A New Method for Atmospheric Temperature and Humidity 3D Profile Inversion Based on FY-4A Satellite Data and Deep Learning

Traditional atmospheric temperature and humidity profile acquisition predominantly relies on in-situ observational networks (e.g., surface stations, radiosondes), which exhibit inherent constraints including sparse spatial coverage, uneven geographic distribution, and insufficient vertical resolution in upper atmospheric layers. Satellitebased profile reconstruction methodologies effectively mitigate these limitations through continuous spatial sampling capabilities. This study develops a machine learning framework utilizing FY-4A GIIRS hyperspectral data for advanced atmospheric profile inversion. The methodology comprises: 1) task-specific dataset construction through spatiotemporal alignment of satellite measurements, 2) PCA-driven spectral channel optimization with variance retention thresholds, and 3) implementation of a modified 3D UNET architecture integrating encoder-stage dense connectivity with depthwise separable convolutions and decoder-integrated self-attention mechanisms in skip connections. Validation against ERA5 reanalysis and independent radiosonde observations confirms enhanced retrieval accuracy, establishing novel benchmarks for operational atmospheric profiling systems.