Enhancing Geostationary Supercooled Water Cloud Identification Using Hyperspectral Infrared Sounder Measurements

Abstract

Supercooled water clouds (SWCs) play an important role in cloud radiative effects, precipitation processes, and aviation safety. Accurate SWC discrimination from passive geostationary observations remains difficult because SWCs are liquid clouds under subfreezing conditions and are influenced by atmospheric background conditions. Although reanalysis products such as ERA5 can provide useful atmospheric background fields, their latency limits near-real-time monitoring. To address this issue, this study investigates whether same-platform hyperspectral infrared observations from FY-4B/geostationary interferometric infrared sounder (GIIRS) can provide effective atmospheric-state constraints for SWC identification without relying on reanalysis inputs. A deep-learning framework, termed graph-based spectral ensemble network (GSE-Net), is developed by combining FY-4B/advanced geostationary radiation imager (AGRI) and GIIRS hyperspectral infrared observations for SWC identification. Within GSE-Net, spectral graph compressor (SGC) refines and compresses GIIRS spectra, while Bi-FiLM bidirectionally modulates spectral and AGRI scene features before fusion. Using CALIPSO as the reference, the proposed framework achieves a mean SWC $F1$ score of 89.33%, improving upon AGRI-only retrievals by 10.41 percentage points and outperforming AGRI+ERA5. In addition, the SGC further improves SWC identification performance over directly using uncompressed GIIRS observations. The results demonstrate that geostationary hyperspectral infrared observations can provide useful atmospheric-state constraints for SWC identification and serve as a practical alternative for near-real-time applications.