Objective Analysis, Initialization, Data Assimilation and Satellite Data
CSIR NET meteorology: Cressman and optimal interpolation, normal-mode and digital-filter initialization, 3D-Var, 4D-Var, Kalman filters and satellite data.
- 21 min
On this page · 14 sections
Prerequisites
- Least squares, weighting by error variance and the minimum-variance estimate: Well- and Ill-Posed Problems and Least Squares.
- Gradient-based minimisation and adjoints in outline: Gradient and Direct Search, Generalized Inversion, SVD and Global Optimization.
- Radiative transfer, absorption bands and weighting functions for satellite sounding: Radiation Laws, Scattering, Albedo, Radiative Transfer and the Greenhouse Effect.
- Gravity waves and the meaning of “balance”: Computational Instability and Filtering.
- Chapter map: Numerical Weather Prediction.
The Hook: Richardson’s Real Problem
The Analysis Problem
An analysis is an estimate of the true state of the atmosphere on the model grid at a given time. Its ingredients:
| Ingredient | Symbol | Error statistics |
|---|---|---|
| Background (first guess): a short forecast from the previous analysis | \( \mathbf{x}_b \) | covariance \( \mathbf{B} \) |
| Observations | \( \mathbf{y} \) | covariance \( \mathbf{R} \) (instrument + representativeness error) |
| Observation operator: maps a model state to what the instrument would see (interpolation, or a radiative-transfer model for radiances) | \( H \) | |
| Analysis | \( \mathbf{x}_a \) | covariance \( \mathbf{A} \) |
The innovation (observation-minus-background, O − B) is \( \mathbf{y} - H(\mathbf{x}_b) \); the analysis increment is \( \mathbf{x}_a - \mathbf{x}_b \).
Representativeness error is part of \( \mathbf{R} \): a point measurement samples scales the model grid cannot represent (a station in a valley, a single cloud), so even a perfect instrument disagrees with a perfect model value.
The cycle repeats:
graph LR
A["Background
(short forecast)"] --> B["Analysis
(background + observations)"]
O["Observations
quality controlled"] --> B
B --> C["Initialization
(remove imbalance)"]
C --> D["Forecast model"]
D -->|"6 h forecast"| A
D --> E["Forecast products"]This loop of background, analysis and forecast is data assimilation. The background carries information from all earlier observations forward in time, so data-sparse regions (oceans, the southern hemisphere) still get a reasonable analysis.
Objective Analysis: From Hand Analysis to Interpolation
Before NWP, analyses were drawn by hand. Objective analysis replaces the analyst by a reproducible algorithm. The main historical steps:
| Method | Idea | Weakness |
|---|---|---|
| Polynomial fitting (earliest) | fit a local polynomial (often a quadratic) to observations around each grid point by least squares | unstable in data-sparse areas; ignores a first guess |
| Successive correction (Bergthorsson–Döös, Cressman) | start from a first guess; add weighted averages of the innovations within a radius of influence; repeat with smaller radii | weights are empirical; ignores observation error and data clustering |
| Optimal interpolation (OI), statistical interpolation (Gandin) | choose weights that minimise the expected analysis error variance, from the error covariances of background and observations | needs covariance models; local data selection; usually treats \( H \) as linear |
| Variational (3D-Var, 4D-Var) | minimise a cost function measuring misfit to background and observations | needs \( \mathbf{B} \) and, for 4D-Var, an adjoint model |
| Kalman filter and ensemble methods | propagate the error covariances with the model | very costly in full form; ensembles approximate it |
Cressman successive correction
At a grid point \( g \), with background \( x_b \) and observations \( k \) at distance \( r_k \) from \( g \):
$$x_a = x_b + \frac{\sum_k w_k\,(y_k - x_{b,k})}{\sum_k w_k}, \qquad w_k = \begin{cases} \dfrac{R^2 - r_k^2}{R^2 + r_k^2}, & r_k < R \\[4pt] 0, & r_k \ge R \end{cases}$$where \( x_{b,k} \) is the background interpolated to observation \( k \), and \( R \) is the radius of influence. A weight of 1 applies at the grid point itself, falling to 0 at \( R \). Several passes are made with decreasing \( R \) (for example from several hundred to about a hundred kilometres), so large scales are fitted first and smaller scales later. The Barnes scheme is a variant with Gaussian weights \( w = e^{-r^2/\kappa} \).
