PaperlayerView exact revision on arXiv

Source-linked AI summary

Tropospheric temperature and humidity profile retrieval from Meteosat Flexible Combined Imager based on deep learning

Alejandro Salgueiro, Johannes Rausch, Julie Thérèse Villinger, Angela Meyer

arXiv:2608.25700v1cs.LGphysics.ao-ph

TL;DR

Broadband imagers provide limited vertical information, making independent all-sky temperature and humidity profiling difficult without NWP background fields. This paper develops a spatially aware deep learning framework using MTG-FCI observations, achieving statistically reliable forecast-independent profiles validated against radiosondes.

  • Problem

    Limited infrared resolution makes imager-based vertical profile retrieval underdetermined and ill-posed, typically requiring NWP priors that reduce retrieval independence.

  • Method

    A spatially aware supervised deep learning model retrieves all-sky temperature and humidity profiles from spatial patches using FCI’s spectral observations.

  • Results

    Temperature biases of -0.3 to 0.1 K and standard deviations of 1.5 to 1.9 K were achieved against independent radiosondes without NWP forecast backgrounds.

  • Takeaways & Limitations

    The framework provides statistically reliable, forecast-independent all-sky tropospheric profiles from MTG-FCI observations for high-resolution atmospheric monitoring.

  • Takeaways & Limitations

    Retrieved fields are spatially smoother than CERRA, with gradients systematically smoothed at horizontal scales of 17 to 550 km.

Abstract

from arXiv · show

The Meteosat Third Generation (MTG) Flexible Combined Imager (FCI) offers new opportunities for tropospheric temperature and humidity profiling, at higher spatio-temporal resolutions and expanded spectral coverage relative to its predecessor. Vertically resolved retrievals from broadband imagers are inherently challenging, and operational retrieval algorithms typically rely on numerical weather prediction (NWP) background fields to compensate for limited infrared spectral resolution, reducing the retrievals' independence. We develop a spatially aware deep learning framework to retrieve all-sky tropospheric temperature and humidity profiles from FCI, without forecast profiles as input. A Residual U-Net that exploits spatial context across all 16 FCI channels was trained on 14 months of collocated FCI observations and CERRA reanalysis targets over Europe. Validated against independent radiosondes, retrieved temperatures show biases below 0.4 K and standard deviations of 1.5-1.9 K. Retrieved relative humidity standard deviations range from 12-20 %, compared to 9-19 % for CERRA. Performance degrades modestly under clouds, with standard deviation increases below 0.4 K and 3 % RH beneath cloud tops despite limited direct radiative information. Ablation experiments show that spatial context improves retrievals, with the largest gains below cloud tops. Feature sensitivity analysis indicates broad consistency with FCI bands' established radiative transfer characteristics. Visible and near-infrared channels contribute despite not being commonly used in physics-based profile inversions. These results demonstrate that spatially aware deep learning models can extract statistically reliable tropospheric profiles from geostationary imager observations, independent of NWP forecast fields, enabling more rapid autonomous monitoring of the atmosphere.

Key points

A spatially aware U-Net retrieves forecast-independent all-sky tropospheric temperature and humidity profiles from MTG-FCI. Radiosonde validation in Europe shows temperature standard deviations below 2.0 K and relative-humidity errors within 20%, while short-wave infrared and near-infrared bands add retrieval value.

  • A spatially aware U-Net retrieves forecast-independent all-sky tropospheric temperature and humidity profiles from MTG-FCI.
  • Temperature STDs below 2.0 K and relative humidity within 20 % were achieved for all-sky retrievals validated against European radiosondes.
  • Short-wave infrared and near-infrared bands added value by constraining the data-driven retrievals.

Plain language summary · 1 Introduction

This study develops a fully data-driven, spatially aware deep learning framework for forecast-independent retrieval of all-sky tropospheric temperature and humidity profiles from FCI observations. It evaluates retrieval accuracy and examines whether learned channel sensitivities accord with radiative-transfer expectations, including contributions from solar channels.

