Résumé
Next-generation ground-based surveys such as the Vera C. Rubin Observatory's Legacy Survey of Space and Time require photometric calibration that is both long-term stable and spatially uniform at the sub-percent level, even during non-photometric conditions. Achieving this precision motivates new approaches to characterize atmospheric transmission, particularly to mitigate gray extinction from clouds. The StarDICE experiment aims to establish a metrology chain linking laboratory standards to astrophysical fluxes with 1 mmag accuracy in the$\textit{griz}$bands, a goal for which controlling variable atmospheric effects is essential. We present a method that corrects photometric measurements using simultaneous radiometric information from an infrared thermal camera. The gray-extinction model is fit on an image-by-image basis using thermal radiance excess and the difference between synthetic and instrumental fluxes of calibration stars, without requiring assumptions about the spatial structure of extinction. The method relies on a forward model that incorporates environmental monitoring, radiative-transfer simulations, and Gaia DR3 stellar catalogs. Using data from a remote observing system that repeatedly monitored two fields under diverse atmospheric conditions, we show that the corrections reduce residuals between corrected and reference magnitudes and produce extinction maps with 2-arcmin resolution and$\sim$ 0.01 mag accuracy. Using this technique, we can recover data acquired under non-photometric conditions with a precision comparable to data obtained under photometric conditions. For the most affected exposures, the mean absolute error improves from 0.64 to 0.11 mag, and temporal extinction variations can be reduced to 0.025 mag per source. We discuss the implications of this technique for future surveys and outline directions for further refinement.