## THE UNIVERSITY OF READING DEPARTMENTS OF MATHEMATICS AND METEOROLOGY Correlated observation errors

### BibTeX

@MISC{Stewart_theuniversity,

author = {Laura M. Stewart and Laura Stewart and I Dr and Sarah Dance and Prof Nancy Nichols},

title = {THE UNIVERSITY OF READING DEPARTMENTS OF MATHEMATICS AND METEOROLOGY Correlated observation errors},

year = {}

}

### OpenURL

### Abstract

Data assimilation techniques combine observations and prior model forecasts to create initial conditions for numerical weather prediction (NWP). The relative weighting assigned to each observation in the analysis is determined by the error associated with its measurement. Remote sensing data often have correlated errors, but the correlations are typically ignored in NWP. As operational centres move towards high-resolution forecasting, the assumption of uncorrelated errors becomes impractical. This thesis provides new evidence that including observation error correlations in data assimilation schemes is both feasible and beneficial. We study the dual problem of quantifying and modelling observation error correlation structure. Firstly, in original work using statistics from the Met Office 4D-Var assimilation system, we diagnose strong cross-channel error covariances for the IASI satellite instrument. We then see how in a 3D-Var framework, information content is degraded under the assumption of uncorrelated errors, while retention of an approximate correlation gives clear benefits. These novel results motivate further study. We conclude by modelling observation error correlation structure in the framework of a one-dimensional shallow water model. Using an incremental 4D-Var assimilation system we observe that analysis errors are smallest when correlated error covariance matrix approximations are used over diagonal approximations. The new results reinforce earlier conclusions on the benefits of including some error correlation structure. i Declaration I confirm that this is my own work and the use of all material from other sources has been properly and fully acknowledged.

### Citations

6050 |
A mathematical theory of communication
- Shannon
- 1948
(Show Context)
Citation Context .... These measures are generally applicable and can be defined without reference to a specific retrieval method [77]. 3.6.1 Entropy reduction Entropy is a real valued functional that characterises pdfs =-=[82]-=-. Recall from Section 2.2, that pdfs can be used as a measure of knowledge of the state and observation vectors. If Pb(x) represents the knowledge of the state vector before the observations and Po(x|... |

349 |
Geophysical Fluid Dynamics
- Pedlosky
- 1987
(Show Context)
Citation Context ...ween conception and operational implementation. They have been shown capable of describing important aspects of the dynamic properties we wish to model, such as geostrophic motion in three-dimensions =-=[71]-=-. In the final chapters of the thesis we will study the behaviour of a data assimilation algorithm under different approximations to the observation error covariance matrix. We will use the SWEs as th... |

243 |
Atmospheric Data Analysis
- Daley
- 1991
(Show Context)
Citation Context ...on errors not only include errors explicitly generated in taking the observations, but also errors in the treatment of these observations. These two sources are sometimes considered in separate terms =-=[20]-=-, but in this work we will contain them in a single error covariance matrix for algebraic convenience. Observation errors are distributed in the horizontal and vertical, and can generally be attribute... |

242 | Atmospheric Modeling, Data Assimilation and Predictability; Cambridge University Press
- Kalnay
- 2010
(Show Context)
Citation Context ... and the background state, weighted by the inverse of their respective error covariances. The cost function minimisation can be solved approximately to obtain the best linear unbiased estimate (BLUE) =-=[49]-=-, xa: x a = x b + K(y − h(x b )), (2.9) K = BH T (HBH T + R) −1 , (2.10) where H is the linearised observation operator given by H = ∂h ∂x | and K is the Kalman xb gain matrix specifying the optimal w... |

172 |
Statistical methods in the atmospheric sciences
- Wilks
- 1995
(Show Context)
Citation Context ...hat is too static may lead to wasting crucial information. NWP centres are now investigating alternative treatments of huge data volumes. One 28promising method is principal component analysis (PCA) =-=[92]-=-. The nature of PCA techniques is to approximate data vectors with many elements (i.e, IASI observations of 8461 channels) by a new correlated set of data vectors containing fewer elements. The proced... |