  • 1 Introduction: Satellite observations provide global, frequent, and consistent atmospheric information that supports NWP assimilation and nowcasting applications.Radiances may be assimilated directly, while retrieved profiles can also be assimilated or used for nowcasting.
  • 1 Introduction: Physics-based profile retrievals solve radiative-transfer equations iteratively, typically using optimal-estimation frameworks constrained by NWP prior information.This dependence motivates methods that can provide forecast-independent estimates.
  • 1 Introduction: Broad imager bands average radiation over thick atmospheric layers, limiting vertical resolving power and making detailed profile retrievals physically and mathematically challenging.Compared with hyperspectral sounders, imagers have fewer infrared bands with wider spectral resolution and highly interdependent observations.
  • 1 Introduction: Data-driven methods can address these limitations, with EUMETSAT’s operational PWLR method providing forecast-independent all-sky profile estimates from IASI.PWLR complements physics-based inversions and has been used operationally for more than a decade.
  • 1 Introduction: The study expects larger spatial context to compensate for missing below-cloud radiative information, capture regional thermodynamic structure and climatology, and address slanted viewing geometry.The motivation also includes naturally incorporating visible and near-infrared information into deep-learning retrievals.
  • 1 Introduction: The proposed framework retrieves all-sky three-dimensional tropospheric temperature and humidity fields from FCI observations using a fully data-driven, spatially aware approach.The work develops and evaluates the framework in a forecast-independent manner.
  • 1 Introduction: Forecast independence means that no prior profiles guide retrieval during inference, while feature-importance analysis assesses individual spectral-channel contributions across vertical levels.The study explicitly investigates retrieval accuracy and learned channel sensitivities.
  • 1 Introduction: The research questions test retrieval accuracy, consistency of learned channel sensitivities with radiative-transfer expectations, and measurable daytime skill from solar channels.These questions define the evaluation scope for the FCI-based framework.

2 Data and pre-processing

The study combines 10-minute, 16-channel MTG FCI observations with CERRA temperature and humidity profiles, ancillary physical variables, and independently quality-controlled radiosonde measurements. Pre-processing harmonizes spatial resolution, vertical levels, cloud conditions, and temporally representative train–validation–test splits for all-sky retrievals.

  • FCI observations: FCI provides hemispheric observations every 10 minutes across 16 channels—3 visible, 5 near-infrared, and 8 infrared—with 0.5–2 km nadir resolutions.Visible and near-infrared radiances are converted to reflectance, while infrared radiances are converted to brightness temperature.
  • Retrieval targets: CERRA supplies tropospheric air-temperature and specific-humidity profiles at 5.5 km horizontal resolution over Europe, represented on 15 coarser pressure levels.Short forecasts at +1 and +2 hours provide hourly estimates between three-hourly reanalysis fields; CERRA relative humidity is converted to specific humidity using Murphy and Koop parametrisations.
  • Evaluation and ancillary data: Independent NOAA IGRA V2.2 radiosonde profiles are quality-controlled and log-linearly interpolated to the same 15 CERRA pressure levels for evaluation.Ancillary inputs include ASTER surface elevation, topographically corrected CERRA surface pressure, and observation-specific surface sun elevation angle.
  • Spatial processing: FCI channels at 1 km resolution are downsampled by nonoverlapping 2×2 block averaging to match the 2 km infrared resolution.The study domain excludes satellite zenith angles of 75° or greater; within the retained domain, pixel areas range from 5.56 to 19.10 km².
  • Cloud treatment: A single all-sky model reconstructs complete temperature and humidity profiles in both clear and cloudy scenes, with performance assessed separately above and below cloud tops.The approach uses measurements above and around clouds, partially clouded scenes, and thermodynamic structures learned from CERRA to approximate conditions beneath cloud layers.
  • Dataset splitting: A rolling split avoids seasonal domain mismatch in the 14-month collocated dataset, using repeating 21-day training, 3.5-day validation, and 7-day testing periods.A secondary independent test set spans 1 December 2025 to 28 February 2026 and is separated from training by a 7-day unused-data buffer.