Optimal interpolation
For one grid value, one observation at the same point, and uncorrelated errors with variances \( \sigma_b^2 \) and \( \sigma_o^2 \), minimising the expected squared analysis error gives
$$\boxed{x_a = x_b + \frac{\sigma_b^2}{\sigma_b^2 + \sigma_o^2}(y - x_b), \qquad \sigma_a^2 = \frac{\sigma_b^2\,\sigma_o^2}{\sigma_b^2 + \sigma_o^2}}$$Equivalently \( \frac{1}{\sigma_a^2} = \frac{1}{\sigma_b^2} + \frac{1}{\sigma_o^2} \): precisions add, so the analysis is always more accurate than either source alone. The weight given to the observation rises as the background becomes less reliable.
In matrix form, for many observations and grid points,
$$\mathbf{x}_a = \mathbf{x}_b + \mathbf{K}\left[\mathbf{y} - H\mathbf{x}_b\right], \qquad \mathbf{K} = \mathbf{B}H^{\mathsf T}\left(H\mathbf{B}H^{\mathsf T} + \mathbf{R}\right)^{-1},$$the best linear unbiased estimate (BLUE). \( \mathbf{K} \) is the gain. The background-error covariance \( \mathbf{B} \) does the essential work: it spreads the information from an observation to nearby points and to other variables (a height observation updates the wind through geostrophically balanced covariances), and it builds balance into the increments. That is why multivariate OI reduced, but did not eliminate, the need for separate initialization.
Initialization
Why it is needed
Any analysis contains small imbalances between the wind and mass fields. In a primitive-equation model they project onto inertia–gravity waves with large amplitudes and high frequencies. The symptoms are spurious oscillations of surface pressure in the first hours of the forecast, corrupted rainfall and vertical velocity during the “spin-up” period, and noisy short-range forecasts that then serve as the next background. Initialization removes the unbalanced part while keeping the slow, meteorological (Rossby) part.
Methods
| Method | How it works | Notes |
|---|---|---|
| Static initialization | Derive the wind from the mass field with a balance relation (geostrophic wind, or the nonlinear balance equation), or the reverse | Simple; geostrophy is poor in the tropics; discards observed wind information |
| Dynamic initialization | Integrate the model forward and backward about the initial time with a scheme that damps high frequencies (for example Matsuno, Euler-backward) | Uses the model’s own dynamics; costly; also damps some meteorological signal |
| Linear normal-mode initialization | Expand the analysis in the model’s linear normal modes (slow Rossby modes and fast gravity modes); set the gravity-mode amplitudes to zero | Removes gravity waves at time zero, but nonlinear terms regenerate them |
| Nonlinear normal-mode initialization (NNMI) | Set the initial tendencies of the gravity modes to zero (Machenhauer’s condition), iteratively | Keeps the gravity modes that the slow flow forces; the operational standard of the 1980s |
| Digital filter initialization (DFI) | Integrate the model a few hours forward and backward around the initial time and apply a low-pass time filter to the sequence of states | Simple, uses the full model and physics; widely used in regional models |
| Implicit (in the assimilation) | Balance constraints in \( \mathbf{B} \), a penalty term in the cost function (\( J_c \)), or 4D-Var fitting observations over a window | Modern variational systems need little or no separate initialization |
Variational Data Assimilation
3D-Var
3D-Var finds the state \( \mathbf{x} \) that minimises
$$\boxed{J(\mathbf{x}) = \tfrac{1}{2}(\mathbf{x} - \mathbf{x}_b)^{\mathsf T}\mathbf{B}^{-1}(\mathbf{x} - \mathbf{x}_b) + \tfrac{1}{2}\left[\mathbf{y} - H(\mathbf{x})\right]^{\mathsf T}\mathbf{R}^{-1}\left[\mathbf{y} - H(\mathbf{x})\right]}$$The first term \( J_b \) penalises departure from the background, the second \( J_o \) misfit to observations, each weighted by the inverse of its error covariance. For linear \( H \) and Gaussian errors the minimum is the same as the OI / BLUE solution. The advantages of 3D-Var over OI are practical:
- all observations are used together in one global minimisation (no local data selection);
- nonlinear observation operators are allowed, so satellite radiances can be assimilated directly through a radiative-transfer model rather than through retrieved temperature profiles;
- constraints (balance, smoothness) can be added as extra terms.