129 |
A strategy for operational implementation of 4D-VAR, using an incremental approach
- Courtier, Thepaut, et al.
- 1994
(Show Context)
Citation Context .... When 4D-Var methods were originally being investigated it was determined that in order to use a 4D-Var algorithm, a significantly faster computer or a substantial algorithmic improvement was needed =-=[18]-=-. The formulation of an incremental 4D-Var algorithm provided this improvement. 2.3.1 Incremental 4D-Var Incremental 4D-Var, as proposed by Courtier et al [18], reduces the cost of the 4DVar algorithm... |

121 |
Analysis methods for numerical weather prediction
- Lorenc
- 1986
(Show Context)
Citation Context ... some prior atmospheric information and a set of observations. The best analysis can be defined as that which gives the best subsequent forecast, without compensating for errors in the forecast model =-=[62]-=-. Consider a discretised representation of the true state of the atmosphere x t ∈ R n , where n is the total number of state variables. The analysis used in NWP will consist of the same model variable... |

110 |
Semi-Lagrangian integration schemes for atmospheric models : a review
- Staniforth, Côté
- 1991
(Show Context)
Citation Context ... a set of water parcels. A set of originally regularly spaced parcels at one time step may evolve to be very close to each other at the next time step, and therefore some areas may be poorly resolved =-=[84]-=-. A semi-Lagrangian scheme tracks a different set of parcels at each time step; chosen so that their positions at the next time step (known as the arrival point) are at regularly spaced grid points. T... |

72 |
Toeplitz and Circulant Matrices: A review, Foundations and Trends
- Gray
(Show Context)
Citation Context ...he form ⎛ Tm = ⎜ ⎝ t0 t−1 t−2 . . . t−(m−1) t1 t0 t−1 . . . t−(m−2) t2 t1 t0 . . . . . . . ... ... ... . tm−1 tm−2 . . . t1 t0 ⎞ , (3.2) ⎟ ⎠ where Tm = [tk,j; k, j = 0, 1, . . .,m −1] and tk,j = tk−j =-=[41]-=-. A standard Toeplitz matrix has 2m − 1 independent entries; a symmetric Toeplitz matrix has only m independent entries. A simpler class of banded Toeplitz matrices can be defined by the constraint th... |

72 |
The Essence of Chaos
- Lorenz
- 1993
(Show Context)
Citation Context ...each component in the assimilation is determined by its associated error. The chaotic nature of the atmosphere requires that the initial conditions be accurately specified to avoid rapid error growth =-=[64]-=-, and thus the correct specification of the weighting errors is vital. 1.1 Motivation In numerical weather prediction (NWP) the governing equations used to describe the behaviour of the atmosphere con... |

72 |
The ECMWF operational implementation of four-dimensional variational assimilation–Part I: Experimental results with simplified physics
- Rabier, Jarvinen, et al.
(Show Context)
Citation Context ...performing the inner loop minimisation at a lower spatial resolution. Incremental 4D-Var is the data assimilation algorithm currently used at several NWP centres such as the Met Office [76] and ECMWF =-=[75]-=-. However, correctly specifying the observation error structure is still an important issue under this new formulation. 2.3.2 Practical implementation We now consider the practical details of implemen... |

60 |
A global three-dimensional multivariate statistical interpolation system
- Lorenc
- 1981
(Show Context)
Citation Context ...servations and the background [49]. Sequential assimilation algorithms approximate (2.9) and (2.10) directly; optimal interpolation is an example of a suitable algorithm previously used operationally =-=[61]-=-. In current operational NWP, the error covariance matrices are too large to be used explicitly in global assimilation problems, and therefore an implicit variational alternative is needed. The statis... |

54 |
The ECMWF implementation of three-dimensional variational assimilation (3DVAR). I : Formulation
- Courtier, Andersson, et al.
- 1998
(Show Context)
Citation Context ...on (3D-Var) [63]. This algorithm seeks the minimum of (2.8) by performing several evaluations of the cost function and its gradient in order to approach the minimum using a suitable descent algorithm =-=[17]-=-. The size of the error covariance matrices is still an issue with the main cost in the cost function evaluation lying in inverting B and R. Although conceptually useful, 3D-Var treats observations as... |

53 | Computing the Nearest Correlation Matrix a Problem From Finance
- Higham
- 2002
(Show Context)
Citation Context ...ar problems have been studied in financial research. In [87] the problem of finding the nearest positive semidefinite Toeplitz matrix (in the Frobenius norm) to an arbitrary matrix was considered. In =-=[44]-=- the nearest correlation matrix to a given symmetric matrix was determined by minimising the distance between the two matrices in a weighted Frobenius norm. By extending the ideas in [44] and [87] we ... |

