Accounting for Radiative Transfer Model Errors in Physics-Guided VAEs for Biophysical Variable Retrieval
Abstract
Monitoring essential climate variables (ECV) from optical remote sensing remains challenging due to the ill-posed nature of biophysical variable retrieval and the limitations of radiative transfer models (RTMs). The recently proposed PROSAIL-VAE framework addresses some of these issues by embedding physical modeling within a variational autoencoder (VAE). Unaddressed RTM deficiencies -- including oversimplified soil spectra, un-modeled canopy clumping, and signal saturation -- introduce systematic residuals that are assumed negligible in most hybrid retrieval frameworks. We adapt a residual physics-guided VAE that explicitly accounts for RTM shortcomings via a refinement network and extended latent space, constrained to balance physical and data-driven contributions. The framework has two outputs: improved retrieval and learned residuals that can be interpreted in physical terms. On two independent in-situ datasets, retrieval improvements over the operational BVNET processor (Leaf Area Index (LAI) RMSE 19% and 5.2%; Canopy Chlorophyll Content (CCC) 47.9%) demonstrate the behavior of the framework. A non-trivial CCC degradation of 7.7% relative to the baseline PROSAIL-VAE emerges as a trade-off of reconstruction-driven optimization, where chlorophyll's weak spectral signature becomes entangled with structural and soil parameters. We further show that this degradation is partially mitigated by the extended latent space and decorrelation constraints. Spectral and spatial analysis of the learned residuals reveals that they concentrate in regimes where PROSAIL is known to be deficient -- particularly soil reflectance modeling and signal saturation -- suggesting that residual modeling can serve both as a retrieval enhancement and as a diagnostic probe of model inadequacy, while acknowledging that a fully validated diagnostic framework remains an open challenge.
Related
- Radiative Transfer Modelling of Ka-band Microwave Interactions with Snow-Covered Freshwater Ice: Implications for SWOT KaRIn Property Retrieval
- Separation of Source, Site, and Path Effects Based on Elastic Radiative Transfer Theory. Part 1: Mapping of Scattering and Absorption Properties and Application to Metropolitan France
- Separation of Source, Site, and Path Effects Based on Elastic Radiative Transfer Theory, Part 2: Generalized Inversion of Source and Site Parameters and Application to Metropolitan France
- AeroTrans: Hourly AOD retrieval over land from MSG-1/SEVIRI imagery integrating Transformer and transfer learning
- UDL-RT: A unified differentiable all-sky microwave radiative transfer model with learnable parameterizations
BibTeX
@misc{sousa2026accounting,
title = {{Accounting for Radiative Transfer Model Errors in Physics-Guided VAEs for Biophysical Variable Retrieval}},
author = {Kevin de Sousa and Jordi Inglada},
howpublished = {HAL},
year = {2026},
month = sep,
url = {https://hal.science/hal-05740441v1},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Accounting for Radiative Transfer Model Errors in Physics-Guided VAEs for Biophysical Variable Retrieval
AU - Kevin de Sousa
AU - Jordi Inglada
PY - 2026
DA - 2026/09/07
PB - HAL
UR - https://hal.science/hal-05740441v1
AB - Monitoring essential climate variables (ECV) from optical remote sensing remains challenging due to the ill-posed nature of biophysical variable retrieval and the limitations of radiative transfer models (RTMs). The recently proposed PROSAIL-VAE framework addresses some of these issues by embedding physical modeling within a variational autoencoder (VAE). Unaddressed RTM deficiencies -- including oversimplified soil spectra, un-modeled canopy clumping, and signal saturation -- introduce systematic residuals that are assumed negligible in most hybrid retrieval frameworks. We adapt a residual physics-guided VAE that explicitly accounts for RTM shortcomings via a refinement network and extended latent space, constrained to balance physical and data-driven contributions. The framework has two outputs: improved retrieval and learned residuals that can be interpreted in physical terms. On two independent in-situ datasets, retrieval improvements over the operational BVNET processor (Leaf Area Index (LAI) RMSE 19% and 5.2%; Canopy Chlorophyll Content (CCC) 47.9%) demonstrate the behavior of the framework. A non-trivial CCC degradation of 7.7% relative to the baseline PROSAIL-VAE emerges as a trade-off of reconstruction-driven optimization, where chlorophyll's weak spectral signature becomes entangled with structural and soil parameters. We further show that this degradation is partially mitigated by the extended latent space and decorrelation constraints. Spectral and spatial analysis of the learned residuals reveals that they concentrate in regimes where PROSAIL is known to be deficient -- particularly soil reflectance modeling and signal saturation -- suggesting that residual modeling can serve both as a retrieval enhancement and as a diagnostic probe of model inadequacy, while acknowledging that a fully validated diagnostic framework remains an open challenge.
ER -