The minimum is found iteratively with the gradient \( \nabla J = \mathbf{B}^{-1}(\mathbf{x} - \mathbf{x}_b) - H^{\mathsf T}\mathbf{R}^{-1}[\mathbf{y} - H(\mathbf{x})] \), using conjugate-gradient or quasi-Newton methods.
4D-Var
4D-Var extends the cost function over an assimilation window (commonly 6 or 12 hours). It seeks the initial state \( \mathbf{x}_0 \) whose model trajectory best fits all observations at their correct times:
$$J(\mathbf{x}_0) = \tfrac{1}{2}(\mathbf{x}_0 - \mathbf{x}_b)^{\mathsf T}\mathbf{B}^{-1}(\mathbf{x}_0 - \mathbf{x}_b) + \tfrac{1}{2}\sum_{i=0}^{n}\left[\mathbf{y}_i - H_i(\mathbf{x}_i)\right]^{\mathsf T}\mathbf{R}_i^{-1}\left[\mathbf{y}_i - H_i(\mathbf{x}_i)\right], \quad \mathbf{x}_i = M_{0\to i}(\mathbf{x}_0).$$- The model \( M \) is a strong constraint: the analysis is a model solution, so it is dynamically consistent and balanced.
- The gradient is computed with the adjoint of the (linearised) model, run backward in time over the window.
- Observations are used at their actual times, which suits asynoptic data (satellites, aircraft).
- The background-error covariance evolves implicitly through the window (“flow-dependent” structure functions).
- It is expensive; operational systems use an incremental formulation with a simplified, lower-resolution linear model for the inner loops.
Kalman filters and ensembles
The Kalman filter is the sequential form of BLUE that also forecasts the error covariance:
$$\mathbf{x}_f^{\,t+1} = M\mathbf{x}_a^{\,t}, \qquad \mathbf{P}_f^{\,t+1} = \mathbf{M}\mathbf{P}_a^{\,t}\mathbf{M}^{\mathsf T} + \mathbf{Q},$$followed by the analysis step with \( \mathbf{K} = \mathbf{P}_f H^{\mathsf T}(H\mathbf{P}_f H^{\mathsf T} + \mathbf{R})^{-1} \) and \( \mathbf{P}_a = (\mathbf{I} - \mathbf{K}H)\mathbf{P}_f \). For a model with \( 10^8 \) or more variables, \( \mathbf{P} \) cannot even be stored. The ensemble Kalman filter (EnKF) estimates \( \mathbf{P}_f \) from an ensemble of forecasts, with localization (cutting off spurious long-distance correlations from a small ensemble) and inflation (counteracting the ensemble’s tendency to become over-confident). Hybrid systems blend a static \( \mathbf{B} \) with ensemble covariances in a variational analysis, and an ensemble of data assimilations also provides initial perturbations for ensemble forecasts.
Comparison
| OI | 3D-Var | 4D-Var | EnKF | |
|---|---|---|---|---|
| Time dimension | single time | single time (observations binned to analysis time) | window; observations at their own times | sequential |
| Background covariance | static, modelled | static | static at start, evolves implicitly over window | flow-dependent, from ensemble |
| Nonlinear \( H \) (radiances) | awkward | yes | yes | yes |
| Needs adjoint model | no | no | yes | no |
| Data selection | local | global | global | local (localization) |
Quality control and bias correction
Before any analysis, observations go through:
- gross error checks (physically impossible values), hydrostatic and consistency checks on radiosondes;
- background check: reject observations whose O − B departure is too large relative to expected error;
- buddy check: compare with neighbouring observations;
- thinning of dense data (satellites) to reduce correlated errors;
- bias correction, especially for satellite radiances, whose systematic biases (instrument calibration, radiative-transfer errors) can be larger than the signal. Modern systems estimate the bias coefficients inside the variational analysis (variational bias correction).
Satellite Data
Satellites now supply most of the observations used in global NWP, and are the only routine source over the oceans and the southern hemisphere.
