Multi-satellite data fusion for upper tropospheric humidity using a novel deep learning architecture

Upper Tropospheric Humidity (UTH) plays a critical role in Earth’s radiative balance, hydrological cycle and climate feedbacks. However, long-term monitoring of UTH is hindered by sparse sampling, sensor heterogeneity and data gaps in satellite observations. This study presents a deep-learning-based multi-satellite data fusion framework to generate a global, daily gridded UTH climate record by integrating microwave humidity sounder observations from the National Oceanic and Atmospheric Administration (NOAA) and the European Organization for the Exploitation Meteorological Satellites (EUMETSAT) (Advanced Microwave Sounding Unit (AMSU)-B and Microwave Humidity Sounder (MHS); 1999–2021) with high-temporal-resolution tropical measurements from ISRO–CNES Megha-Tropiques/Sondeur Atmosphérique du Profil d’Humidité Intertropicale par Radiométrie (SAPHIR) (2012–2021). A hybrid architecture combining U-Net++ with a Variational Autoencoder (VAE) and attention-based fusion is developed to address sensor heterogeneity, asynchronous sampling, spatial gaps and uncertainty quantification. Sensor-specific quality weighting and physics-informed constraints guide the construction of fused training targets, with enhanced emphasis on SAPHIR in the tropics. The framework produces spatially complete, uncertainty-aware daily UTH fields at 1° × 1° resolution. Validation against 3,193 radiosonde observations from 23 stations demonstrates high accuracy (R ≈ 0.95, MAE ≈ 8.4%, RMSE ≈ 10.0%), with errors close to the intrinsic radiosonde uncertainty, indicating reliable multi-satellite fusion and faithful representation of large-scale UTH variability. The proposed hybrid U-Net++–VAE outperforms conventional weighted fusion and deterministic U-Net, particularly in data-sparse tropical regions and during strong convective events. The resulting long-term, high-resolution UTH dataset provides improved spatial continuity and uncertainty characterization, offering a robust observational constraint for climate monitoring, trend detection and model evaluation.