Integrating image inpainting into an ensemble score filter to track surface quasi-geostrophic dynamics under partial observations

The authors build an ensemble score filter (EnSF) that incorporates image inpainting to solve high-dimensional nonlinear data assimilation problems with partial observations. The original EnSF had been demonstrated with full observations, but because it does not use a covariance matrix to capture dependence between observed and unobserved state variables, extending it to partial observations is nontrivial; this work adds image inpainting to predict the unobserved states. The abstract describes the method design: at each filtering step the diffusion model estimates observed states by integrating likelihood information into the score function, and image inpainting methods then predict the unobserved state variables.

The EnSF relies on an exclusively designed training-free diffusion model to solve high-dimensional nonlinear data assimilation problems. Compared with machine-learning assimilation methods that require training, the emphasis here is on a diffusion model design that needs no training. The abstract describes the diffusion model as 'exclusively designed training-free diffusion models' and provides no training details or comparative experimental data.

The authors demonstrate the performance of the EnSF with inpainting by tracking Surface Quasi-Geostrophic (SQG) model dynamics under a variety of scenarios. By combining image inpainting with the score filter, the validation setting extends from full observations to partial observations across multiple scenarios. The abstract calls this a 'successful proof of concept' on SQG dynamics and reports no specific error metrics or sample sizes in the abstract.

The result targets data assimilation settings with partial observations, demonstrated as a proof of concept on Surface Quasi-Geostrophic (SQG) model dynamics across a variety of scenarios. It is meant for researchers who want to predict unobserved states via image inpainting without using a covariance matrix to capture dependence between observed and unobserved variables. The authors state that this successful proof of concept paves the way for more in-depth investigations exploiting modern image inpainting techniques for practical geoscience and weather forecasting problems.

The abstract provides no specific error metrics, comparison baselines, or observation sparsity levels, so the accuracy gain of the EnSF with inpainting under partial observations relative to other methods remains to be confirmed in the main text; differences among the various image inpainting techniques, and the feasibility of moving from the SQG proof of concept to practical weather forecasting systems, are also questions readers should keep watching.