Machine learning-based retrieval of total column water vapor over land using GMI-sensed passive microwave measurements

Abstract

ABSTRACT The Global Precipitation Measurement (GPM) Microwave Imager (GMI) is a microwave (MW) radiometer that has near-global coverage and frequent revisit time. To date, operational total column water vapor (TCWV) data records from the GPM GMI sensor have been exclusively offered over oceanic regions. It is challenging to retrieve TCWV over land from satellite MW measurements because of varying land surface characteristics. In this paper, a novel Light Gradient Boosting Machine-based retrieval algorithm is proposed to derive TCWV over land from GMI-sensed MW brightness temperature (BT) observations. The GMI-observed MW BT at 18.7 GHz and 23.8 GHz, differential BT between 18.7 GHz and 23.8 GHz, latitude, longitude, and month are selected and utilized as the input variables of the retrieval approach, because of their strong correlation with satellite-sensed MW TCWV retrievals. Instead of surface emissivity data or radiative transfer model, we take into account the spatial and temporal elements, namely latitude, longitude, and month. The training of the retrieval method is performed based on ground-based TCWV estimates from worldwide 4,471 Global Navigation Satellite System (GNSS) stations in 2017. The performance of the newly proposed retrieval algorithm is independently validated in a worldwide coverage using reference TCWV from additional 4,341 GNSS stations in 2018–2020 and 605 radiosonde stations in 2017–2020. The newly retrieved TCWV estimates over land have a correlation coefficient of 0.76 and 0.83, a root-mean-square error (RMSE) of 5.82 mm and 6.02 mm, a relative RMSE of 34.91% and 34.36%, and a mean bias of 0.02 mm and −0.42 mm compared to reference TCWV from GNSS and radiosonde, respectively. The performance of the retrieval algorithm is satisfactory when compared to that of land-purpose TCWV of other satellite missions, though we have not used either surface emissivity data or radiative transfer model. This result increases confidence in retrieving TCWV over land from satellite-sensed MW BT measurements based on machine learning using ground-based TCWV observations. The newly developed retrieval algorithm has the potential for integration into operational satellite missions or meteorological services, thereby enhancing weather forecasting, climate modeling, and other relevant applications.