A New Deep-Learning-Based Framework for Ice Water Path Retrieval From Microwave Humidity Sounder-II Aboard FengYun-3D Satellite

The derivation of ice water path (IWP) from microwave (MW) radiometer measurements is challenging. This study presents a deep learning framework for global retrieval of IWP using observations from the Microwave Humidity Sounder-II (MWHS-II) aboard the FengYun-3D (FY-3D) satellites. Two deep learning models, deep forest (DF21) and quantile regression neural network (QRNN), are constructed to detect ice cloud flags and retrieve IWP. By collocating MWHS-II observations with 2C-ICE, a joint product of CloudSat and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO), deep learning models learn the characteristics of IWP from MWHS-II brightness temperatures (BTs). The test results show that the MWHS-II channels provide more information on IWP than the MWHS channels, particularly the 89-GHz channel and the 118-GHz channels with an offset of
$\geq 0.8$
GHz. Combining the QRNN and DF21 models, the IWP retrieval results in a root mean square error (RMSE) of 707.346 g/m2, mean absolute percentage error (MAPE) of 65.122%, mean bias error (MBE) of −104 g/m2, determination coefficient (
$R^{2}$
) of 0.683, and Pearson correlation coefficient (PCC) of 0.831. Application of the models to MWHS-II observations of tropical cyclone CILIDA shows better agreement with 2C-ICE. All the datasets exhibit a similar feature on the monthly mean scale, but the magnitudes of IWP differ. Compared with GMI-GPROF, MODIS, and ERA5 IWP products, MWHS-II results are closest to 2C-ICE. Similar results are also shown for the zonal mean data. These results show that deep learning methods efficiently and probabilistically retrieve IWP from long-term observation data of MWHS/MWHS-II.