1. Introduction
Images constitute important data for studying geophysical flows since they can characterise a large range of spatial scales in comparison to sparse information provided by standard point measurement techniques.
Correlation-based techniques are common inverse approaches to extract sparse velocity fields from image sequences. These methods penalise deviations from constant flows over a spatial window and from the brightness preservation of the tracer along its trajectory (Leese et al., ). Such methods have demonstrated their robustness and accuracy for velocity measurement in experimental studies (Adrian, ). However, correlation is not a physical-based image observation model and is particularly limited in the context of geophysical flows (e.g., because of small-scale intermittency and 3-D effects). Moreover, the assumption of local constant motion gives rise to an intrinsic measurement scale associated to the correlation window size and an artificial regularity of the flow. In particular, for low-contrast observations such as satellite images, correlation-based methods are considerably limited and provide only very sparse motion fields.
In computer vision, estimating the projected apparent motion of a 3-D scene onto the image plane, referred in the literature as optical flow, has been an intensive subject of research since the 1980s and the work of Horn and Schunck (). The inverse motion modelling proposed in this seminal work is formulated in a differential framework enabling the easy introduction of a physical-based direct observation model (relying on brightness preservation, mass conservation, scalar transport, etc.) adapted to the particular image data at hand (Fitzpatrick, ; Heas et al., ; Heas and Memin, ; Liu and Shen, ). The model links the motion field to the image intensity function. However, these models are insufficient for fluid flow estimation since they are ill conditioned, as they are generally constituted at each point of a scalar equation for two unknowns. The introduction of regularization constraints is required to remove the motion ambiguities and achieve inversion. Common spatial regularizers may introduce measurement inaccuracy since they impose in a small spatial neighbourhood a prior smoothness, which may improperly describe the regularity followed by fluid flows. Moreover, they strongly depend on a tuning parameter weighting the amount of smoothing required to cure the ill-posed nature of the inverse motion problem. For large values of this parameter (which is almost unavoidable when facing noisy images or with low photometric contrast), the solution tends to be too smooth and presents sometimes lack of accuracy at large scales. Alternatively, dynamical constraints embedded in variational assimilation frameworks have recently been proposed for geophysical flow estimation (Papadakis and Memin, ; Corpetti et al., ). Unfortunately, these methods rely only on a large-scale dynamical model, which is not valid at the pixel level.
This work aims at showing the possibilities to alleviate these limitations in coupling two main ingredients: turbulence statistical modelling and Bayesian inference. In the one hand, important advances have been done in the statistical modelling of turbulence since the precursor work of Kolmogorov in 1941 (see, e.g., Kraichnan, ; Frisch, ; Kurien et al., ; Falkovich et al., ; Lindborg and Cho, ). In particular, it has been shown from the Navier–Stokes equations for different flows that turbulent motion regularity can be characterised using some (assumed universal) scaling properties of the probability distribution function (PDF) of motion increments. On the other hand, Bayesian modelling enables the reconstruction of complex physical processes in the presence of incomplete observations and prior geophysical scaling (Berliner et al., ; Wikle et al., ; Hoar et al., ). Moreover, this framework results in the design of non-parametrical methods and provides reliable model selection tools given some data (Gull, ; MacKay, ; Jaynes, ; Robert, ).
To be more precise, in this work we propose a method for motion estimation and turbulence characterisation, which exploits simultaneously two ideas: (1) the use of turbulence scaling laws for motion regularization in optic-flow inverse problems and (2) the selection by Bayesian evidence maximisation of the most appropriate power law model describing the image intensity function. The resulting frame-to-frame regularization is built from the physics of fluids and involves neither dynamical constraints nor an adjoint model within a variational assimilation framework. It is intrinsically multi-scale as it controls the solutions within a prescribed scale range. It is also non-parametric in the sense that it does not involve the tuning of any smoothing parameters. Finally, it allows as a by product the recovery of some important quantities for turbulence characterisation such as energy flux across scales.
After validating the method using synthetic images from numerical simulations of 2-D turbulence, we use the proposed regularization model to analyse data from meteorological image sequences. Recovered atmospheric quantities such as the energy and enstrophy flux are shown to be in good agreement with in situ measurements. Moreover, reconstructed second- and third-order structure functions and PDFs of velocity increments are consistent with previous analysis using aircraft data (Lindborg, ; Lindborg and Cho, ). We believe that those first results demonstrate the potential of image-based techniques in complementing other measurements for geophysical studies.
The study is organised as follows. In Section 2, we briefly overview optic-flow methods and review the motion scaling properties developed in theoretical works on turbulence. In Section 3, self-similar regularizers are briefly introduced. Finally, the synthetic and real-world experiments on meteorological image data are presented, followed by the conclusions. Details of the model, of the multi-scale constraint minimisation problem and of the use of Bayesian evidence for the optimal selection of the prior self-similar model are given in the Appendices.
2. Related work
2.1. Optic-flow methods
This section briefly describes some solutions for the computation of velocity fields from image sequences. Interested readers may refer to an extended review (Heitz et al., ) for a more complete account of different techniques proposed in the literature and the description of some new trends based on variational assimilation processes.
2.1.1. Image observation model.
Motion perceived through image intensity I(x, t) variations and the projection on the image plane of the velocity field v=(u, v) are identical when considering rigid motion and stable lighting conditions. In this situation, motion v respects the standard optical flow constraint (OFC) equation, which reads:
2.1.2. Standard regularization constraints.
To deal with this under-constrained estimation problem, an additional set of constraints have to be satisfied by the sought velocity field.