49 |
The statistical structure of short-range forecast errors as determined from radiosonde data. Part I : the wind field, Tellus 38A
- Hollingsworth, Lönnberg
- 1986
(Show Context)
Citation Context ...önnberg approach The most commonly used estimation technique is the observational method, otherwise known as the Hollingsworth-Lönnberg method after the authors who popularised its use in meteorology =-=[47]-=-. This method uses background innovations statistics from a dense observing network, under the assumption that the background errors carry spatial correlations while the observation errors do not. The... |

48 |
What is an adjoint model
- Errico
- 1997
(Show Context)
Citation Context ...omatic adjoint’ method [37]. The principal application of adjoint models is in sensitivity analysis, and further details on the derivation, properties, and applications of adjoint models are given in =-=[27]-=-, [28]. In Chapter 6 we comment on how the TLM and adjoint code are constructed for a one-dimensional shallow water model, and describe the coding tests needed to determine their suitability for inclu... |

47 |
N.: Introduction to Atmospheric Radiation
- Liou
- 2002
(Show Context)
Citation Context ...or temperature sounding and the retrieval of N2O and CO column amounts. IASI measurements of radiances, r, are expressed as black-body equivalent brightness temperatures, T, through Planck’s function =-=[59]-=- r = 2hν3c exp{ hcν kT } − 1, where k is Boltzmann’s constant, h is Planck’s constant, c is the speed of light and ν is the wavenumber. Planck’s function is used in the radiative transfer equation (2.... |

40 |
Dynamic data assimilation: a least squares approach. Cambridge Univ Pr
- Lewis, Lakshmivarahan, et al.
- 2006
(Show Context)
Citation Context ... (2.1) and (2.2). This analysis is sometimes known as the maximum a posteriori estimate and can be derived in terms of probability distribution functions (pdfs) using Bayesian methods. Bayes’ theorem =-=[56]-=- states that the posterior probability of event A, given that event B occurs, is proportional to the prior probability of A multiplied by the probability of event B given that event A occurs; P(A|B) ∝... |

30 |
Four-dimensional assimilation in the presence of baroclinic instability
- Rabier, Courtier
- 1992
(Show Context)
Citation Context ...wo tests: the correctness test and the validity test. The correctness test checks if the evolution of a perturbation in the TLM is comparable with that of the same perturbation in the nonlinear model =-=[73]-=-. The validity test identifies the time window for which the assumption of linearity in the TLM is valid. Details on the successful application of these tests to the TLM in the inherited SWM code are ... |

22 |
Data assimilation concepts and methods
- Bouttier, Courtier
- 1999
(Show Context)
Citation Context ..., and the optimality of the solution generated are all important issues in their operational use. Generally these techniques can be classified as sequential or variational, intermittent or continuous =-=[8]-=-. Sequential assimilation algorithms solve the system of equations needed for an optimal solution explicitly; variational algorithms solve the equations implicitly through the minimisation of a cost f... |

20 | An investigation of incremental 4d-var using non-tangent linear models
- Lawless, Gratton, et al.
(Show Context)
Citation Context ...ajectory. The iterative minimisation procedure is known as the inner loop; the update step is known as the outer loop. Full details of the procedure are described in the following iterative algorithm =-=[52]-=- where k is the iteration number: 1. At the first timestep (k = 0) define the current guess x (0) 0 = xb . 2. Run the nonlinear model to calculate x (k) i at each time step i. 3. Calculate the innovat... |

20 |
Iterative Methods for Toeplitz Systems
- Ng
- 2004
(Show Context)
Citation Context ...olve the resultant Toeplitz systems. In this work we are not concerned with solving general Toeplitz systems ex48plicitly, but a detailed discussion of the iterative techniques available is given in =-=[69]-=-. 3.2.2 Circulant matrices A circulant matrix is a Toeplitz matrix where each column is a circular shift of its preceding column. A circulant matrix C can be written in the form ⎛ C = ⎜ ⎝ c0 c1 c2 . .... |