Orbits
| Polar-orbiting (sun-synchronous) | Geostationary | |
|---|---|---|
| Altitude | low Earth orbit, typically about 800 km | about 35 800 km above the equator |
| Coverage | whole globe, each place about twice a day per satellite | fixed disc, up to about 60–70° latitude; no polar view |
| Repeat time | about 12 h per satellite at a place | minutes (rapid scans) |
| Strength | high-resolution sounders (IR and microwave), global coverage | continuous imaging: cloud motion, convection, rapid change |
| NWP use | temperature and humidity sounding radiances, scatterometer winds, radio occultation | atmospheric motion vectors, clear-sky radiances, nowcasting |
What satellites measure and how it is used
| Observation | Physics | Information |
|---|---|---|
| Infrared sounder radiances (hyperspectral sounders such as AIRS, IASI, CrIS; earlier HIRS) | emission in the CO₂ absorption band near 15 μm (temperature) and water-vapour bands | temperature and humidity profiles, mainly in clear sky |
| Microwave sounder radiances (AMSU-A, ATMS; humidity: AMSU-B, MHS) | emission by O₂ in the 50–60 GHz band (temperature) and H₂O near 183 GHz (humidity) | temperature and humidity profiles through most clouds |
| Microwave imagers | emission at window frequencies | total column water vapour, cloud liquid water, rain rate, surface wind speed over sea |
| Atmospheric motion vectors (AMVs) | tracking clouds and water-vapour features in successive geostationary images | wind at the height assigned to the feature (height assignment is the main error source) |
| Scatterometer winds (for example ASCAT; the Indian Oceansat-2 scatterometer and SCATSAT-1) | radar backscatter from wind-roughened sea surface | ocean surface wind vectors |
| GNSS radio occultation | bending of GPS signals passing through the atmosphere’s limb | refractivity, hence temperature in the upper troposphere and stratosphere; no calibration drift, so it anchors radiance bias correction |
| Ocean altimetry, SST | radar altimeter, IR and microwave SST | ocean initial states for coupled models |
Weighting functions and the retrieval problem
A sounder channel at frequency \( \nu \) measures radiance emitted from a layer of the atmosphere described by its weighting function \( W(p) = d\tau_\nu/d\ln p \), where \( \tau_\nu \) is the transmittance from level \( p \) to space. Channels near the centre of an absorption band peak high in the atmosphere (opaque, emission escapes only from near the top); channels in the wings peak lower. Each channel sees a broad layer, and neighbouring channels overlap, so recovering a temperature profile from radiances (retrieval) is an ill-posed inverse problem: many profiles give the same radiances.
Two ways to use them:
| Retrieval assimilation | Direct radiance assimilation | |
|---|---|---|
| What enters the analysis | retrieved temperature/humidity profiles | the radiances themselves |
| Observation operator | simple interpolation | a fast radiative-transfer model (and its adjoint) |
| Prior information | from the retrieval’s own first guess (often climatology or a different model), with complicated errors | from the NWP background, used consistently with \( \mathbf{B} \) |
| Status | older approach | standard in variational systems |
Direct assimilation is preferred because the errors of radiances are simpler and better known than those of retrievals, and because the background used is the same as in the rest of the analysis. Its difficulties are clouds and precipitation (IR channels see cloud tops; microwave channels are affected by rain), surface emissivity over land and snow, and bias correction.
Impact
Observing-system experiments (forecasts run with and without one type of data) show that satellite data, especially microwave and hyperspectral IR sounder radiances, are among the largest contributors to global forecast skill, with the largest impact in the southern hemisphere and over the oceans. The narrowing of the gap between northern and southern hemisphere skill is attributed largely to better use of satellite radiances. Over India and the surrounding seas, geostationary imagery and soundings from the INSAT-3D series, together with scatterometer and microwave data, fill the gaps between the sparse radiosonde network and the open ocean.
Conventional data for comparison
| Platform | Variables | Remarks |
|---|---|---|
| Radiosondes | \( T \), humidity, wind, pressure profiles | main synoptic hours 00 and 12 UTC; backbone of upper-air data over land |
| Surface (SYNOP), ships, buoys | pressure, temperature, humidity, wind | surface pressure is especially valuable |
| Aircraft (AMDAR) | temperature, wind; some humidity | dense along flight routes and near airports |
| Wind profilers, Doppler weather radar | wind profiles; reflectivity and radial wind | radar is central to convective-scale assimilation and nowcasting |
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Objective analysis (spatial interpolation at one time) | Data assimilation (analysis + forecast cycling in time) | |
| Innovation \( \mathbf{y} - H\mathbf{x}_b \) | Analysis increment \( \mathbf{x}_a - \mathbf{x}_b \) | |
| Linear NMI: gravity-mode amplitudes set to zero | Nonlinear NMI: gravity-mode tendencies set to zero | |
| Static initialization (balance relation) | Dynamic initialization (damping time integration) | |
| 3D-Var: one time, no adjoint model | 4D-Var: window, model as constraint, adjoint | |
| \( \mathbf{B} \): background-error covariance | \( \mathbf{R} \): observation-error covariance (includes representativeness) | |
| Retrieved profile | Radiance assimilated directly through a radiative-transfer operator | |
| Polar-orbiting, about 800 km, global | Geostationary, about 35 800 km, fixed disc, frequent images |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- What is the Cressman weight of an observation located exactly at the radius of influence?