In the standard computer vision literature, dealing with frame-to-frame motion estimation in natural scenes, these constraints are usually chosen to enforce some spatial regularity of the solution. Global regularization schemes are convenient to enforce global coherence via local spatial dependencies. More precisely, estimation is performed through the minimisation of an energy function composed of two terms:
1Let us outline that in the case of large displacements, the brightness constancy model (1) is generally replaced by its time-integrated form I(x+vdt,t+dt)−I(x,t)=0 linearised around some initial solution x+vdt. For the sake of clarity we will not describe the time-integrated model further even if its application is in practice essential to recover long-range velocity fields (see Heitz et al. () and references therein).
In the context of geophysical flows, second-order spatial regularizers on motion vorticity and divergence have been proposed to overcome these limitations (Corpetti et al., ; Yuan et al., ; Heas et al., ). It is important to outline that these spatial regularization constraints do not rely on physical kinematics properties. At best they mimic qualitative regularity properties of the physical system of interest, i.e., coherent blobs of vorticity and divergence. Recently, Corpetti et al. () proposed to rely on large-scale dynamical constraints enforcing the temporal coherence of the solution, since in geophysics there exist some fair approximations of the physical equations governing the flows. The authors introduced a 4-D variational assimilation technique to solve the motion estimation problem on a whole image sequence subject to layered shallow-water dynamical constraints. Unfortunately, this choice of constraints does not entirely remove the under-determination of the estimation problem and still yields an infinite set of solutions. To close the problem, Corpetti et al. () needed to rely on an additional local spatial regularization on motion measurements (Lucas and Kanade, ) in consistency with a filtered version of shallow-water dynamics. By doing so, they restricted themselves to the determination of the solution at large scales (on a grid of a 100 km resolution).
In summary, in frame-to-frame estimation or variational image assimilation, there is actually a lack of physical constraints to solve at small scales the under-constrained fluid motion estimation problem. However, as we will show in the following, turbulence statistical models can provide the needed closure.
2.2. Turbulence statistical modelling
Since experimental results at the beginning of the 20th-century, and later Kolmogorov's theoretical works, turbulent motion increments have been known to be structured as scale invariant spatial processes. In many turbulence models, a quantity of interest is the longitudinal velocity increment, which, in the direction of a unitary vector nθ, is given by
represents a spatial increment. In turbulent studies, homogeneity and isotropy are often assumed; that is to say, it is considered that statistical properties of the velocity field are invariant under translation of spatial location x and rotation of direction nθ.2The hypothesis of isotropy can be relaxed; see, e.g., Taylor et al. (). The method to recover isotropic statistics described there motivates the use of different directions for the displacements in Section 3.1.
can be approached by spatial integration:
For 3-D isotropic turbulent flows, Kolmogorov () demonstrated from the Navier–Stokes equations that the third-order moment of the PDF 
, namely the third-order structure function, is linear w.r.t. scale and follows the well-known ‘4/5 law’: 
in the inertial range (see, e.g., Frisch () for details on this demonstration). The inertial range is defined as the range of scales [η, ℓ0], where η represents the largest molecular dissipative scale and where ℓ0 is smaller than the diameter L of the largest vortex. Within this range, an energy flux cascades from large to small scales. The kinetic energy dissipation rate 
corresponds to this energy flux passed across scales which is then evacuated at small scales by molecular viscosity.
Analogously, for pure 2-D turbulence with energy injection at scale ℓ0, Kraichnan () demonstrated that two different cascades are coexisting: a direct cascade of enstrophy (the L2 norm of the vorticity ω=∇×v), which satisfies 
within the inertial range [η, ℓ0], and an inverse cascade of energy where 
within the range [ℓ0,L]. An enstrophy flux ɛω passes in the direct cascade from large to small scales, whereas an energy flux ɛ passes in the inverse cascade from small to large scales.
In the case of atmospheric turbulence, the actual scaling laws constitute still open questions. Lindborg () and Lindborg and Cho () consider scaling laws inferred from aircraft measurements and propose an answer to the question: ‘can atmospheric flow statistics be explained by two-dimensional turbulence?’. The authors showed that, under some isotropy assumption, the self-similar processes observed in the aircraft data (Nastrom et al., ) at small and at large scales can be modelled by the superposition of a 3-D direct energy cascade and a 2-D direct enstrophy cascade so that
Going further, Kolmogorov assumed that the longitudinal velocity increments were strictly self-similar to explain the scaling laws observed in many experiments. If this is true, the normalised PDF of motion increments is self-similar through scales. This implies that, in a given cascade, the pth order structure function follows a power law:
, although power law behaviour in inertial range still holds. The special case of the inverse cascade of 2-D turbulence must be remarked, as growing evidence indicates that it is self-similar (Benzi and Scardovelli, ; Bernard et al., ).
Finally, another result that is used in the following sections is that any 2-D or 3-D turbulent flow is regular in the dissipative range; using Taylor expansion we then have 
within ℓ∈[0, η]. All these general power law dependencies motivate the use of self-similar priors for motion estimation as discussed in the next section.
3. Regularization of motion increments
Exhibiting self-similarity properties of fluid flows enables to define prior spatial regularization functionals. Besides providing a closure for motion estimation, power law priors have several advantages compared with traditional smoothing functions: they constitute physical sound regularizers for fluid motion and they provide intrinsic multi-scale prior models, which structures motion across scales. In this section, we discuss the use of the self-similar regularization of motion increments with a prior power law and postpone to Section 4 how Bayesian modelling is used to select the most appropriate power law from the images (instead of imposing one).
