A feasibility study to Reconstruct Atmospheric Rivers using space- and ground-based GNSS observations

Abstract. Atmospheric rivers (ARs) are long filaments that transport large amounts of water vapor from the Tropics to mid- and high latitudes. They are directly related to heavy precipitation and extreme weather leading to flooding and mud slides. Accurate identification of AR structures over the ocean is important to improve the forecast of their landfall location and timing. GNSS radio occultation (RO) is a space-based technique that can measure meteorological variables with high vertical resolution. While RO can observe structures like ARs in individual RO profiles, RO observations have non-uniform and sparse spatial and temporal sampling, so it is not yet possible to fully characterize AR morphology using RO alone.

In this work, we use previous research in which we applied machine learning (ML) to enhance the spatial and temporal resolution of RO observations. Here, we train neural networks (NNs) to map RO observations and help resolve ARs. Analyses using existing RO data, such as from the COSMIC-2 mission, showed that the sampling density is insufficient to resolve and geo-locate ARs. Adding observations from the other available missions (for example METOP) improved matters, but was still insufficient to reliably reconstruct AR structure.

We undertake a study to determine how many LEO RO satellites would be needed to quantify the structure, location, and timing of ARs. We simulate RO observations as would be obtained with Walker constellations of 12, 24, 36, 48 and 60 LEO RO satellites. First, we investigate possible constellations for proper AR monitoring. We aim for constellations that lead to hourly RO counts that change as little as possible during the AR (up to several days). This allows us to resolve ARs with similar accuracy during the scenario. We conclude that 3 or 6 orbital planes and inclinations between 85° and 90° perform best. Second, we make use of 12-h forecasts of the European Centre for Medium-range Weather Forecasts (ECMWF) system to interpolate the forecasts to the simulated RO constellation sampling coordinates. Third, we use the ECMWF-based RO observations to train ML models and map them to the ECMWF grid. We compare ML-mapped RO sampled grids to ECMWF products in a closed-loop validation. Initially, we map RO refractivity at 2 km geopotential height, where small-scale structures related to water vapor are visible. We find that at least 36 RO satellites are needed to characterize the morphology of ARs in the Pacific basin with useful precision and accuracy (from the ML produced maps). Then, we use a framework with two consecutive NNs to map column-integrated water vapor (IWV) from profiles of RO. The first NN maps the refractivity into IWV, and the second NN maps the IWV spatially. In this case, we find that a constellation of 48 satellites is needed to continuously map IWV fields accurately and thus reconstruct the morphology of ARs with useful precision and accuracy. Finally, when using RO, we find that mapping refractivity into IWV is less accurate over land than over oceans. To further improve the AR mapping over land, we made use of IWV from ground-based (GB) GNSS. The significantly higher spatial and temporal resolutions of GB data compared to RO lead to much improved IWV fields and thus AR path and shape over land.