How can we make better use of satellite observations for weather forecasting?

By Rishabh Bhatt, Postdoctoral Researcher in the Department of Meteorology

Observations are used in numerical weather prediction (NWP) through a process known as data assimilation (DA), which combines observations with a short-range model forecast (known as the background) to produce the best estimate of the atmospheric state (known as the analysis). Satellite observations play an important role in weather forecasting by providing wide-ranging, high-resolution measurements of the Earth’s atmosphere. Making the best use of this wealth of information is becoming increasingly important as NWP systems move towards kilometre-scale resolutions.

Despite the abundance of available satellite data, only a fraction is currently used in operational weather forecasting. For example, at the European Centre for Medium-Range Weather Forecasts (ECMWF), only around 40-50% of the available Advanced Microwave Sounding Unit-A (AMSU-A) observations are assimilated. AMSU-A is a microwave temperature sounder that measures atmospheric temperature from the surface to the stratosphere and is consistently one of the most influential observing systems for improving forecast skill. While some observations are rejected during quality control (QC) procedures, many more are intentionally discarded through a process known as observation thinning.

Observation thinning is primarily used to mitigate the effects of unaccounted-for spatially correlated observation errors.  At ECMWF, AMSU-A observations are typically thinned to a spacing of approximately 125 km, meaning that only one observation is retained within each 125 km by 125 km grid box. If these spatial observation-error correlations can be accurately represented within DA systems, the thinning distance can be reduced to around 31 km, increasing the number of assimilated AMSU-A observations by almost a factor of five (see Figure 1). Our recent research addresses this challenge by estimating the spatial observation-error correlations in AMSU-A observations, identifying their main physical sources, and providing guidance on where more observations can be assimilated in current operational forecasting systems.

What are spatial observation-error correlations?

Observation-error correlations describe how the errors in two observations are related. When the observations are taken at different locations, spatial observation-error correlations describe how this relationship changes with the distance between them.

To illustrate this concept, imagine temperature observations taken in Reading, London and Edinburgh. If the observation error is large in London, it is more likely to be large in nearby Reading than in distant Edinburgh (assuming positive correlations). This is because nearby observations are more likely to share common sources of error, such as instrument calibration biases, radiative transfer modelling errors, or unresolved atmospheric variability.

The distance over which these shared errors remain important is described by a correlation length scale. In simple terms, it tells us how far apart two observations can be before their errors stop looking similar. A larger correlation length scale means that these shared errors extend over a greater distance. Figure 2 illustrates the physical meaning of a correlation length scale.

Figure 2 Illustration of the physical meaning of a correlation length scale. A correlation length scale of 120 km means that observations within a 120 km radius of the black observation are likely to have similar errors (shown in red and yellow). Beyond this distance, the similarity in their errors becomes much weaker (shown in pale yellow).

Because of these shared errors, neighbouring observations are not fully independent and do not always provide completely new information. If the DA system ignores these correlations, it can overestimate the information content of densely spaced observations, giving them too much weight and potentially degrading the analysis. By explicitly accounting for spatial observation-error correlations, observations can be weighted more appropriately, allowing more of the available satellite observations to be used while maintaining the quality of the analysis.

Correlations in AMSU-A observations

In our recent study (Bhatt et al., 2026), we estimated spatial observation-error correlations in AMSU-A observations assimilated under all-sky conditions (i.e. including both clear and cloudy scenes) using one month of operational data from ECMWF and the UK Met Office. The correlations were estimated using the widely adopted method of Desroziers et al. (2005), which estimates observation-error statistics from differences between the observations, the background and the analysis.

Figure 3: Estimates of spatial observation-error correlations in AMSU-A observations for tropospheric channels 4 (left), 5 (middle), and 6 (right) assimilated at ECMWF (red circles) and the Met Office (blue squares). The bars indicate the number of observation pairs used to compute the estimates. Channel 4 is not assimilated at ECMWF.

Our results show positive spatial observation-error correlations whose strength depends on both the surface type and the presence of clouds. Figure 3 shows the estimated spatial observation-error correlations for the AMSU-A tropospheric channels 4-6 onboard Meteorological Operational Satellite-C (MetOp-C). Each AMSU-A channel is most sensitive to temperature at a different height in the atmosphere. Channels 4-6 mainly sense temperature in the lower and middle troposphere, but are also influenced by clouds, precipitation, and to some extent, surface emission. The estimated correlation length scales for these channels (defined here as the separation distance at which correlations fall below 0.2) range between 75 and 125 km. Particularly strong correlations are found for channel 4 at the Met Office and for channels 5 and 6 at ECMWF. These stronger correlations are expected because these channels observe the part of the atmosphere where clouds and precipitation are most common. Since these features are difficult to represent accurately in forecast models, nearby observations often end up sharing similar errors.

More generally, the strongest correlations occur for the tropospheric channels over land and under cloudy conditions at both operational centres. This suggests that neighbouring observation errors are often affected by same sources of uncertainty. Our findings indicate that errors in the specification of surface emissivity, skin temperature, clouds and precipitation in the forecast model are the main drivers of the observed spatial correlations. These findings provide useful guidance on where observation thinning could potentially be reduced for AMSU-A observations in the Met Office and ECMWF systems.

Current status and outlook

At present, most operational global forecasting centres do not account for spatial observation-error correlations. Instead, they rely on observation thinning to reduce the impact of these neglected correlations.

The biggest challenge is computational cost. Accounting for these correlations in DA means working with a very large matrix (known as the observation-error covariance matrix) that describes how the errors in all observations are related to one another. Because this matrix is enormous and changes every time new observations are assimilated, using it directly is currently too expensive for operational weather forecasting.

Our current research focuses on developing faster numerical techniques that make these calculations practical. If successful, this would allow many more satellite observations to be assimilated while properly accounting for their correlated errors. Since microwave observations such as from AMSU-A are among the most influential observation types for NWP, making better use of them has the potential to further improve weather forecasts.

References

Bhatt, R., Bonavita, M., Bormann, N., Dance, S.L., Fowler, A., Hólm, E., Merchant, C.J., Mittaz, J., Newman, S. and Waller, J., 2026. Spatial observation‐error correlations for AMSU‐A in all‐sky assimilation: An ECMWF and UK Met Office intercomparison. Quarterly Journal of the Royal Meteorological Society, p.e70238. https://doi.org/10.1002/qj.70238

Desroziers, G., Berre, L., Chapnik, B., & Poli, P. (2005). Diagnosis of observation, background and analysis‐error statistics in observation space. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography, 131(613), 3385-3396. https://doi.org/10.1256/qj.05.108