3.1. Self-similar constraints
Let us first formalise the use of self-similar constraints to specify regularization functionals. Although there is no exact prediction on its scaling law for non-strictly self-similar flows, we chose to use the second-order structure function S2(ℓ) because it constitutes a convenient quadratic constraint for our purpose. Also, this quantity is related with the energy spectrum and the two-point correlation function, quantities of interest in experiments and simulations. Since the exponent of the power law is unknown, we take into account in Section 4 deviations from the predicted laws by selecting the most likely power law defined by the two parameters γ2 and ζ2 given the image data (γ and ζ for simplicity in the following).
S2(ℓ) is an expectation which can be obtained by spatial integration over the image domain and over all directions as presented in eq. (7). A self-similar constraint is then defined at each scale ℓ as the difference between the second-order structure function and (for the moment) a given power law depending on some parameters γ and ζ:
3.2. Constrained motion estimation problem
Referring to Section 2.1, the minimisation of the ill-conditioned optic-flow estimation problem can be written as
and where 
represents the Lagrangian multipliers. Further details are given in Appendix A.
4. Selection of a prior power law
So far, the proposed model recovers motion from images conditioned by some prior power law model 
for the second-order structure function. This power law can be given by turbulence statistical models describing the self-similar structure of turbulence, which is assumed universal in the works of Kolmogorov (), Kraichnan () and Lindborg and Cho (). However, we now want to relax this universal assumption by allowing ζ and γ to change and select the most appropriate power law model for motion estimation given only the image data. With this prior model selection, we cope with uncertainties in turbulence theoretical predictions associated to, e.g., intermittency effects. Moreover, it will enable us to reveal directly from the images important physical quantities in turbulence such as power law exponents, flux across scales, or energy and enstrophy dissipation rates. To this end, we now reformulate the constrained motion estimation problem in a probabilistic framework, and we show how probability of the prior model 
can be evaluated given image data. Bayes’ rule provides a nice framework to evaluate this model probability.
4.1. Probabilistic model
A probabilistic reformulation of the global motion estimation problem yields a model relating the image intensity function variable I, the motion field variable v, an observation model variance 
and, finally, a prior power law model 
. Noticing conditional independences, the joint probability can be factorised in three primary statistical models yielding the Bayesian hierarchical model:
represents the image observation PDF or likelihood, 
denotes the motion prior PDF and 
and 
denote the prior PDF, respectively, on the likelihood hyper-parameter and the power law model. Let us note that Bayesian hierarchical model can also be used to decompose multi-dimensional probability distributions of complex spatio-temporal processes in geophysics (Berliner et al., ; Wikle et al., ; Hoar et al., ). Using hierarchical models, it is then possible to infer the joint maximum a posteriori (MAP) estimate by maximisation of the joint probability 
w.r.t. v, β and 
. However, it is well-known in Bayesian analysis that selection with a joint MAP estimator tends to over-fit the data, i.e., to favour models with large degrees of freedom (Robert, ). Bayesian analysis suggests to rely instead on marginalised posterior distributions for model and hyper-parameter selection. Therefore, by successive marginalisation of the joint posterior distribution 
, we perform three levels of inference yielding estimates: (1) the velocity field; (2) the variance of the observation model error; and (3) the power law model. Although, variable β does not affect the first level of inference, we will see that it is necessary to consider the variance of the observation model to perform model selection. In the following for notational convenience, the dependence of the Lagrangian multipliers to model 
will be omitted. The three levels are detailed in the next sections.
4.2. First inference level: motion field estimation
In this first level of inference, we consider the maximisation of the joint probability w.r.t. v given some data realisation I0, the hyper-parameter estimate 
and the selected scaling model 
:
. Let us show that solving the problem 
introduced in the previous section is equivalent to inferring a velocity field 
according to an MAP criterion. This can be shown using Bayes’ relation to define the motion field posterior PDF:
denotes a normalisation constant also called partition function. The constraints 
weighted by inverse variances 
form the energy of a multi-scale prior PDF:
denotes the partition function. At the point λ*, the Lagrangian function 
thus defines, up to a multiplicative factor, the energy of the posterior PDF:
) denotes the associated partition function. It is now obvious that the maximum v* of the posterior is reached at the minimum of the Lagrangian:
4.3. Second inference level: estimation of observation model variance
A second level of inference considers the maximisation w.r.t. β of the joint posterior PDF, marginalised over v, given the data realisation I0 and the selected scaling model 
:
. Using again Bayes’ relation one can write:
of (21) is simply the maximum likelihood (ML) estimate, or in other words the maximum of the evidence PDF 
. This probability is defined by a marginalisation w.r.t. the velocity field:
maximising the evidence is obtained by cancelling out its derivative.
4.4. Third inference level: selection of power law
A third level of inference considers the maximisation w.r.t. M of the joint posterior PDF, successively marginalised over v and β, given some data realisation I0:
for 
(also called evidence for 
) appeared earlier as the normalisation constant of the second level of inference (eq. 21). Its evaluation implies marginalising out variable β:
as detailed in Appendix C. The model maximising the posterior probability (26) of the power law model:
of the power law parameters 
.
As a result, it is important to outline that, a priori, no particular power law is imposed on the data, but rather the second-order structure function is constrained to follow some power law in some range of scales (which can be selected). If the smallest resolved scales in the images correspond to the dissipative scales of the flow, the constraint reduces to a multi-scale first-order smoothing, as in that range 
. This kind of smoothing function constitute a multi-scale generalisation of regularizer (5) computed with finite differences. On the other hand, if the smallest resolved scales in images are in the inertial range of the turbulent flow, the constraint gives a better representation of the physics of the problem than ad hoc regularizers. Finally, note that since laminar flows possess vanishing increments, their structure functions can be modelled with power laws with prefactors equal zero (
). Appendix D shows with a toy example how the proposed methodology succeeds to recover this particular case. Therefore, the proposed methodology can be expected to improve the physical representation of the flow in turbulent cases and enforces smooth solutions if the hidden flow is smooth or laminar.
