Context

Infrared (IR) imagers onboard geostationary satellites observe the Earth continuously, day and night, including over oceans and mountains where ground radar coverage is sparse or absent. However, they sense cloud tops rather than surface rain, so the link between the measured signal and precipitation is indirect. Retrieving rain from IR brightness temperatures is an ill-posed inverse problem: many different rain fields are compatible with the same cloud-top signal. Deep learning models for satellite precipitation estimation [1, 2, 3] are usually trained with regression losses (MSE, MAE, quantiles). They predict a conditional mean rather than a realistic rain field: the output is overly smooth and underestimates heavy rain [4].

Recent generative models, such as flow matching [5], offer an attractive alternative. Instead of regressing a single estimate, they learn to turn noise into samples from the distribution of possible rain fields given the observations, so each generated field looks like real rain, with its sharp structures and intense cells [4, 6]. Drawing several samples for the same satellite image also gives several plausible rain fields, a natural way to represent uncertainty. However, retrieving each time step independently offers no guarantee of temporal coherence: nothing ensures that consecutive retrieved fields form a plausible storm evolution.

Moving beyond instantaneous maps, video diffusion models [7] have recently been proposed to generate the dynamics of precipitating events: HiDFiP generates sequences of precipitation fields from satellite precipitation estimates and atmospheric variables, with storm dynamics consistent with ground radar. Following this line, this internship will investigate video flow matching to generate temporal sequences of rain fields (from the US radar mosaic MRMS) from infrared observations of GOES (the US geostationary IR satellites).

The internship will be carried out in collaboration with S. Upadhyaya (Indian Institute of Technology Hyderabad) and P. Kirstetter (University of Oklahoma), who are jointly developing a GOES/NEXRAD version of this type of retrieval. The team is also part of the ARCHES project team (Inria Paris–LATMOS).

Objectives

The overall goal is to improve satellite precipitation estimation by explicitly accounting for the dynamics of precipitating systems, while building expertise in state of the art generative modelling.

• Model development. Develop a video flow matching model that learns the fine-scale spatial and temporal dynamics of precipitation from GOES IR sequences and MRMS reference data.
• Validation. Evaluate the generated sequences against MRMS: accuracy of rain rates and extremes, realism of spatial structures, temporal coherence and storm evolution, and quality of the uncertainty carried by the generated ensembles.

We encourage the submission of a paper to a conference at the end of the internship.

Profile and prerequisites

Master’s or engineering degree in computer science, applied mathematics, data science or environmental/geosciences with a strong machine learning component;
solid skills in Python (PyTorch) and deep learning, and interest in generative models;
ability to work independently, rigor and good writing skills;
knowledge of remote sensing or meteorology is a plus, but not required.

How to apply

Please send your CV and cover letter by email to matthieu.meignin@latmos.ipsl.fr.

References

[1] N. Viltard, V. Sambath, P. Lepetit, A. Martini, L. Barthès, C. Mallet. Evaluation of DRAIN, a Deep Learning Approach to Rain Retrieval from GPM Passive Microwave Radiometer. IEEE Transactions on Geoscience and Remote Sensing, 2023. doi:10.1109/TGRS.2023.3293932.

[2] S. Pfreundschuh, C. D. Kummerow, J. Tan, G. J. Huffman. GPROF-IR: An improved single-channel infrared precipitation retrieval for merged satellite precipitation products. arXiv:2605.07167, 2026.

[3] A. Amell Tosas, L. Hee, S. Pfreundschuh, P. Eriksson. Probabilistic near-real-time retrievals of rain over Africa using deep learning. J. Geophys. Res. Atmos., 130(20), e2025JD044595, 2025. doi:10.1029/2025JD044595.

[4] M. Meignin, C. Mallet, N. Viltard. Generative and deterministic deep learning models comparison for fine-scale precipitation retrievals from infrared brightness temperature. arXiv:2610.09859, 2026.

[5] Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, M. Le. Flow matching for generative modeling. Int. Conf. on Learning Representations (ICLR), 2023. arXiv:2210.02747.

[6] M. Meignin, C. Mallet, N. Viltard. RainFlow: Flow Matching for Infrared-Only Quantitative Precipitation Estimation. 2026. hal.science/hal-05855393.

[7] R. Li, Y. Xia, Y. Qu, D. Fu, C. Guilloteau, J. Hoffman, S. Mandt, P. Gentine, E. Foufoula-Georgiou. Video diffusion for satellite-based high-dynamical-fidelity precipitation (HiDFiP) field generation. arXiv:2608.21812, 2026.