3 Methods

The method frames retrieval as supervised image-to-image regression on spatial patches, using a residual U-Net to predict temperature and humidity profiles from FCI observations and ancillary variables. Ablations, radiosonde-based evaluation, sharpness analysis, and channel substitution probe spatial context, reliability, and feature use.

  • Each 128×128-pixel input patch contains 26 channels, producing temperature and specific-humidity profiles across 15 pressure levels.The inputs comprise 16 FCI spectral bands and 10 ancillary variables, while both output tensors have 15 channels.
  • The main model is a Residual U-Net augmented with squeeze-and-excitation channel attention and a dilated multi-scale bottleneck.Its encoder uses residual convolutional blocks, attention, and pooling, with learnable skip connections supporting gradient flow.
  • A pixel-wise 1x1-Net baseline isolates the contribution of spatial context by predicting independently from each pixel’s spectral channels.Because it lacks neighbouring information, its performance gap relative to the U-Net also reflects sensitivity to geometric parallax displacement.
  • Training minimises mean squared error on standardised temperature and specific-humidity profiles, equalising their contributions despite differing physical scales.The models use batches of randomly sampled patches, while alternative sharpness-preserving losses are reserved for future work.
  • Evaluation uses independent radiosondes for bias and residual standard deviation, supplements point metrics with spatial sharpness analysis against CERRA, and tests channel sensitivity by cross-sample substitution.Relative humidity derived from the retrieved profiles is constrained to 0–100 %, while substitution preserves marginal distributions but disrupts scene-level physical consistency.
  • Cross-sample channel substitution can underestimate a channel’s contribution when correlated inputs allow the model to compensate with alternative channels.The passage identifies SHAP as more robust for correlated inputs but computationally more demanding and harder to interpret for three-dimensional outputs.

4 Results and discussion · 4.1 Statistical reliability of the retrieved profiles

The FCI U-Net retrieves statistically reliable all-sky tropospheric profiles without NWP background fields, outperforming climatological and single-pixel baselines while approaching CERRA accuracy. Skill varies modestly with clouds, illumination, season, and temporal domain, with humidity more cloud-sensitive and near-surface generalisation weakest.

  • 4.1.1 All-sky retrieval accuracy: Shared positive humidity-bias structures between retrieved profiles and CERRA indicate inherited training-source structure rather than retrieval-specific error.This interpretation is supported by direct retrieval-versus-CERRA checks and motivates refinement of the moist bias.
  • 4.1.1 All-sky retrieval accuracy: Temperature skill improves monotonically across the information hierarchy, with the U-Net outperforming climatological and ancillary-only baselines and benefiting from spatial context.The climatology has a -1.2 to -0.5 K cold bias, while climatology and ancillary-only STDs reach 4.5 to 4.7 K; spatial context adds 0.7 to 1.3 K.
  • 4.1.1 All-sky retrieval accuracy: 12 to 20 % U-Net relative-humidity STDs compare with 14.5 to 23.3 % for the 1x1-Net and 9 to 19 % for CERRA.The climatology and ancillary-only model have STDs of 16 to 31 %, while all products show a positive humidity bias increasing below 700 hPa to approximately 24 %.
  • 4.1.2 Above vs. below cloud tops: Below-cloud temperature STDs remain within 1.5 to 2.0 K and biases below 0.5 K, while below-cloud humidity errors are generally larger than above-cloud errors.Relative-humidity STDs are 12 to 19 % above cloud and 11 to 23 % below cloud through the mid-troposphere; biases reach approximately 23 % and 33 %, respectively, below 700 hPa.
  • 4.1.2 Above vs. below cloud tops: Spatial context reduces the 1x1-Net’s cloud-regime relative-humidity errors by 0.2 to 4.6 % below cloud and up to 3.5 % above cloud.Regime differences persist against CERRA, indicating effects beyond inherited training-source structure.
  • 4.1.3 Seasonal and diurnal variability: Daytime retrievals outperform nighttime, with temperature STDs lower by 0.5 to 1.2 K and relative-humidity STDs by 1 to 5 %, although seasonal and CERRA effects contribute.Temperature biases remain below 0.5 K in both illumination conditions, and uneven seasonal sampling may favour SON results, which has 2297 hourly training scenes.
  • 4.1.4 Temporally independent generalisation: 400 to 900 hPa skill remains closely comparable across test periods, while near-surface temperature STDs increase by 0.2 to 0.9 K and specific-humidity STDs by 0.01 to 0.12 g/kg.Smaller temperature degradation reaches up to 0.27 K at 400-300 hPa, with warmer and moister near-surface biases in the second period.
  • 4.1.4 Temporally independent generalisation: The localized near-surface degradation indicates limited meteorological representativeness rather than uniform synoptic-event overfitting.The second period contains more cold and dry near-surface observations below −10 °C and 1 g/kg at 950 and 1000 hPa, conditions under-represented in training.