5. Numerical and experimental validation
5.1. Synthetic turbulence of horizontal velocity fields
To evaluate the performance of the reconstruction of the horizontal velocity field from intensity images using the proposed self-similar regularization, synthetic image sequences were generated based on forced 2-D turbulence obtained from direct numerical simulations (DNS) of the 2-D and incompressible Navier–Stokes equations with a Reynolds number of 3000. The pseudo-spectral code used for this study solved the vorticity conservation equation and the Lagrangian equation for particle tracers transported by the flow or, in the case of scalar images, the advection–diffusion equation, i.e., model (2) in the case of divergence-free velocities, with a Pecklet number of 2100. The flow was stirred by an external force acting between wavenumbers k = 5 and 10. Figure 1 shows the isotropic (angle averaged) energy spectrum and the energy flux for one snapshot of the velocity field. A flat spectrum is observed at scales larger than the forcing scale, resulting from an incipient inverse cascade of energy. The direct cascade range shows a spectrum steeper than 
expected from Kraichnan phenomenology, which is not unusual in numerical simulations of 2-D turbulence and which, in this particular case, can also be associated to the moderate spatial resolution and Reynolds number considered (the slope is in fact close to –5). The simulation shows positive enstrophy flux toward small scales for wavenumbers larger than the forcing wavenumbers and a few wavenumbers with negative energy flux associated to the beginning of the inverse cascade. Also, the energy flux in the direct cascade range is negligible when compared with the enstrophy flux (see Boffetta, , for examples of energy and enstrophy fluxes in high-resolution simulations of 2-D turbulence). A precise description of the simulation can be found in Carlier and Wieneke (). This database has proven its relevance in previous studies in the field of fluid flow estimation (Yuan et al., ; Heas et al., ; Papadakis and Memin, ; Heas et al., ).
Isotropic energy spectrum and flux in the numerical simulation. Left: energy spectrum, as a function of the wavenumber in units of the inverse of the length of the box, and in units of 1/pixel (between parenthesis). Right: enstrophy flux (solid) and energy flux (dotted), as a function of the wavenumber using the same notation.
Figure 2 presents a sample of the synthetic particle and scalar images of 
pixels generated using the numerical simulation. The corresponding DNS velocity field is also presented in Fig. 2. The motion estimates were obtained by minimising discrepancies to the brightness consistency (1) (resp. to the scalar diffusion equation, i.e., model (2) in the case of divergence-free velocities) for particle images (resp. for scalar images) under power law constraints. The self-similar regularizer was applied only in the range of 
pixels (corresponding either to dissipation or to the end of the inertial range). A multi-resolution approach was used to cope with large displacements (Bergen et al., ). Motion estimates can be compared with the DNS motion fields and with the operational correlation-based software from LaVision company in Fig. 2. The true and reconstructed velocity fields look similar. It can also be noticed that although correlation-based methods are accurate measurement tools for particle imagery, they usually perform badly for scalar imagery. Quantitative details of the estimation are given next.
Particle and scalar images, DNS and motion estimates. From left to right: Particle (above) and scalar (below) image at time 
; Turbulent motion field obtained by DNS in vectorial (above) and colour (below) representation; Motion field inferred by the proposed method and correlation-based velocity fields provided by LaVision in the context of the Fluid FET Open European project (Heitz et al., ) for particles (above) and scalar (below). For a better visual analysis, the velocity fields are displayed in a colour representation where colour and intensity code respectively vector orientations and magnitudes (Baker and coauthors, 2007). The corresponding colour map circle is displayed on the left.
5.1.1. Evaluation of power law selection
A flat prior was chosen on the power law parameters in (27) and, therefore, inference was performed by maximising the evidence probability (26) w.r.t. γ and ζ. This probability was evaluated by sampling the exponent κ near the theoretical value of 2 (the value expected for a dissipative range or for a direct enstrophy cascade from the theoretical predictions), and by sampling the prefactor γ around a first guest given for instance by a least square (LS) fitting of the solution of a rough Horn and Schunck () estimation.
Figure 3 shows the shape of the evidence according to the power law model parameters. The model probability global maximum is located at the coordinates ζ=0.0030 and γ=1.975 (resp. б=0.00275, γ=1.850) for the particle (resp. scalar) images. Three observations assess experimentally the Bayesian criterion used for model inference. First, Fig. 3 shows that, specially for particle images, the parameters yielding high model probabilities correspond to low root mean square (RMS) motion reconstruction errors. Second, this correspondence is not so remarkable in the case of scalar images, because minimising an RMS error criterion does not always lead to the underlying power law. Indeed, as shown in the plots of Fig. 4, the inferred power law parameters are for both experiments very close to the values б=0.003, γ=1.948 obtained by an LS fit on the ground truth motion field, whereas on the contrary in the case of scalar images, choosing the power law minimising the RMS error induces a slight bias. Remarkably, the reconstructed structure functions S2(ℓ) in x- and y-directions (horizontal and vertical in the images) as well as in the diagonal directions adjust well the structure functions computed directly from the simulation (indicated by the straight lines) in a wide range of scales, and well beyond the range used for the regularizer. Third, between 1 and 8 pixels, the power law of all structure functions is close to ∼ℓ2, which is the expected value from the theory in the inertial or dissipative range. It is also observed that the simulation approaches isotropic statistics at small scales since structure functions calculated in horizontal–vertical and in diagonal directions are nearly identical.
