Training course: EUMETSAT/ECMWF NWP-SAF satellite data assimilation

Global Observing System overview

This is an introductory seminar to briefly introduce the main concepts in the satellite global observing system, the types of satellite observation and how they are used and monitored at ECMWF, and how the global observing system as a whole is designed, managed and integrated by the World Meteorological Organization (WMO).

The presentation will be split into three parts. In the first part the concepts of polar orbiting and geostationary orbiting satellites will be briefly presented, and the challenges and opportunities posed by both techniques discussed. Also other possible orbital configurations, such as highly elliptical orbits will be mentioned. The changes in satellite data used over the last 15 years will be discussed, with some of the key missions highlighted. A very brief introduction to how we monitor the performance of the satellite data in the assimilation will be given.

In the second part the website maintained by WMO, named OSCAR, will be presented and it will be explained why this is such a useful facility and how it can be used to plan for and understand the changing observing system.

The session will end with a practical session using OSCAR. We will log on to the OSCAR website and examine the information available.

What do satellites measure?

The primary purpose of this lecture is to explore the implications of the fact that satellites can only measure radiation at the top of the atmosphere and do not measure the geophysical variables we require for NWP (e.g. temperature, humidity and wind).

The link between the atmospheric variables and the measured radiances is the radiative transfer equation - the key elements of which are discussed. It is shown how - with careful frequency selection - satellite measurements can be made for which the relationship to geophysical variables is greatly simplified. For example, in the case of sounding radiances (where the primary absorber is a well mixed gas of known concentration), the relationship reduces to the measured radiance being a broad vertical average of the atmospheric temperature profile. Despite these simplifications, it is shown that the extraction of detailed profile information from downward looking radiance measurements is a formally ill posed inverse problem. The ways in which retrieval schemes use prior information to make the inverse problem more tractable are discussed as well as how this approach generalises into data assimilation.

Data assimilation algorithms and key elements

Numerical models provide an almost complete description of the atmosphere, but errors grow rather rapidly in time. Observations provide fairly incomplete, but up-to-date accurate information. Data assimilation is introduced as the process where these two sources of information are combined to produce a best or optimal estimate of the atmospheric state.

The state is optimal in the sense that background information from the model is combined with observed information - respecting the uncertainty in both (encapsulated by the background error covariance B and observation error covariance R). This lecture looks in detail at the main elements of the assimilation scheme (such as the chain of observation operators for radiances) and its key statistical inputs. For example it is shown that an incorrect specification observation errors (R) can result in the assimilation process producing an analysis that has larger errors than the background.

Also, the importance of vertical correlations of background error is discussed - with respect to the assimilation of low vertical resolution downward looking radiance data. The role of the bias correction system and the data selection/quality control in producing an input population of observations that have statistical properties consistent with R is described.

The infrared spectrum, measurement, modelling and information content

This lecture gives an overview of the atmospheric infrared spectrum. The main spectral regions are discussed: the long-wave (15um) and short-wave (4um) CO2 bands providing temperature information, the ozone band at 9um, the water vapour band around 6um and finally the window regions.

The measurement of the infrared spectrum by the primary operational sounders (IASI and CrIS) is described, highlighting the different compromises made between measurement noise and spectral resolution. The key challenges associated with using the IR spectrum are discussed with particular emphasis on cloud and non-linearity. The lecture includes several practical exercises where ECMWF model fields and the RTTOV radiative transfer scheme are used to simulate IASI spectra and Jacobians in different atmospheric conditions.

Introduction to GPS radio occultation

GPS radio occultation (GPS-RO) measurements are a relatively recent addition to the global observing system. GPS-RO is an active measurement technique with a limb geometry. The GPS-RO measurements complement the information provided by satellite radiances, because they have good vertical resolution, and they can be assimilated without bias correction to the NWP model. This means they "anchor" the bias correction of the radiances.

This lecture explains the physical basis of the GPS-RO technique, and outlines the main preprocessing steps required to map from the raw measurements to quantities used in NWP. Information content studies are used to demonstrate that the measurements provide excellent temperature information in the upper-troposphere and lower/middle stratosphere.

The approach used at ECMWF to assimilate the GPS-RO measurements is described. We show how the assimilation of GPS-RO was able to correct a long-standing ECMWF temperature analysis problem in the polar regions, and explain the "null-space" of the measurements. The ability of GPS-RO to constrain the surface pressure field, and the retrieval planetary boundary layer height information from GPS-RO, are discussed briefly.

Wind information from satellites  (PMW, SCAT, AMV and PT)

In this session we will present the main sources of dynamical information available from satellites. Dynamical information is currently extracted from satellite data in three ways.

Firstly, features can be tracked in a sequence of satellite images and the movement of the tracked feature used to deduce wind information.

Secondly, the wind stress induced roughening of the ocean's surface can modify both active and passive microwave observations and near-surface marine wind information can be obtained, usually through direct observation assimilation or through a retrieved 10m neutral equivalent windspeed (or vector).

