FY satellite data and sounding data

Acquiring high-resolution 3D atmospheric humidity profiles remains a major challenge for improving numerical weather prediction and global environmental monitoring. This study proposes a deep learning–based method to retrieve 3D humidity profiles directly from FY-2F geostationary satellite infrared images. First, radiosonde data validated the reliability of the official FY-2F humidity product as training targets. We then developed an end-to-end CNN-Transformer hybrid model that takes dual-channel IR2 and IR3 images as input and outputs 3D relative humidity fields across six pressure levels (1000–400 hPa). Results show excellent performance in the mid-to-upper atmosphere (700–400 hPa), with an average correlation coefficient >0.93 and R² >0.84, while performance is limited in the lower atmosphere (e.g., R² = 0.35 at 925 hPa) due to satellite sensitivity constraints and near-surface complexity. The model maintains low error (MAE 0.01–0.05%) and high computational efficiency, demonstrating its potential for operational, near-real-time humidity monitoring within Digital Earth frameworks.