Energy of probability 
and RMS error behaviour w.r.t. power law parameters. Plot of minus of the log of the power law model evidence probability (left) and the motion RMS error (right) w.r.t. power law exponent ζ (ordinate) and prefactor γ (absciss) for the particle (above) and scalar (below) image sequences. The coordinates of the minima are plotted with yellow squares. These coordinates have to be compared with the true power law parameters 
, given by an LS fit of the DNS data.
Second-order structure function reconstruction. Left plots: Inferred power law model represented by a blue dashed line, and estimated (resp. true) second-order structure function in horizontal-vertical directions plotted with stars (resp. continuous line) and in diagonal directions with cross (resp. dashed line). Right plots: identical legend for the motion estimate obtained with the prior power law model minimising the RMS error.
Therefore, maximising the model evidence constitutes a theoretically sound and reliable criterion for selecting the prior power law.
5.1.2. Evaluation of motion estimation.
A comparison with state-of-the-art methods based on the RMS reconstruction error
3 To compute RMS error on the pixel grid, true velocity fields have been interpolated since they are given on a shifted (by half a pixel) grid by DNS.
Motion estimation accuracy. Evolution of RMS error w.r.t. time index for an operational correlation-based method, a robust first order regularizer (Horn and Schunck, ), a second-order regularizer (Yuan et al., ), and the inferred self-similar constraints for the particle (left) and scalar (right) image sequences. The results obtained with correlation approach were provided by LaVision (www.lavision.de) company from their operational PIV software (Davis) in the context of the Fluid FET Open European project.
To assess quantitatively the accuracy of the reconstruction of the velocity increment PDFs, we plotted in Fig. 6 the exponents and the prefactors of structure functions w.r.t. their order. The exponents correspond to the dissipative range as they were computed by LS fitting of the reconstructed motion fields for increments of the structure functions between 1 and 8 pixels. Analysing the DNS motion field, it can be noticed that, in this range, the field is smooth and the exponents are on a straight line, as the structure function of order p, Sp(ℓ), goes as ∼ℓp. The other estimates introduce spurious deviations from the straight line that may be incorrectly interpreted as intermittency. It is striking that the proposed method provides a very good estimation of the exponents and prefactors while all other methods fail to characterise exponents above the fifth order. This is a testimony of the power of the proposed motion estimation methods in recovering accurately the multi-scale structure of turbulence.
Longitudinal structure function exponents and prefactor w.r.t. their order for the scalar image sequence. The proposed self-similar regularization (crosses +) provides exponents (left plots) and prefactors (right plots) (in an LS sense) very close to the ground truth (solid line) in comparison to estimates of (Horn and Schunck, ) (stars *) or (Yuan et al., ) (× symbols). Correlation-based measurements are not presented here since they do not provide motion increments in the bottom of the scale range of [1, 8] pixels.
5.2. Atmospheric turbulence
We now assess the inverse motion modelling approach using power laws on real data. The objective of this section is to estimate atmospheric parameters from satellite images, even when limitations associated to images and to required approximations may only give orders of magnitude of quantities of interest. We also compare results with previous observations using in situ techniques. The analysis is done using METEOSAT second-generation meteorological image sequences acquired above the North Atlantic Ocean at a rate of an image every 15 min during part of one day (5 June 2004), from 1200 to 1430 hour UTC (universal time, similar to GMT, the Greenwich Mean Time). Image spatial resolution is 3 × 3 km2 at the centre of the whole Earth image disk. This benchmark data, which has been provided by the EUMETSAT consortium, is composed of images of top of cloud pressure and cloud-classification images. Cloud classifications were used to segment images into two broad layers, at low and intermediate altitude.
4Stratus and stratocumulus (resp. altostratus, atltocumulus and nimbostratus) are classified in the lower layer (resp. the intermediate layer).
5Let us remark that one can not guarantee that the data model which has been proposed in Heas et al. () is valid for regions where the model assumptions are violated (e.g., for non-layered structures) or for too noisy input observations (e.g., due to cloud classification errors or pressure retrieval errors). However, the local weakness of the observation model is balanced in our approach by the introduction of the relevant prior information on turbulent motion regularity.
Depression over the north-Atlantic Ocean. Sequence of images depicting sparse pressure difference maps of layers at intermediate (above) and low (below) altitude. The images characterise the layers evolution in a time interval of 15 min and with an average spatial resolution of 3 km. Black regions correspond to missing observations and white lines represent meridians (20o, 30o and 40o), parallels (50o and 60o) and the coastal contours of south of Greenland (upper left corner).
In the following, based on the work of Lindborg and Cho (), we first introduce prior knowledge on the sought power law using the Bayesian relation (27). The power law is chosen to model the longitudinal structure function of indifferently the zonal and meridional horizontal winds. Power law selection reduces then in this case to estimating the energy flux of the direct energy cascade toward small scales. We then relax this assumption and perform model selection considering a flat prior distribution on both unknowns: the power law prefactor and the power law exponent. Finally, we provide some statistical analysis and compare wind fields obtained with the two methods.
5.2.1. Power law selection with a Lindborg's prior.
In this section, we assume that the scaling б=2/3 predicted by Lindborg in the direct energy cascade holds in the range I=[1,4] pixels, equivalent to I=[3,12] km [the direct energy cascade is visible in the second-order structure function for separations of about 10 km in the data of Lindborg ()]. This hypothesis can be formulated in the Bayesian relation of (27) by choosing a prior such as 
, where δ and 
denote, respectively, the Dirac and the flat distributions. We thus only need to infer the parameter γ. Note that no power law is imposed for separations larger than 12 km.