18 |
A new dynamical core for the Met Office’s global and regional modelling of the atmosphere
- Davies, Cullen, et al.
- 2005
(Show Context)
Citation Context ...s are discretised using a two-time-level semi-implicit, semi-Lagrangian scheme (SISL). The SISL scheme is chosen to match closely the numerical integration scheme used operationally at the Met Office =-=[22]-=-. In a Lagrangian scheme the advection in a shallow water system is studied by tracking the position of a set of water parcels. A set of originally regularly spaced parcels at one time step may evolve... |

18 |
Sensitivity analysis using an adjoint of the psu-ncar mesoscale model
- Errico, Vukicevic
- 1992
(Show Context)
Citation Context ... adjoint’ method [37]. The principal application of adjoint models is in sensitivity analysis, and further details on the derivation, properties, and applications of adjoint models are given in [27], =-=[28]-=-. In Chapter 6 we comment on how the TLM and adjoint code are constructed for a one-dimensional shallow water model, and describe the coding tests needed to determine their suitability for inclusion i... |

17 |
W.: The Met. Office global three-dimensional variational data assimilation scheme, Quart
- Lorenc, Ballard, et al.
(Show Context)
Citation Context ...refore an implicit variational alternative is needed. The statistical method used operationally by the UK Met Office between 1999 and 2004 was three-dimensional variational data assimilation (3D-Var) =-=[63]-=-. This algorithm seeks the minimum of (2.8) by performing several evaluations of the cost function and its gradient in order to approach the minimum using a suitable descent algorithm [17]. The size o... |

17 |
Three- and four-dimensional variational assimilation with an ocean general circulation model of the tropical pacific ocean : Part i. formulation, internal diagnostics and concistency checks
- Weaver, Viliard, et al.
(Show Context)
Citation Context ...servations as if they were valid at the same point in time, which is clearly an unrealistic assumption. An extension to the standard 3D-Var is the First Guess at the Appropriate Time (3D-FGAT) method =-=[91]-=-. This technique calculates the observation increments (y − h(x)) at their appropriate measurement times, but then applies the increments at a single analysis time. A further sophisticated extension o... |

16 |
The Met Office global four-dimensional variational data assimilation scheme
- Rawlins, Ballard, et al.
(Show Context)
Citation Context ...lation window (JB term) plus the squared distance between the observations (red stars) and the forecast state throughout the assimilation window (JO,i terms). Although demonstrably superior to 3D-Var =-=[76]-=-, the size of the problem in NWP inhibits the direct solution of the 4D-Var cost function (2.12). When 4D-Var methods were originally being investigated it was determined that in order to use a 4D-Var... |

13 |
Development of a four-dimensional variational analysis system using the adjoint method at GLA. Part I: Dynamics
- Chao, Chang
- 1992
(Show Context)
Citation Context ...olves over time under a linear approximation. By treating the TLM code as a sequence of linear operations, the adjoint model code can be derived directly. The details of this process are described in =-=[12]-=-. 6.2 Data assimilation system The data assimilation system in which we will run the SWEs is the incremental 4D-Var method described in Section 2.3.1. This method simplifies the minimisation problem p... |

12 |
Diagnosis of observation, background and analysiserror statistics in observation
- Desroziers, Berre, et al.
- 2005
(Show Context)
Citation Context ...ite instrument, and describe how its measurements are processed in the Met Office incremental 4D-Var assimilation scheme. Using a postanalysis diagnostic based on variational data assimilation theory =-=[25]-=- and statistics from the Met Office system, we successfully quantify the cross-channel error correlations between IASI measurements. Diagnosed error covariances are given for the pre-processing 51D-V... |

12 |
S.P.Ballard. A comparison of two methods for developing the linearization of a shallow-water model
- Lawless, Nichols
(Show Context)
Citation Context ... the results in Chapter 5, Chapter 6 describes the mathematical framework needed to extend this investigation to a 4D-Var setting. We introduce a set of one-dimensional shallow water equations (SWEs) =-=[54]-=-, used to represent simplified atmospheric dynamics, and describe the continuous analytical and discretised numerical models. We then develop a new incremental 4D-Var data assimilation system for the ... |

11 |
Nonlinear shallow fluid flow over an isolated ridge
- Houghton, Kasahara
- 1968
(Show Context)
Citation Context ... described here, the added time dimension leads 116to a two-dimensional problem. The one-dimensional model has previously been used to represent atmospheric phenomena such as air flow over mountains =-=[48]-=-, and practical problems such as hydraulic flow in power plants [11]. A thorough description of inviscid multi-dimensional shallow water theory is given in [71]. 6.1.1 The continuous analytical model ... |