4.2 Spatial structure and resolution of retrieved fields

The retrievals reproduce large-scale tropospheric structures across pressure levels and variables, while remaining smoother than CERRA at smaller scales. Spectral and spatial-error analyses identify reduced mesoscale variability and viewing-angle-related degradation as key limitations.

  • Spatial structure: Large-scale temperature and specific-humidity structures are properly reproduced at both 300 and 900 hPa, including 900 hPa humidity despite weak infrared water-vapour sensitivity.The comparison uses retrieved fields, CERRA, FCI’s 7.3 µm water-vapour brightness temperature, and cloud-top pressure.
  • Spatial structure: Retrieved fields are smoother than CERRA, with less pronounced gradients, particularly where cloud cover blocks radiative information below cloud top.Broad vertical weighting functions also limit the representation of sharp gradients.
  • Spatial structure: Finer structures emerge in zoomed fields and appear to inherit smaller-scale patterns from FCI measurements.The retrieved fields show local similarities to brightness-temperature patterns near cloud borders.
  • Spectral resolution: ∼10 −20 pixels (∼28 −55 km) marks the approximate scale down to which CERRA carries increasingly more spectral energy than the retrieval.Synoptic-scale structures are broadly captured, whereas mesoscale gradients are smoothed by limited FCI spectral information and MSE attenuation.
  • Spatial error distribution: Increasing satellite zenith angle likely raises lower-tropospheric temperature RMSE toward the northeastern domain edge, while humidity RMSE shows a weaker viewing-angle signature.At 300 hPa, no viewing-angle gradient is apparent in the humidity field.

4.3 Channel contributions to retrieval performance

Feature sensitivities are broadly consistent with radiative-transfer expectations, with 12.3 and 13.3 µm channels especially important for cloud-top characterization in all-sky retrievals. The 10.5 µm clean window provides a baseline, while visible, near-infrared, and ancillary inputs make additional contributions.

  • Analysis approach: The feature-importance analysis evaluates retrieval sensitivity as the RMSE increase caused by substituting individual FCI channels or ancillary inputs.Sensitivities are interpreted using radiative-transfer theory and comparable-channel weighting functions.
  • Cloud-top and infrared contributions: The 12.3 and 13.3 µm channels dominate temperature and relative-humidity sensitivities because they help characterize cloud tops, the effective lower boundary in all-sky retrievals.Their combined information is essential for moisture correction and for semi-transparent or partially clouded pixels.
  • Cloud-top and infrared contributions: The 10.5 µm clean window contributes across the full profile and serves as a brightness-temperature baseline for differencing moisture-affected channels.Other window bands can partly compensate when the 10.5 µm channel is shuffled.
  • Visible and near-infrared contributions: The 2.2, 1.6, and 1.3 µm near-infrared bands show relatively small contributions, with 1.3 µm contributing least to moisture retrieval because of strong water-vapour absorption.The 0.8/0.9 µm pair and 3.8 µm band nevertheless show strong sensitivity.
  • Ancillary-input contributions: Satellite zenith angle is the most relevant ancillary input, with sensitivity across the profile that peaks at upper levels, while surface pressure contributes modestly to lower-to-mid-tropospheric retrievals.Sun elevation, surface elevation, latitude, and longitude show negligible sensitivity.