5.2.1.1. Direct energy flux estimation
The model proposed in Lindborg () provides power law models for the second-order structure function as:
where b and c denote parameters and 
is a Kolmogorov constant. The energy flux ε (which is equal to the energy dissipation rate) can be related in the scale range I to the power law prefactor by 
. Therefore, maximising the power law posterior probability also reveals the most likely energy flux. Figure 8 shows that this probability is maximised for power laws 
with energy flux:
Selection of most likely energy flux for Lindborg's model. Minus log of the power law posterior probability 
vs. energy flux ε in m2 s−3 for horizontal winds at low (solid line) and at intermediate (dashed line) altitude.
which correspond to estimates at low and intermediate altitudes. These estimates have the same order of magnitude as previously reported by results based on aircraft data analysis.
6A collection of other aircraft measurements (Nastrom et al., ) shows a typical value in the troposphere close to ∼3×10−6 m2s−3 (Lindborg, ). Nevertheless, note that analogously to stratospheric observations (Dewan, ; Nastrom and Eaton, ; Cho and Lindborg, ), one should consider a significant variability in measurements.
can be identified in the scale range [10, 40] km, and the agreement of the structure function for larger scales with the theoretical prediction seems reasonable.
Second- (above) and third-(below) order structure functions at low (left column) and intermediate (right column) altitudes. Second-order structure functions (+ symbols) are plotted with their associate models (dashed line). Third-order structure functions (plotted with × symbols for positive values and with + symbols for negative values) can be compared to their associate models (fine dashed line for positive values and coarse dashed line for negative values).
5.2.1.2. Direct enstrophy flux estimation
An LS estimate of the enstrophy flux 
can be obtained by fitting w.r.t. 
the third-order structure function to its model given by eq. (8) in the converged scale range (ℓ>50 km). In particular, at the upper range of the cascade, positive cubic power laws can be observed for both layers. The LS average enstrophy flux estimates are
7These values agree well with an early estimate of ∼10−15 s−3 obtained in Charney (). Authors in Tung () then re-obtained this same result from simulations. Note, however, that enstrophy flux is slightly lower than the estimated∼10−13 s−3 of Lindborg () using Nastrom et al. () aircraft data.
5.2.1.3. Moments convergence
To verify that we obtained converged statistics for the evaluation of the structure functions, we follow Gotoh et al. () and examine the convergence toward a flat curve of the accumulated moments
is the PDF of the normalised increments 
with denoting the standard deviation. To check convergence of the second- and third-order structure functions, we examine the accumulated moments and see if there is convergence toward flat curves for increments associated to the tails of the PDF. Curves 
and 
show that the second- and third-order moments are reasonably converged. However, we observe that accumulated moments show a faster convergence for the lower layer motion. This obviously results from sparsity of image data, which increases with altitude. Furthermore, curves show that the third-order moments are better converged for large-scale increments (ℓ≳0 km) than for small-scale increments (ℓ≲0 km). This is in agreement with the fact that the normalised PDFs of velocity increments show stronger tails as the scale studied is decreased. These tails, which depart from Gaussian behaviour, are the result of intermittency: turbulence comes in gusts, and scarce regions with strong gradients develop in the flow as the result of non-linearities. The development of these strong events also makes the convergence of high-order moments slower, as more data are needed to have enough statistics to resolve the impulsive (but scarce) events. An extension of the present study to consider more images may enhance convergence.
5.2.2. Power law selection with a flat prior.
In this section, we now relax the previous scaling assumption and consider instead a uniform prior 
defined on the bi-dimensional parameter space (γ, б)∈ℝ2 of the power law. Thus, we now also infer the most likely exponent 
of the power law together with the prefactor 
. Figure 10 shows that the selected prefactor (i.e., the energy flux) has the same order of magnitude as in the previous case. However, the selected power law exponent does not correspond to the model in (Lindborg and Cho, ). Indeed, the second-order structure function maximising the evidence probability does not scale as 
but rather as ℓ2 (at least, in the range of scales in which the constraint is applied; as will be shown next the structure functions scale in a similar way at scales larger independently of whether a scaling assumption or a uniform prior are used). Moreover, note that the model in (Lindborg and Cho, ) still corresponds to a region of high probability, i.e., a valley of the probability energy (marked by an iso-contour in Fig. 10). In the absence of ground truth, it is not obvious to decide whether one power law model is better than the other. One may conclude that using the images alone favours (in the particular case of these types of images) the choice of a low-complexity model that does not over-fit the data. In other words, the scaling of a regular motion is preferred, although more complex scaling is not drawn aside. When prior knowledge is available, the Bayesian modelling framework is able to enforce relevant regions of lower probability and recover accurately complex turbulence scaling laws.
Power law evidence probability. Minus log of model probability 
w.r.t. parameters: exponent ζ (ordinate) and energy flux ɛ in m2 s−3 (absciss) (i.e., prefactor 
), for horizontal winds at low (left) and at intermediate (right) altitude. Iso-contours of decreasing values around the minima are plotted in dark blue, yellow and turquoise.
5.2.3. Comparison of wind field estimates.
Deviation from strict self-similarity of the two motion estimates can be evaluated based on the behaviour w.r.t. their order of the structure function exponents (obtained by LS fitting of the reconstructed motion fields in the scale range of 1 to 6 pixels, corresponding to 3 to 18 km). Analysing the results in Fig. 11, it can be seen that we obtain a stronger deviation from linearity with the model of Lindborg and Cho (). However, in both cases the motion field is multi-fractal and not strictly self-similar.