10 | 2001: Diagnosis and adaptive tuning of observation-error parameters in a variational assimilation. Quart - Desroziers, Ivanov |

10 |
An Experiment in Objective Analysis
- Gilchrist, Cressman
- 1954
(Show Context)
Citation Context ...iently for large scale problems. The complexity of data assimilation in NWP has increased significantly since the first objective analysis algorithms were introduced by Gilchrist and Cressman in 1954 =-=[38]-=-. The Cressman analysis scheme set the analysis equal to a background state plus a weighted contribution of the observations dependent on their spatial proximity to model grid points. The background s... |

10 |
M.H.Wright. Practical Optimization
- Gill
- 1981
(Show Context)
Citation Context ...the computed solution is to the ‘true’ solution of the problem. By testing the gradient ∇J (k) at the converged solution of the kth outer loop, we can determine the limiting accuracy in this solution =-=[39]-=-. The best numerical accuracy we would expect in the converged solution x is the same order as the normalised gradient ‖∇J(k) (x)‖2 |J(k)(x)| . Experiments showed that as we decrease the 127outer tol... |

10 |
P.: Eight Years of High Cloud Statistics Using HIRS
- Wylie, Menzel
- 1999
(Show Context)
Citation Context ...ed radiances, will only become possible operationally once these error characteristics can be better represented. Assimilation of ‘cloudy’ radiances Approximately 70% of the globe is covered by cloud =-=[93]-=-, and therefore much of the satellite data available to NWP centres contains contributions from cloud. Because of the highly nonlinear relationship between satellite retrievals and cloud properties, d... |

9 |
Statistics for technology
- Chatfield
- 1983
(Show Context)
Citation Context ...r cross-covariances on the off-diagonals. A variance is defined as the mean-square deviation about the mean of the error data; a covariance is a measure of the association between two error variables =-=[10]-=-. In operational systems the error covariance matrix must be symmetric positivedefinite. Consider a simple 2 × 2 case. Suppose we have a direct measurement y = (y1 y2) T of a variable x = (x1 x2) T , ... |

9 | Correlated observation errors in data assimilation
- Stewart, Dance, et al.
(Show Context)
Citation Context ...matrix. In this work both information measures were found to be significantly larger when the full error covariance matrix was used in preference to a diagonal approximation of the same variances. In =-=[85]-=- approaches A1, A2 and A3, described in Section 3.5, were used to evaluate information content under different diagonal and eigendecompostion approximations to a SOAR distributed error correlation mat... |

8 |
Inner-loop stopping criteria for incremental four-dimensional variational data assimilation
- Lawless, Nichols
(Show Context)
Citation Context ...te gradient method [36]. Because the minimisation is essentially an approximation to the full nonlinear cost function, we need not solve it too accurately, but it needs to have converged sufficiently =-=[53]-=-. Several stopping criteria options exist for the conjugate gradient minimisation, and a good review of their use in a similar model framework is given in [53]. In these experiments we will use the re... |

7 |
IASI: An advanced sounder for operational meteorology
- Chalon, Cayla, et al.
- 2001
(Show Context)
Citation Context ...d Atmospheric Sounding Interferometer (IASI) data The IASI instrument is an infrared Fourier transform spectrometer which measures the infrared radiation emitted by the earth’s surface and atmosphere =-=[9]-=-. The first IASI 68instrument was launched on the MetOp-A satellite in 2006 as part of the EUMETSAT European Polar System (EPS). Its spectral interval of 645-2760cm−1 is divided into three bands and ... |

7 |
The optimal density of atmospheric sounder observations in the met office NWP system
- Dando, Thorpe, et al.
(Show Context)
Citation Context ...erage as a single observation value. Under such assumptions, increasing the observation density beyond some threshold value has been shown to yield little or no improvement in analysis accuracy [60], =-=[21]-=-. Although discarding available information may be appropriate when the spatial resolution of the observations is denser than the model grid, recent technological advances have challenged the practica... |

