Image-Based Retrieval of All-Day Cloud Physical Parameters for FY4A/AGRI and Its Application Over the Tibetan Plateau

Satellite remote sensing serves as a crucial means to acquire cloud physical parameters. However, existing official cloud products from the advanced geostationary radiation imager (AGRI) onboard the Fengyun-4A geostationary satellite lack spatiotemporal continuity and important micro-physical properties. In this study, an image-based transfer learning ResUnet (TL-ResUnet) model was applied to realize all-day and high-precision retrieval of cloud physical parameters from AGRI thermal infrared measurements. Combining the observation advantages of geostationary and polar-orbiting satellites, the TL-ResUnet model was pre-trained with official cloud products from advanced Himawari imager (AHI) and transfer-trained with official cloud products from moderate resolution imaging spectroradiometer (MODIS), respectively. For comparison, a pixel-based transfer learning random forest (TL-RF) model was trained using the equally distributed data sets. Taking MODIS official products as the benchmarks, the TL-ResUnet model achieved an overall accuracy of 79.82% for identifying cloud phase and root mean squared errors of 1.99 km, 7.11 μm, and 12.87 for estimating cloud top height, cloud effective radius, and cloud optical thickness, outperforming the precision of AGRI and AHI official products. Compared to the TL-RF model, the TL-ResUnet model utilized the spatial information of clouds to significantly improve the retrieval performance and achieve more than a 6-fold increase in speed for single full-disk retrieval. Moreover, AGRI TL-ResUnet products with spatiotemporal continuity and high precision were used to accurately describe the spatial distribution characteristics of cloud fractions and cloud properties over the Tibetan Plateau, and provide the diurnal variation of cloud cover and cloud properties across different seasons for the first time.