Automated Selection of Meteorological Ensemble Members for Inversions

Atmospheric inversions rely on accurately modeled atmospheric conditions, especially wind speeds, often for days in the past. Most atmospheric model products - ECMWF, GFS, and HRRR in our case - report on a scale of 0.1 - 0.5 degrees, or ~10 - 50km. These products can work quite well for inversions at coarse mesoscale resolutions, that model winds for ~100 - 1000km and beyond. However, inversions working at finer scales face significant problems in trying to model winds over a single or even a few meteorological model grid cells.

The MethaneSAT mission attempts to run an atmospheric inversion at a scale of 4km x 4km through the CORE algorithm (Described by Knapp et al at EGU2026), and in trying to extract emissions from instrument observations, we sometimes see plumes that run 30 degrees or more off of model wind directions. Wind speed and especially direction errors at this scale often lead to failed inversions, accounting for a roughly 10% - 20% loss of scenes collected by MethaneSAT, second only to cloudiness.

Luckily, many atmospheric model products are distributed as both a single estimate and as an ensemble product, containing dozens of perturbed forecasts for every time step. To minimize the CORE inference error induced by error in wind estimates, we present a technique for quickly and efficiently analyzing the wind fields in comparison to the concentration Level3 maps in order to select the most likely ensemble member. We utilize a novel Total Variation approach to quantify the expected alignment of wind fields with measured methane gradients, and select the ensemble product whose winds are most likely to produce the observed concentration map.

We demonstrate this technique on both simulated and real MethaneSAT data, and discuss the effects this has on both success rate of the inversion and on residual errors from the inversion, though the approach is not specific to methane, and is broadly applicable in any case where winds are the primary drivers of transit. Special care is taken to identify pathological cases of wind error, and how these could be addressed in the future.