Tropospheric temperature and humidity retrievals from Meteosat Third Generation Imager (MTG-FCI) based on deep neural networks

The Meteosat Third Generation (MTG) Flexible Combined Imager (FCI) offers enhanced capabilities for monitoring atmospheric processes. Its higher spatial and temporal resolution, together with new spectral channels and advances in deep learning, enables opportunities to re-explore its potential for retrieving temperature and humidity profiles. Vertically resolved inversion of temperature and humidity is inherently ill-posed and becomes more challenging with imagers due to their limited infrared spectral resolution. Operational algorithms rely on numerical weather prediction background fields to guide the retrieval, reducing the independence and added value. Additionally, expanding data volumes also make optimal iterative inversion approaches computationally slow, limiting real-time applications. This study aims to investigate whether spatially aware deep learning models can extract frequent, spatially detailed, and forecast-independent profiles of tropospheric temperature and specific humidity that could complement traditional retrievals, be used for higher-level products or be more frequently assimilated in numerical weather prediction systems.

We developed and evaluated a deep learning framework to retrieve temperature and humidity profiles from FCI measured radiances. The model is based on a U-Net encoder–decoder architecture that exploits spatial context from 128x128 pixel image patches and integrates all 16 FCI spectral channels with ancillary surface and time variables. The network was trained on 14 months of collocated FCI observations as input and profiles of temperature and humidity across 15 pressure levels from CERRA reanalysis as target. Results were validated against temporally independent radiosonde measurements.

When evaluating retrievals and CERRA profiles against radiosondes, temperature retrievals remain within 1 K standard deviations (STDs) of CERRA, while derived relative humidity STDs remain below 5% of CERRA's over the whole profile. Performance degrades under cloudy conditions and at night, but the model preserves physically consistent vertical structures. Compared to above cloud retrievals, we measured a maximum STD increase of 0.3 K temperature and 7% relative humidity below clouds, despite limited direct radiative information. The retrieved profiles substantially improve upon a 30-year climatology baseline from ERA5 against radiosondes.

A permutation feature importance analysis reveals the spectral sensitivities learned by the model. The split window and CO2 infrared channels at 12.3 and 13.3µm have the strongest importance on both temperature and humidity profiles, as they measure both thermodynamic information combined from the cloud top, troposphere and surface. The water vapour channels around 6.3 and 7.3 µm, as expected, also show large importance for mid- and upper-tropospheric humidity. Visible and near-infrared channels despite not being used in traditional radiative transfer, show some sensitivities, with the 0.8 and 0.9µm bands contributing most to lower-tropospheric humidity during daytime.

The results demonstrate that deep learning models can extract potentially useful tropospheric profiles of temperature and humidity by combining radiance measurements with image-based context and learned climatological structures.