These models are based on the assimilation of existing measurements to provide estimations of atmospheric parameters. Therefore, their accuracy may be reduced over regions with a low resolution of radiosonde or continuous GNSS stations. By emerging and increasing the Low-Earth-Orbiting (LEO) satellites that measure atmospheric parameter profiles using the Radio Occultation (RO) technique, new opportunities have appeared to acquire high-resolution atmospheric observations at different altitudes. This study aims to apply these RO observations to improve the accuracy of the GPT3w model over Iran, which is sparse in terms of long-term GNSS and radiosonde measurements. The temperature, pressure, and water vapor pressure parameters from the GPT3w model have been used as the input layers of the Extremely Learning Machine (ELM) technique. The wet refractivity indices from the RO technique are considered target parameters in the output layer to train the ELM. The RO observations of 2007–2020 are applied for training, and those of 2020–2022 for evaluating the performance of the developed ELM. Our numerical results indicate that the developed ELM decreases the Root-Mean-Square Error (RMSE) values of the wet refractivity indices by about 17 percent, compared to the original GPT3w RMSE values. Additionally, the wet refractivity indices from ELM have revealed correlation coefficients of about 0.64, which is about 1.9 times those related to the original GPT3w model. The performance of ELM has also been examined by comparison with the data of six located radiosonde stations covering the year 2020. This comparison shows an improvement of about 14 percent in the average RMSE values of the estimated wet refractivity indices.
Udgave nummer1
StatusUdgivet - 4 jan. 2023


Dyk ned i forskningsemnerne om 'Improving the Wet Refractivity Estimation Using the Extremely Learning Machine (ELM) Technique'. Sammen danner de et unikt fingeraftryk.