7 |
AIRS near real-time products and algorithms in support of operational numerical weather prediction
- Goldberg, Qu, et al.
(Show Context)
Citation Context ...ASI observations of 8461 channels) by a new correlated set of data vectors containing fewer elements. The procedure retains most of the variability and information of the initial data. Goldberg et al =-=[40]-=- demonstrated that PCA produces an efficient retrieval of atmospheric temperature, moisture and ozone, and an accurate reconstruction of over 2000 AIRS channels from 60 principal component scores. Als... |

7 |
Use of discrete Fourier transforms in the 1D-Var retrieval problem
- Healy, White
(Show Context)
Citation Context ...n alternative approach to dealing with observation error correlations is needed. 2Approximating observation error correlation is a relatively new direction of research but progress has been made. In =-=[43]-=- circulant matrices were used to approximate a Toeplitz observation error covariance matrix. Results showed that incorrectly assuming uncorrelated observation errors gave misleading estimates of infor... |

7 | 2002: Channel selection methods for Infrared Atmospheric Sounding Interferometer radiances
- Rabier, Fourrié, et al.
(Show Context)
Citation Context ...h to accurately represent atmospheric variability but small enough to be assimilated efficiently within NWP systems. Specific methods of channel selection based on objective criteria are described in =-=[74]-=-. The iterative method proposed by Rodgers [77] was found to generate the best results in respect to the lowest standard deviations of errors over the vertical profile. This method takes entropy reduc... |

6 |
On the choice of observation errors for the assimilation of AIRS brightness temperatures: A theoretical study
- Collard
(Show Context)
Citation Context ...ss the second question posed in Section 1.2 and present possible matrix representations of the observation error covariance structure. Three different approximating structures are described: diagonal =-=[14]-=-, circulant [78], [43], and eigendecomposition [34] approximations. We pay careful attention to the feasiblity of including these matrices operationally. The effectiveness of each approximation can be... |

6 |
Estimation of entropy reduction and degrees of freedom for signal for large variational analysis systems
- Fisher
(Show Context)
Citation Context ... the information provided by an observation or an observing system is used in the development of satellite instruments [74], [89] as well as in the assessment of operational data assimilation systems =-=[33]-=-. The two most frequently used information measures are entropy reduction (or Shannon Information Content) and degrees of freedom for signal. These measures are generally applicable and can be defined... |

6 |
Development of linear models for data assimilation in numerical weather prediction
- Lawless
- 2001
(Show Context)
Citation Context ...In Section 2.3.1 we described how an incremental 4D-Var assimilation requires a linear approximation to a nonlinear model. The linear approximation used in this work is the tangent linear model (TLM) =-=[50]-=-. To generate the TLM we assume that a nonlinear model and its linearised version exhibit similar behaviour for a period known as the validity time. We then consider the nonlinear model mi applied to ... |

6 |
Four-dimensional data assimilation using the adjoint of a multi-level primitive-equation
- Thepaut, Courtier
- 1225
(Show Context)
Citation Context ...del Mi, mi(x + δx) = mi(x) + Mi(x)δx. (2.20) 18Adjoint model The adjoint model MT provides us with a system of model equations, solvable backwards in time to obtain the gradient of the cost function =-=[90]-=-. In practice the discrete adjoint equations are constructed directly from the tangent linear model code using an ‘automatic adjoint’ method [37]. The principal application of adjoint models is in sen... |

5 |
Accounting for correlated observation error in the ECMWF analysis
- Fisher
- 2005
(Show Context)
Citation Context ... were used to approximate a Toeplitz observation error covariance matrix. Results showed that incorrectly assuming uncorrelated observation errors gave misleading estimates of information content. In =-=[34]-=- Fisher proposed assigning a block-diagonal structure to the observation error covariance matrix, with (uncorrelated) blocks corresponding to different instruments or channels. Using this technique, i... |

5 | Characteristics of high resolution versions of the Met. Office Unified Model for forecasting convection over the UK, Monthly Weather Review - Lean, Clark, et al. |

5 |
An improved general fast radiative transfer model for the assimilation of radiance observations
- Matricardi, Chevallier, et al.
(Show Context)
Citation Context ...n is used in the radiative transfer equation (2.21) described in Secton 2.4.1. The radiative transfer model used in the assimilation of IASI radiances is the radiative transfer model for TOVS (RTTOV) =-=[66]-=-. IASI data is an important component of the global observing system. The assimilation of IASI radiances is operational at the Met Office [45], Meteo-France [72] and ECMWF [13], and is in the testing ... |