A comparison of data-driven and physics-informed statistical learning models for atmospheric remote sensing applications
Abstract
Accurate models of gas radiation are required for the analysis of atmospheric remote sensing data. These models need to be CPU efficient in order to comply with the exponential growth of data provided by the most recent generations of satellite instruments. In this paper, we describe two methods to treat gaseous absorption in satellite-based remote sensing scenarios. The two approaches are founded on statistical/machine learning (SL/ML) strategies. Compared to most works that only focus on a single SL/ML model, a detailed comparison is provided in this paper between a physics-informed SL method founded on a double mathematical induction strategy, called the dependent ALD (Augmented L-Distribution) model, and a purely data driven ML model (DDML). Results of the two approaches are assessed against reference LBL calculations and compared with a Correlated K-Distribution (CKD) model both in terms of accuracy and CPU cost. The two SL/ML models are shown to reach an accuracy higher than 0.5% at a small fraction (three orders of magnitude below) of the CPU cost of high-resolution calculations. The two methods open new possibilities for fast though accurate forward modeling of satellite remote sensing data.
Related
- Statistical Downscaling of Daily Precipitation from ERA5 Atmospheric Fields to IMERG over West Africa
- Evaluating a candidate predictor beyond statistical significance: Atmospheric pressure and psychiatric day-care attendance as a worked example
- Atmospheric Predictability beyond 30 Days with Machine Learning
- The Equilibrium Response of Atmospheric Machine‐Learning Models to Uniform Sea Surface Temperature Warming
- SATFRONT: Detecting Atmospheric Fronts with Satellite Data using Machine Learning
BibTeX
@article{andre2026comparison,
title = {{A comparison of data-driven and physics-informed statistical learning models for atmospheric remote sensing applications}},
author = {Frédéric André and Xijiade Aihemaiti and Tao Ren and Loïc Guilmard and François Thieuleux and Nada Mourtaday and Antoine Rimboud and Somesh Roy and Céline Cornet and Laurent Labonnotte and Jérôme Riedi},
journal = {Journal of Quantitative Spectroscopy and Radiative Transfer},
year = {2026},
month = oct,
doi = {10.1016/j.jqsrt.2026.110011},
url = {https://hal.science/hal-05805159v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A comparison of data-driven and physics-informed statistical learning models for atmospheric remote sensing applications
AU - Frédéric André
AU - Xijiade Aihemaiti
AU - Tao Ren
AU - Loïc Guilmard
AU - François Thieuleux
AU - Nada Mourtaday
AU - Antoine Rimboud
AU - Somesh Roy
AU - Céline Cornet
AU - Laurent Labonnotte
AU - Jérôme Riedi
PY - 2026
DA - 2026/10/26
JO - Journal of Quantitative Spectroscopy and Radiative Transfer
DO - 10.1016/j.jqsrt.2026.110011
UR - https://hal.science/hal-05805159v1
AB - Accurate models of gas radiation are required for the analysis of atmospheric remote sensing data. These models need to be CPU efficient in order to comply with the exponential growth of data provided by the most recent generations of satellite instruments. In this paper, we describe two methods to treat gaseous absorption in satellite-based remote sensing scenarios. The two approaches are founded on statistical/machine learning (SL/ML) strategies. Compared to most works that only focus on a single SL/ML model, a detailed comparison is provided in this paper between a physics-informed SL method founded on a double mathematical induction strategy, called the dependent ALD (Augmented L-Distribution) model, and a purely data driven ML model (DDML). Results of the two approaches are assessed against reference LBL calculations and compared with a Correlated K-Distribution (CKD) model both in terms of accuracy and CPU cost. The two SL/ML models are shown to reach an accuracy higher than 0.5% at a small fraction (three orders of magnitude below) of the CPU cost of high-resolution calculations. The two methods open new possibilities for fast though accurate forward modeling of satellite remote sensing data.
SN - 0022-4073
SN - 1879-1352
ER -