- Which data assimilation method requires the adjoint of the forecast model?
- Which spectral band is used by microwave sounders to measure temperature profiles?
Answers
- Zero. \( w = (R^2 - R^2)/(R^2 + R^2) = 0 \).
- 4D-Var, to compute the gradient of the cost function with respect to the initial state.
- The oxygen absorption band at 50–60 GHz.
Level 2: Part B Style
Practice Q1. With a radius of influence \( R = 200 \) km, the Cressman weight of an observation 100 km from a grid point is:
(a) 0.25 (b) 0.50 (c) 0.60 (d) 0.75
Answer: (c) 0.60
\( w = (200^2 - 100^2)/(200^2 + 100^2) = 30\,000/50\,000 = 0.6 \). Option (d) would be the linear weight \( 1 - (r/R)^2 \); option (b) is \( 1 - r/R \).
Practice Q2. A background temperature of 20.0 °C has error standard deviation 1.0 °C. An observation at the same point reads 22.0 °C with error standard deviation 0.5 °C. Errors are unbiased and uncorrelated. The optimal analysis and its error standard deviation are:
(a) 21.0 °C, 0.75 °C (b) 21.6 °C, 0.45 °C (c) 22.0 °C, 0.50 °C (d) 21.6 °C, 0.20 °C
Answer: (b) 21.6 °C, about 0.45 °C
Weight \( = \sigma_b^2/(\sigma_b^2 + \sigma_o^2) = 1/(1 + 0.25) = 0.8 \). \( x_a = 20 + 0.8(2) = 21.6 \) °C.
\( \sigma_a^2 = (1)(0.25)/1.25 = 0.2 \), so \( \sigma_a = 0.447 \) °C. Option (d) quotes the variance as the standard deviation. The analysis error is smaller than both input errors.
Practice Q3. In nonlinear normal-mode initialization:
(a) The amplitudes of the Rossby modes are set to zero (b) The initial time tendencies of the gravity modes are set to zero (c) The model is integrated forward with the forward (Euler) scheme to amplify gravity waves (d) The wind is replaced by the geostrophic wind everywhere
Answer: (b)
NNMI (Machenhauer’s condition) sets the gravity-mode tendencies to zero, keeping the slowly varying gravity-mode component forced by the Rossby flow. (d) is static initialization; (a) would remove the weather itself.
Practice Q4. A geostationary satellite differs from a sun-synchronous polar-orbiting satellite in that it:
(a) Orbits at about 800 km and sees the whole globe twice a day (b) Orbits at about 35 800 km above the equator and images the same disc every few minutes, but cannot see the poles (c) Measures only microwave radiances (d) Provides GNSS radio-occultation profiles
Answer: (b)
Geostationary orbit is about 35 800 km above the equator, period one sidereal day; it gives frequent images of a fixed disc (hence atmospheric motion vectors) but poor views poleward of about 60–70°. (a) describes a polar orbiter.
Level 3: Part C Style
Practice Q5. A grid point has background value 10.0 (units of the field). Two observations lie within a Cressman radius of influence of 200 km: observation A at 50 km with innovation (O − B) of +4.0, and observation B at 150 km with innovation +2.0. The Cressman analysis at the grid point (single pass) is closest to:
(a) 12.0 (b) 13.0 (c) 13.5 (d) 14.0
Answer: (c) about 13.5
\( w_A = (40\,000 - 2\,500)/(40\,000 + 2\,500) = 37\,500/42\,500 = 0.882 \).
\( w_B = (40\,000 - 22\,500)/(40\,000 + 22\,500) = 17\,500/62\,500 = 0.280 \).
Correction \( = (0.882 \times 4 + 0.280 \times 2)/(0.882 + 0.280) = (3.529 + 0.560)/1.162 = 3.52 \).
\( x_a = 10 + 3.52 = 13.52 \). Option (b) is the simple mean of the innovations (3.0) added to 10; Cressman favours the nearer observation.