Thirdly, dynamical information can be analysed in a 4D-var assimilation system through radiance assimilation combined with the adjoint of the forecast model. 4D-var finds a solution that fits the observations by moving features (i.e. adding a wind increment) rather than changing the feature.

There is a fourth method being developed, with a view to a first launch in 2017, which uses doppler lidar to infer vertical wind profiles, but this will be mentioned only briefly.

We will begin by discussing the atmospheric motion vectors and radiance assimilation, concentrating on the technique, main challenges in assimilation and the current state of the art. The doppler wind lidar will then be briefly introduced. Finally the passive and active microwave techniques will be described, again starting with the basic science but moving quickly on to the details of how data is currently used and the issues that need to be resolved. In all cases the data currently used and forecast impact will also be shown.

Observation errors

The role of observation errors in the assimilation of observations is introduced. Together with background errors, the observation errors determine the weight of an observation in the analysis. Observation errors include measurement errors, but also forward model errors, and errors of representativeness.

Various methods exist to diagnose observation errors from output of assimilation systems, and these will be introduced and illustrated with practical examples. It will be shown that satellite data often exhibit some error correlations, either spatially or, for radiances, between different channels.

The specification of observation errors in practice in real assimilation systems will also be introduced. Mostly, observation errors are assumed to be diagonal. Concepts to either counteract the effect of error correlations or to explicitly take them into account will be introduced, based on examples taken from the ECMWF system.

Microwave observations in clear-sky conditions

The main uses of the microwave spectrum in weather prediction are temperature and moisture sounding and the estimation of surface properties. Typical microwave instruments like AMSU-A and SSMIS will be introduced.

The interaction of microwave radiation with the surface will be described in the context of microwave imaging channels. The spectroscopy of the 60 GHz oxygen line will be examined along with the temperature sounding channels that it makes possible.

The AMSU-A 60 GHz temperature channels are perhaps the most important single source of observations in weather-forecasting and their practical use at ECMWF will be described.

Microwave observations in all-sky conditions

Microwave radiation has wavelengths of the order of millimetres, very similar to the size of atmospheric particles such as rain and snow. This means that the interaction of cloud and precipitation particles with microwave radiation is complex and frequency-dependent, but it means that the microwave is one the best ways to observe cloud and precipitation in all its forms.

Weather forecasting centres are increasingly trying to assimilate this information and to extend temperature and moisture-sounding information further into areas affected by cloud and precipitation.

This talk will briefly look at solving the radiative transfer equation in scattering conditions before examining the practical details of microwave assimilation in all-sky conditions. All-sky assimilation is still not practical in every case, particularly temperature sounding, so methods for cloud screening will also be covered.

The detection and assimilation of clouds in infrared radiances

The potentially extreme impact of cloud on infrared observations represents a huge challenge to the successful exploitation of these data for NWP. This impact of cloud is illustrated in practical exercises where cloud is introduced to the RTTOV radiance simulations.

The lecture then considers two options to handle clouds in the observations. The first option is the detection (and rejection) of cloud-affected data, and the lecture introduces a variety of algorithms, from simple window channel checks to complex pattern recognition, that may be used to identify when a particular scene is contaminated by cloud.

The additional use of collocated imager diagnostics (e.g. AVHRR/IASI) for scene homogeneity is also described. Potential pitfalls when cloud detection algorithms can go wrong are discussed - for example when emission from the underlying surface is poorly modelled. The second option (as an alternative to the wasteful rejection of cloudy data) is the explicit treatment of cloud as part of the analysis control vector - where parameters describing the cloud are estimated from the observed spectra simultaneously with other atmospheric variables such as temperature and humidity.

The large magnitude of the cloud signal, non-linearity of Jacobians and the complexity of the cloud parameters are presented as significant scientific issues. The operational ECMWF scheme treating overcast cloudy radiances with an extremely simple parametric description of cloud is presented. Prospects for a complex treatment of cloud for all-sky infrared radiances - fully interactive with the forecast model physics are discussed.

Estimation and correction of systematic errors

Biases in modern satellite measurements are generally rather small but, due to the global nature of the data, they can still be very damaging to a data assimilation scheme if not corrected. A variety of sources of bias are discussed - from the instrument itself to the radiative transfer model used to simulate the data, as are the difficulties in estimating biases using an NWP model state that may itself have systematic errors.

Various approaches to the correction of biases are described in this lecture, from simple static offsets to highly variable adaptive (automatic) correction schemes. The dangers of a correction that is over- or under adaptive are considered, as is the potential for a bias correction scheme to interact (negatively) with departure-based quality control decisions. Constraining correction schemes with anchoring observations (i.e. data assimilated with no bias correction applied) is presented as a useful safety measure. Finally, the importance of constantly monitoring (automatically) the corrections generated by fully adaptive schemes is illustrated.