4.4 Comparison to other retrievals

The FCI retrieval achieves error margins broadly comparable to established IASI and ABI products despite using only 16 broadband channels and no NWP background fields. Comparisons remain contextual because the instruments, datasets, resolutions, and evaluation methodologies differ substantially.

  • Comparative performance: Our FCI retrieval achieves error margins comparable to IASI and ABI, although instruments and evaluation methodologies differ substantially.The comparison provides order-of-magnitude context rather than a strictly controlled benchmark.
  • Instrument and dataset differences: 8,461 spectral channels give IASI substantially more atmospheric information than the 16 channels used by the FCI retrieval.IASI validation also uses approximately 12 km spatial resolution and different collocation criteria.
  • Instrument and dataset differences: The ABI retrieval uses short-range GFS NWP forecasts as a background prior and is restricted to clear-sky conditions, unlike the FCI retrieval.The FCI framework operates in all-sky conditions without prior NWP information, exposing it to cloud attenuation and the absence of a guided background state.
  • Comparative performance: Despite these constraints, FCI retrievals show statistical reliability somewhat comparable to IASI and ABI using only 16 broadband channels and no NWP forecasts.The higher relative humidity biases than ABI are consistent with the moist bias inherited from the reanalysis.

5 Conclusions

The study demonstrates that a spatially aware, fully data-driven deep learning framework retrieves statistically reliable all-sky tropospheric temperature and humidity profiles from MTG-FCI without NWP forecast backgrounds. The retrieval reproduces large-scale structures, recovers fine-scale variance, and provides fast, high-resolution atmospheric monitoring.

  • Performance: -0.3 to 0.1 K temperature biases and 1.5 to 1.9 K standard deviations were achieved using 14 months of FCI observations and CERRA profiles without NWP forecast backgrounds.The framework retrieves all-sky temperature and humidity profiles from MTG-FCI using a U-Net-based model.
  • Spatial structure: 17 to 550 km horizontal scales show systematically smoothed gradients, while fine-scale variance below CERRA’s effective resolution is recovered from FCI’s 2 km nadir observations.Spatial and power spectra confirm that large-scale temperature and humidity structures are well reproduced.
  • Limitations: Higher satellite zenith angles increase retrieval errors because oblique viewing amplifies atmospheric path lengths, with additional degradation over high-elevation areas.These effects identify viewing geometry and terrain as limitations of the retrieval.
  • Feature sensitivity: The 12.3 and 13.3 µm channels dominate sensitivity throughout the column, while the 0.9/0.8 µm pair and 3.8 µm band also contribute.The sensitivity patterns broadly match known FCI radiative transfer characteristics, with cloud-top characterisation acting as the effective lower boundary condition.
  • Implications: MTG-FCI’s spatial and spectral information enables fast, forecast-independent temperature and humidity profiles at high spatial and temporal resolution.The results support more frequent and autonomous tropospheric monitoring and may complement sounder-based retrievals.

6 CRediT authorship contribution statement

The authors divided contributions across drafting, visualization, validation, methodology, investigation, analysis, data curation, conceptualization, supervision, resources, funding, and review/editing.

  • Alejandro Salgueiro led the original draft, visualization, validation, methodology, investigation, formal analysis, data curation, and conceptualization.
  • Angela Meyer contributed review and editing, supervision, resources, methodology, funding acquisition, and conceptualization.
  • Johannes Rausch and Julie Th´er`ese Villinger contributed review and editing, methodology, and conceptualization.
Loading 2608.25700v1…