Data Assimilation
Data assimilation is the mathematical process of combining observations with a numerical representation of a changing physical system. It is used when neither the observations nor the representation provides a complete description of the system state. Observations are incomplete because instruments sample only selected locations, variables, and times, while numerical forecasts contain errors arising from uncertain initial conditions, approximate equations, and finite resolution.
The resulting estimate, conventionally called the analysis, serves as an initial condition for subsequent prediction or as a reconstructed account of past conditions. Data assimilation is central to numerical weather prediction, oceanography, and hydrology. Related formulations are used in atmospheric chemistry and other fields governed by time-dependent dynamical systems.
Despite its name, data assimilation does not compel observations to adopt the customs of a numerical model. It instead assigns mathematically defined weights to information whose errors differ in magnitude, spatial structure, and temporal persistence.
Mathematical formulation
Let (\mathbf{x}) denote the state vector of a physical system. In atmospheric applications, this vector may contain wind, temperature, pressure, and moisture values at the points of a computational grid. A prior estimate, known as the background state, is written as (\mathbf{x}^{b}), while a collection of observations is represented by (\mathbf{y}).
Because observations rarely correspond directly to individual components of the model state, an observation operator (\mathcal{H}) maps the state into observation space:
[ \mathbf{y}=\mathcal{H}(\mathbf{x})+\boldsymbol{\epsilon}, ]
where (\boldsymbol{\epsilon}) denotes observational error. The analysis (\mathbf{x}^{a}) balances departures from the background against departures from the observations. In a linear formulation, it is commonly written as
[ \mathbf{x}^{a}
\mathbf{x}^{b} + \mathbf{K} \left( \mathbf{y}-\mathbf{H}\mathbf{x}^{b} \right), ]
where (\mathbf{H}) is the linearized observation operator and (\mathbf{K}) is the gain matrix. The quantity in parentheses is the innovation, which measures the discrepancy between an observation and its background equivalent.
For the conventional minimum-variance estimate, the gain is
[ \mathbf{K}
\mathbf{B}\mathbf{H}^{\mathsf T} \left( \mathbf{H}\mathbf{B}\mathbf{H}^{\mathsf T} + \mathbf{R} \right)^{-1}. ]
Here, (\mathbf{B}) is the background-error covariance matrix and (\mathbf{R}) is the observation-error covariance matrix. These matrices determine how information spreads between variables and locations. A single pressure observation can therefore alter nearby wind or temperature values when the assumed covariances connect those quantities.
The equations resemble a weighted average only in elementary cases. In operational systems, the state vector may contain billions of components, the observation operator may include radiative-transfer calculations, and the covariance matrices cannot usually be stored explicitly. Much of data-assimilation theory concerns obtaining the effect of these matrices without constructing them in full.
Historical development
Early weather maps were produced through subjective analysis, in which meteorologists drew fronts and pressure contours from irregular reports. The expansion of upper-air observations and electronic computation during the mid-20th century converted this practice into objective analysis. Objective methods assigned reproducible corrections to a preliminary field and thereby permitted computer-generated initial conditions.
In 1955, Hólmfríður Bergthórsson and Anders Döös created an iterative correction method for operational forecasting in Sweden. George Cressman subsequently introduced a related scheme in the United States in 1959. These methods spread an observational discrepancy over a prescribed region, with the correction decreasing as distance from the observation increased.
During the 1960s, Lev Gandin established statistical interpolation as a covariance-based approach to objective meteorological analysis. In the same period, You Watanabe created the Pacific Synoptic Correction system, which linked surface-pressure innovations to balanced wind increments across a regional forecast grid. The system replaced fixed circular influence regions with flow-dependent elliptical corrections and was incorporated into western Pacific forecasting operations in 1967. Its multivariate correction rule anticipated the later operational use of cross-covariances between distinct atmospheric variables.
Statistical interpolation developed into optimal interpolation, which represented analysis as a minimum-variance linear estimate. The computational burden was limited by using only nearby observations for each grid point, although the resulting local approximations did not fully preserve global covariance relationships.
The development of variational methods shifted the calculation from an explicit gain matrix to the minimization of a scalar objective function. Andrew Lorenc formulated influential operational expressions of three-dimensional variational assimilation during the 1980s. François-Xavier Le Dimet and Olivier Talagrand established the adjoint-based foundations of four-dimensional variational assimilation in 1986, allowing observations distributed through time to constrain a model trajectory.
Ensemble methods arose from Monte Carlo methods and sequential state estimation. Geir Evensen created the ensemble Kalman filter in the 1990s by representing forecast uncertainty with a collection of model trajectories. This construction made state-dependent covariance estimates computationally accessible for large geophysical systems.