Practice Q6. At one grid point, a 3D-Var analysis combines a background temperature of 280 K (\( \sigma_b = 2 \) K) with two independent observations: 282 K (\( \sigma = 1 \) K) and 283 K (\( \sigma = 2 \) K), both at the grid point. Consider:
- The analysis is about 281.8 K.
- The analysis error variance is about 0.67 K².
- The analysis lies closer to the first observation than to the background because the first observation has the smallest error.
- Removing the background term from the cost function would leave the analysis unchanged.
Which are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 1 and 4 only (d) 1, 2, 3 and 4
Answer: (b) 1, 2 and 3 only
Minimising \( J = \frac{(x - 280)^2}{2 \cdot 4} + \frac{(x - 282)^2}{2 \cdot 1} + \frac{(x - 283)^2}{2 \cdot 4} \) gives the precision-weighted mean:
\( x_a = \dfrac{280/4 + 282/1 + 283/4}{1/4 + 1 + 1/4} = \dfrac{70 + 282 + 70.75}{1.5} = 281.83 \) K. (1 true)
\( \sigma_a^2 = 1/1.5 = 0.667 \) K², \( \sigma_a = 0.82 \) K. (2 true)
281.83 is 0.17 K from 282 and 1.83 K from 280. (3 true)
Without \( J_b \): \( x = (282/1 + 283/4)/(1 + 1/4) = 352.75/1.25 = 282.2 \) K, a different answer. (4 false)
Practice Q7. Consider these statements about initialization and satellite data:
- Digital filter initialization applies a low-pass filter in time to a short model integration around the initial time.
- 4D-Var reduces the need for separate initialization because its analysis is a model trajectory fitted over a window.
- Direct radiance assimilation needs a fast radiative-transfer model as the observation operator.
- Retrieving a temperature profile from sounder radiances is a well-posed problem because each channel senses a thin layer.
Which are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) 1, 2, 3 and 4
Answer: (b) 1, 2 and 3 only
1, 2 and 3 are true as described on this page. 4 is false: weighting functions are broad and overlapping, so the inverse problem is ill-posed; prior information (a background) is needed to choose among profiles that fit the radiances.
Quick Revision Table
| Idea | Formula or fact |
|---|---|
| Analysis ingredients | background \( \mathbf{x}_b \) (\( \mathbf{B} \)), observations \( \mathbf{y} \) (\( \mathbf{R} \)), operator \( H \) |
| Cressman weight | \( (R^2 - r^2)/(R^2 + r^2) \), 0 beyond \( R \); successive passes, shrinking \( R \) |
| Barnes weight | \( e^{-r^2/\kappa} \) |
| OI (scalar) | \( x_a = x_b + \frac{\sigma_b^2}{\sigma_b^2 + \sigma_o^2}(y - x_b) \); \( 1/\sigma_a^2 = 1/\sigma_b^2 + 1/\sigma_o^2 \) |
| BLUE gain | \( \mathbf{K} = \mathbf{B}H^{\mathsf T}(H\mathbf{B}H^{\mathsf T} + \mathbf{R})^{-1} \) |
| 3D-Var | \( J = J_b + J_o \); same minimum as OI for linear \( H \); allows radiances |
| 4D-Var | window, model as strong constraint, adjoint for the gradient, incremental form |
| Kalman / EnKF | forecasts \( \mathbf{P} \); EnKF uses ensemble with localization and inflation |
| Static / dynamic initialization | balance relation / damping forward-backward integration |
| Linear vs nonlinear NMI | gravity amplitudes = 0 / gravity tendencies = 0 (Machenhauer) |
| DFI | low-pass time filter on a short integration |
| Polar vs geostationary | about 800 km, global / about 35 800 km, fixed disc, frequent |
| Temperature sounding | CO₂ 15 μm (IR), O₂ 50–60 GHz (microwave, through cloud) |
| Humidity sounding | H₂O 6.3 μm (IR), 183 GHz (microwave) |
| Other satellite data | AMVs, scatterometer winds, GNSS RO, microwave imager TPW and rain |
| Weighting function | \( d\tau/d\ln p \); broad, overlapping ⇒ retrieval ill-posed |
Key Takeaways
What’s Next
- Numerical Weather Prediction: Quick Revision: the whole chapter on one page.
- Short-, Medium- and Long-Range Prediction: how analysis quality turns into forecast skill at each range.
- Inverse-problem tools behind 3D-Var and retrievals: Gradient and Direct Search, Generalized Inversion, SVD and Global Optimization.
- Back to the Earth Sciences overview.