Jiménez et al. (2025) Automatic optical depth parametrization in radiative transfer model RTTOV v13 via LASSO-induced sparsity

Identification

• Journal: Geoscientific model development
• Year: 2025
• Date: 2025-11-14
• Authors: Felipe Jiménez, Juan Carlos De los Reyes
• DOI: 10.5194/gmd-18-8511-2025

Research Groups

• MODEMAT Research Center in Mathematical Modelling and Optimization, Quito, Ecuador
• Department of Mathematics, Escuela Politécnica Nacional, Quito, Ecuador

Short Summary

This study introduces a novel methodology for automatic and sparse parametrization of atmospheric optical depths within the Radiative Transfer for TOVS (RTTOV) version 13 model, utilizing statistical thresholds and LASSO regression. The approach significantly reduces computational costs and parameter counts while maintaining accuracy, making it highly efficient for satellite data assimilation.

Objective

• To develop and validate a novel methodology for automatically parameterizing atmospheric optical depths in the RTTOV v13 model by inducing sparsity through statistical thresholds and LASSO regression, aiming to reduce computational costs and parameter numbers without compromising accuracy.

Study Configuration

• Spatial Scale: Atmospheric profiles across 101 pressure levels (from top-of-atmosphere to surface pressure); satellite zenith angles from 0 to 63.61 degrees (6 path secant angles); Visible Infrared Imaging Radiometer Suite (VIIRS) M-bands (M7-M16) covering near, medium, and long infrared spectral ranges.
• Temporal Scale: Training data based on 83 vertical atmospheric profiles from the ECMWF83 database; application intended for operational satellite data assimilation systems requiring fast and continuous processing.

Methodology and Data

• Models used:
  • Radiative Transfer for TOVS (RTTOV) v13 (standard and proposed variants: SI, BIC+L1, L0+L1).
  • Least Absolute Shrinkage and Selection Operator (LASSO) regression.
  • Bilevel optimization (for optimal regularization parameter selection, using ℓ0-regression cost function or Bayesian Information Criterion (BIC)).
  • Ordinary Least Squares (OLS) regression (for SI variant and Post-LASSO).
  • Line-by-Line Radiative Transfer Model (LBLRTM) v12.15.1 (for generating reference optical depths and radiances).
• Data sources:
  • Satellite: Visible Infrared Imaging Radiometer Suite (VIIRS) M-bands (M7-M16) Spectral Response Functions (SRF J2).
  • Observation/Reanalysis: ECMWF83 database (83 vertical atmospheric profiles with temperature and gas concentrations across 101 pressure levels).
  • Line parameter database: AER Line Parameter Database v3.8.1 (primarily HITRAN 2016).
  • Continuum models: AER Continuum MT CKD v4.1.1.

Main Results

• The proposed methods (SI, BIC+L1, L0+L1) significantly increase sparsity in optical depth parametrization compared to standard RTTOV v13. For instance, L0+L1 (with ϵ1 = 10⁻⁶) reduced non-zero parameters from a range of 46.36 %–80.00 % (RTTOV v13) to 3.78 %–23.87 %.
• Computational runtime for evaluating parameterized transmittances was significantly reduced, with L0+L1 (ϵ1 = 10⁻⁶) achieving runtimes between 13.77 % and 41.39 % of standard RTTOV v13.
• Accuracy in approximating total transmittances showed minimal degradation, with Root Mean Square Error (RMSE) generally ranging from O(10⁻⁶) to O(10⁻⁵) for most channels, comparable to RTTOV v13.
• Average relative brightness temperature (BT) errors ranged from O(10⁻⁵) to O(10⁻⁴), with maximum relative errors from O(10⁻⁴) to O(10⁻³), demonstrating accuracy comparable to RTTOV v13, especially at lower statistical thresholds.
• For emissive bands (M11-M16), all proposed methods fully satisfied the instrument noise equivalent delta temperature (NEdT) criterion for 100% of profiles. For solar reflective bands (M7-M10), performance was comparable to RTTOV v13 at stricter statistical thresholds.
• The methodology automatically identified relevant gases, pressure levels, and predictors, and in many cases, eliminated the need for a correction term.

Contributions

• Introduction of a novel, automatic, and sparse optical depth parametrization method for RTTOV v13, combining statistical thresholding with LASSO regression and bilevel optimization.
• Demonstrated significant reduction in computational cost and parameter count (sparsity) for radiative transfer models, crucial for operational satellite data assimilation.
• Automated the selection of absorbing gases, relevant pressure levels, and key predictors, a process traditionally requiring expert manual intervention.
• First application of LASSO regression to the RTTOV model for automatic gas and parameter selection, enhancing efficiency without compromising accuracy.
• Provides a framework extensible to other Fast-RT models and satellite instruments, improving the efficiency and accuracy of atmospheric profile retrievals.

Funding

• Escuela Politécnica Nacional de Ecuador (award PIGR-22-01: "Asimilación de datos satelitales para el sistema de pronóstico meteorológico METEO: selección óptima de predictores y localización óptima de observaciones").
• PhD Program in Applied Mathematics, Escuela Politécnica Nacional de Ecuador (partial support for Franklin Vargas).

Citation

@article{Jiménez2025Automatic,
  author = {Jiménez, Felipe and Reyes, Juan Carlos De los},
  title = {Automatic optical depth parametrization in radiative transfer model RTTOV v13 via LASSO-induced sparsity},
  journal = {Geoscientific model development},
  year = {2025},
  doi = {10.5194/gmd-18-8511-2025},
  url = {https://doi.org/10.5194/gmd-18-8511-2025}
}