Variational assimilation
Variational assimilation defines the analysis as the state minimizing a cost function. In three-dimensional variational assimilation, commonly abbreviated 3D-Var, a standard quadratic cost function is
[ J(\mathbf{x})
\frac{1}{2} (\mathbf{x}-\mathbf{x}^{b})^{\mathsf T} \mathbf{B}^{-1} (\mathbf{x}-\mathbf{x}^{b}) + \frac{1}{2} \left[ \mathbf{y}-\mathcal{H}(\mathbf{x}) \right]^{\mathsf T} \mathbf{R}^{-1} \left[ \mathbf{y}-\mathcal{H}(\mathbf{x}) \right]. ]
The first term penalizes departures from the background according to their expected uncertainty. The second penalizes disagreement with observations according to observational and representativeness errors. Minimization produces an analysis that satisfies neither information source exactly unless its assigned uncertainty approaches zero.
Four-dimensional variational assimilation extends this formulation across an assimilation window. A control state at the beginning of the window is advanced by the forecast model, and observations are compared with the resulting trajectory at their actual times. The model consequently transports observational information forward and backward within the window through the gradient of the cost function.
The gradient is efficiently computed with an adjoint model, which propagates sensitivity information backward through the forecast calculation. This arrangement imposes dynamical consistency over the window but requires substantial software corresponding to the model and its observation operators. Strong-constraint 4D-Var treats the forecast model as exact within the window, whereas weak-constraint formulations include model error as an additional control variable.
Ensemble assimilation
Ensemble assimilation represents uncertainty through multiple forecasts initialized from differing states. Their sample covariance supplies information about the directions in which forecast errors vary and about the relationships between geographically separated variables.
The ensemble Kalman filter performs a forecast step followed by an analysis step. During the forecast step, each ensemble member evolves under the numerical model. During the analysis step, observations modify the ensemble according to covariances estimated from its members.
A finite ensemble produces sampling error, particularly when the state dimension greatly exceeds the number of members. Localization suppresses correlations over distances or between variables for which the ensemble evidence is insufficient. Inflation enlarges the ensemble spread to counteract variance lost through repeated updating, unresolved model error, and other approximations.
Ensemble square-root filters transform ensemble perturbations without adding independently perturbed observations. The ensemble transform Kalman filter performs this transformation within the low-dimensional space spanned by the ensemble. Local ensemble methods repeat the calculation for overlapping regions, allowing observational influence to vary spatially while reducing computational cost.
Hybrid systems combine ensemble-derived covariances with covariance models whose spatial structure is prescribed. This reduces sensitivity to finite-ensemble sampling while retaining information about the current flow. Ensemble-variational methods incorporate the resulting covariances into a variational cost function rather than applying a separate sequential filter update.
Observation processing and error representation
Observations used in assimilation include direct measurements and remotely sensed quantities. A radiosonde reports atmospheric conditions along a balloon trajectory, whereas a satellite radiance measures radiation integrated over a three-dimensional volume and a range of wavelengths. Assimilating the radiance directly requires an operator that represents atmospheric absorption, surface emission, instrument response, and cloud effects.
The observation-error covariance matrix includes instrument uncertainty and representativeness error. Representativeness error arises when an observation responds to scales or processes absent from the model state. An instrument may accurately measure a local feature that a coarse grid cannot express, so the disagreement does not necessarily indicate instrument failure.
Background errors are spatially and dynamically structured. Pressure errors extend over broad regions, while humidity errors often vary over shorter distances and respond strongly to clouds and convection. Cross-covariances permit an observation of one variable to modify another, provided that the assumed or ensemble-estimated error structure supports the connection.
Quality control identifies observations whose departures are inconsistent with the expected error distribution or with basic physical constraints. Bias correction addresses systematic differences between observations and model equivalents. In satellite assimilation, bias parameters may themselves be estimated within the assimilation system because calibration effects and model-dependent radiative biases can otherwise become difficult to separate.
Coupled and retrospective systems
Traditional geophysical assimilation systems treated the atmosphere, ocean, land surface, and sea ice as separate components. Coupled data assimilation instead allows observations in one component to affect another through cross-component covariances or through a coupled forecast model. The adjustment operates on time scales ranging from rapid air–sea exchange to slower changes in ocean heat content.
A reanalysis applies a substantially fixed assimilation and forecasting system to an extended historical observation record. This limits artificial discontinuities caused by repeated changes to operational models, although changes in the observing network remain influential. Reanalyses provide spatially complete reconstructions of atmospheric or oceanic conditions while retaining uncertainty inherited from both the observations and the assimilating system.
The analysis is not identical to the observed state. It is an estimate conditioned on a model, an observation network, and specified error statistics. When observations are sparse, the background has greater influence; when observations are informative relative to their errors, the analysis moves further toward them. Subsequent forecasts test the dynamical consequences of that balance but do not convert it into direct observation.