Deviation from strict self-similarity at small scales (3–18 km) and at larger scales (30-80 km). Longitudinal structure functions exponents w.r.t. their order at low (solid curve) and intermediate (dashed curve) altitude using the model in (Lindborg and Cho, ) (on the left) or a flat prior (on the right) for power law models. The dashed straight lines represent strict self-similar behaviour, i.e., a linear relation of exponents w.r.t. order for the two models.
Exponents in Fig. 11 were computed for a similar range of scales as the one used for power law selection (3 to 12 km). It is thus not surprising that the second-order exponent changes substantially (as well as the other exponents), as in that range the method with the flat prior selected scaling ∼ℓ2. This change in the exponents should be understood in the same light as the change in the reconstructed power law of the second-order structure function: images are favouring a model of lower complexity with regular motion at small scales.
However, structure functions show similar slopes and features for both reconstructions outside the range used for power law selection (e.g., for scales larger than 40 to 80 km depending on the altitude). It is not the motivation of our work to do a detailed analysis of scaling exponents at intermediate scales for atmospheric data, but just as an example, in Figs. 11 and 12 we show the third-order structure functions for low- and mid-altitude using both methods, and the scaling exponents at larger scales (10 to 30 pixels, or 30 to 90 km) for the model of Lindborg and Cho () and for a flat prior. Two features are worth mentioning. In Fig. 12, note that not only amplitude of structure functions is similar in both cases (as already mentioned, as the amplitude of these functions is associated to the energy flux) but also that structure functions change sign at similar scales. In Fig. 11, exponents are similar at intermediate scales independently of the method used to reconstruct the field.
Third-order structure functions’ absolute value at low (left) and intermediate (right) altitudes for the two methods. Structure functions obtained with a (Lindborg and Cho, ) (resp. a flat) prior are plotted with×symbols (resp. □symbols) for positive values and with + symbols (resp. * symbols) for negative values. Estimate obtained with the (Lindborg and Cho, ) prior can be compared to their associate models (fine dashed curve for positive values and coarse dashed curve for negative values). At large scales, estimate obtained with both priors scale as ∼ℓ3 (straight dashed line).
Figure 13 compares motion fields obtained with our method (prior power law in ℓ2) with a vector field obtained by a correlation-based method. It is conspicuous that, for these very sparse and noisy images, correlation yields spurious and incoherent motion estimates while our approach yields a coherent and physically plausible reconstruction. Scaling in ℓ2 tends to produce vortex and divergent structures of larger magnitude than for scaling in ℓ2/3. Motion fields may present what could be interpreted as local abrupt discontinuities when visualised at larger scales. Indeed, the proposed optic-flow method provides a velocity vector per pixel. Therefore, small vortex and divergent structures of only a few pixels size may be confused with noise. In order to remove the confusion, we plot results at the pixel level. As shown in Figs. 14 and 15, velocity fields at this resolution appear to be very smooth and structured. In particular, vortex, sinks and sources are coherent structures of comparable order as claimed in Lindborg, .
Estimated horizontal wind fields compared to correlation results. Dense wind fields at low (left) and intermediate (right) altitude where obtained using the flat prior, i.e., a scaling in ℓ2. The correlation results were obtained using the operational PIV software (Davis) from LaVision (http:\\www.lavision.de) company.
Detail of horizontal winds at low altitude. From top to bottom: input image (zoom between the 20o and 30o meridians and the 50o and 60o parallels), solenoidal (2 following lines) and divergent part (2 last lines) of motion estimated with a scaling of ℓ2/3 (second and fourth line) or ℓ2 (third and fifth line).
Detail of horizontal winds at intermediate altitude. From top to bottom: input image (zoom between the 20o and 30o meridians and the 50o and 60o parallels), solenoidal (2 following lines) and divergent part (2 last lines) of motion estimated with a scaling of ℓ2/3 (second and fourth line) or ℓ2 (third and fifth line).
6. Conclusions and perspectives
In this study, we presented a physically based, multi-scale method that does not involve any tuning of regularization parameters for fluid motion modelling from images. It relies on a Bayesian hierarchical model, which simultaneously provides optimal solutions for two problems: motion estimation and regularization model selection. The regularization models arise from Kolmogorov's work on turbulent flow self-similarity and recent results in the study of turbulent flows. Structure functions are used to exhibit some regularity in velocity field reconstruction, without a priori imposed power law. Besides achieving motion inversion, the method recovers several quantities of physical interest in turbulence. Moreover, for images of turbulent flows (or for laminar flows) where dissipative range is resolved, small scales are smooth and the proposed method reduces in fact to a smoothing with, e.g., null second-order structure function, or with the second-order structure function proportional to the square of the spatial increment. Experiments on a synthetic sequence show that the method is more accurate than the best fluid-dedicated motion estimators. As a result, we believe the method constitutes a valuable tool for physical characterisation of turbulence from images. In particular, we were able to recover consistent flux across scales in atmospheric turbulence at different altitudes from a meteorological image sequence. Results are also consistent with theoretical predictions and with previous studies using in situ measurements.
We believe that this work opens interesting perspectives for observational characterisation of turbulence. Characterisation of turbulence cascades could also be performed by analysing other meteorological (or laboratory) images, depicting for instance evolution of passive scalars, water vapour, or temperature fields. Of course, the method requires a proper direct image observation model for each problem of interest, and as a result may have to be complemented with other measurement techniques for validation. However, we believe that the results presented here are promising and give a useful tool to complement other measurements in the difficult problem of turbulence characterisation. Inferring directly the scale range on which the prior power law should be exhibited according to the observed images represent an interesting perspective which is in principle straightforward considering a Bayesian framework. It would present the advantage to extend the set of possible power law multi-scale models considered for selection to traditional single scale regularizers, which represent power law particular cases. Another interesting perspective could concern the addition of a prior dynamical model to ensure dynamical consistency along time. Several works relying on variational image assimilation techniques have recently been proposed in this direction (Papadakis and Memin, ; Corpetti et al., ).
related to any quadratic data term w.r.t motion v (e.g., the model of (4)) can be expressed in the matricial form A0v−b0, when discretized on an image grid S of m points. The two discretized components of v∈ℝn now represent a field of n=2m variables supported by the grid S, A0 is n×n symmetric positive definite, and b0 ∈ℝn represents a vector of size n. The discrete data term can then be rewritten as:
and 
and 4 diagonal directions 
of the bi-dimensional plane. The continuous motion field discretisation yields to rewrite the integral over the spatial domain Ω with a sum over the pixel grid S. On this regular grid, the longitudinal velocity increment function is available at scale ℓ either for 
on horizontal nh and vertical nv directions, or for 
on diagonal directions 
. To avoid using boundary conditions, we exclude of the sum the grid points 
which belongs to image borders of width ℓ. A node subset 
is thus defined depending on scale. Therefore, at scale ℓ, the discrete second-order structure function reads 
:
. The quadratic constraint derivatives can be expressed in the vectorial form 
, where 
are symmetric positive semi-definite matrices and 
are vectors of size n. Indeed, manipulating the derivative of the expectation of eq. (A2) w.r.t. the motion field v, one obtains for grid points in the subset Sℓ a new expression for the sth component of the constraint derivatives 
represents a discrete Laplacian operator with a centred second-order finite difference scheme defined on a grid rotated 
of with mesh point spacing equal ℓ to and where 
, 
and 
are similar to the discretisation of second-order spatial derivatives in a centred second-order finite difference scheme on a mesh with spacing of ℓ. The constraints can finally be written in their discrete form as,
.
is introduced as
denotes the dual function. As the functions fd and 
are convex and as the constrained group is not empty, for positive Lagrangian multipliers 
, L is convex and the minimisation problem (P) has a unique saddle point, i.e., an optimal solution v* which is unique. More details of convex functionals and the convergence of the related minimisation problems can be found in Weickert and Schnörr (
, which represent the regularization parameters at scales 
, are then optimally given by the coordinates of the saddle point. Note that for negative Lagrangian multipliers, the convexity of the functional is no longer insured, and there is no guarantee of the solution unicity. Nevertheless, there still exist local optimal solutions.
of a locally convex Lagrangian function at point 
can be obtained by cancelling the gradient
are constituted by the superposition of a collection of discrete operators obtained in a centred second-order finite difference scheme on a grid of mesh ℓ, which corresponds to second-order derivatives at different scales. Since we have no guarantee that the matrix 
is positive-definite depending on the sign of Lagrangian multipliers 
, the solution of the large system given by eq. (A9) is efficiently achieved using a conjugate gradient squared (CGS) method with an incomplete LU pre-conditioner. The dual function is then given by
. We employ a classical gradient method to find 
which maximises the dual function and thus obtain the solution 
. Finally, the constrained motion estimation method results in an Uzawa algorithm, which is used to converge toward the saddle point 
, i.e., the optimal motion estimate under self-similar constraints. An important remark is that once the regularization coefficient vector 
has been estimated for two consecutive images of the sequence, assuming motion stationarity, only one step of the Uzawa algorithm is needed to process the following image pairs. Therefore, the complexity of the algorithm reduces to the resolution of the linear system by CGS, that is, 
, where is the conditioning number of 
. The Uzawa algorithm is briefly presented hereafter.
solution of
eq. (A9).
by a gradient ascent step:
At convergence: 
represents a vector whose components are 
and 
and 
denote the step size and a positive-definite matrix defined at iteration k. The latter parameter and matrix can be fixed at each iteration to formalise a quasi-Newton algorithm satisfying the strong Wolf conditions (Nocedal and Wright, 
of the observation model error maximising the evidence, we first detail analytically the three different partition function composing the evidence, then give an analytical expression of its derivate w.r.t. β and, finally, provide an expression for the ML estimate 
cancelling out the evidence derivative.
energy related to a quadratic observation model 
is composed of the energy of a multi-dimensional Gaussian with m uncorrelated components of variance 
, where m is the number of point composing the image grid. Its partition function reads:
of the posterior PDF of (19) is an integral, which can be calculated using Laplace's approximation (see, e.g., MacKay, 
is the associated set of Lagrangian multipliers. Noting that the term 
vanishes since constraints are fulfilled, this yields:
of the self-similar prior of (18) does not exist since the PDF is degenerated and has an infinity set of maxima, corresponding to the infinite set of admissible velocity field solutions respecting the self-similar constraint. To make this prior well-defined, we use Dirichlet boundary conditions for evaluating the evidence.
(resp. 
) does not (resp. does) depend on β. Using (B9), the derivative of log evidence (B7) w.r.t. β can be written as:
random samples of an n-dimensional normalised and centred Gaussian distribution 
are used to approximate the trace operator:
is the solution provided by the CGS algorithm of the inverse problem 
, where 
is the unknown.
. Using Laplace's Gaussian approximation, the prior model evidence reads:
is defined as a second-order derivative w.r.t. β of the evidence energy:
at 
:
:
, the parameter vector 
is sampled uniformly in ℝ2. The ML estimate 
is the minimiser of (C4) and represent the selected model.
constitute proper models for their structure functions. In the following we use a toy example in order to demonstrate the capacity of the proposed algorithm to recover accurately this particular case. Wu used a synthetic image (of 64×64 pixels) seeded with particles, which is evolved by a laminar flow (a translation by 1 pixel) to form the second image of the pair.