1. Introduction
The Taylor (1938) frozen-turbulence hypothesis is commonly used in physical science and engineering studies to infer the spatial distribution of flow variables from time series of those variables at fixed locations, study the statistical properties of turbulence, and mitigate analysis errors arising from the nonsimultaneity of data collections. According to this hypothesis, small-scale features in the velocity field can be idealized as patterns of unchanging form that are transported by a larger-scale “wind stream.” The validity of this hypothesis is a topic of continued interest (Uddin et al. 1997; Burghelea et al. 2005; Dennis and Nickels 2008; Li et al. 2009; Castro et al. 2011; Higgins et al. 2012; Wilczek et al. 2014; Creutin et al. 2015). Not surprisingly, the scales of motion bear on the validity of the hypothesis, and the validity degrades for longer time intervals.
The pattern-translation components U and V used in a variety of radar-, hydro-, and mesoscale-meteorological applications have been estimated visually (Heymsfield 1978; Carbone 1982; Hildebrand and Mueller 1985; Austin 1987; Dowell and Bluestein 1997; Wurman et al. 2007), obtained as wind profiles from a numerical weather prediction model (Mittermaier et al. 2004), calculated by cross-correlation analysis (Zawadzki 1973; Bellon and Zawadzki 1994; Fabry et al. 1994; Liu and Krajewski 1996; Anagnostou and Krajewski 1999; Gerstner and Heinemann 2008; Harrison et al. 2009; Nielsen et al. 2014; Thorndahl et al. 2014; Seo and Krajewski 2015), and derived from the minimization of a cost function in which (1.1), (1.2), or (1.3) is imposed as a weak constraint in a variational framework 1 (Gal-Chen 1982; Liu et al. 2004; Shapiro et al. 2010a, hereafter SWP10a). SWP10a repurposed a single-Doppler velocity retrieval (SDVR) algorithm similar to the frozen-turbulence-based SDVR algorithms of Tuttle and Foote (1990), Qiu and Xu (1992), Shapiro et al. (1995), Zhang and Gal-Chen (1996), Liou (1999, 2007), Lazarus et al. (2001), Liou and Luo (2001), and especially the variational echo tracking procedure of Laroche and Zawadzki (1995). In repurposing SDVR as advection correction, SDVR “winds” are interpreted as pattern-translation components rather than air velocity components. Moreover, while the tracer may be of limited interest in SDVR (beyond its use in obtaining the “winds”), it is of primary interest in advection correction—it is the variable being analyzed. SWP10a speculated that, in many cases, a frozen-turbulence-based SDVR procedure may work better as an advection-correction procedure than as a wind retrieval.
A vortex of unchanging form translating past a radar provides a simple example of a radial velocity field that does not satisfy the frozen-turbulence hypothesis (Dυ r /Dt ≠ 0) even though the velocity field does satisfy that hypothesis (D u/Dt = 0). The same velocity vector that yields an inbound radial component (green, υ r < 0) when the vortex is west of the radar yields an outbound radial component (red, υ r > 0) when the vortex is east of the radar.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
Unlike the expanded form of DR/Dt [which is (1.2)], the expanded form of (1.4) [see (2.2)] includes second derivatives with respect to space and time.
Attempts to mitigate analysis errors stemming from the nonsimultaneity of radar data collections often use advection-correction procedures in which U and V are treated as constant throughout the analysis domain or on each subdomain of a partitioned analysis domain. A notable exception is the SWP10a procedure, in which (1.3) is applied as a weak constraint along with smoothness constraints to obtain spatially varying U(x, y) and V(x, y) fields and an analyzed reflectivity field R(x, y, t) consistent with those fields and two time levels of input data. In a companion paper (Shapiro et al. 2010b, hereafter SWP10b), reflectivity data from a supercell storm were advection corrected to a time between the two data input times, and compared to data at that time. The spatially variable procedure produced analyses with smaller root-mean-square errors and larger correlation coefficients than analyses obtained with any pair of constant U and V parameters. Additional tests in SWP10b focused on advection correcting υ r using U and V obtained from (i) the spatially variable reflectivity-based procedure, or (ii) a range of imposed constant U and V values. Although the reflectivity-based procedure yielded generally good results, the errors in υ r were slightly larger than the errors obtained using the best of the constant U and V analyses. SWP10b suggested that better results might be obtained using a υ r -based procedure, that is, a procedure based on (1.4) instead of (1.3).
In this study we develop a frozen-turbulence-based advection-correction procedure for the radial velocity field in which the retrieved pattern translation components are spatially variable. An important application is to the analysis of nonsimultaneously collected radial velocity data from two or more radars to common analysis times for use in multiple-Doppler wind syntheses of squall lines, supercells, tornadoes, and other convective phenomena (Clark et al. 1980; Gal-Chen 1982; Chong et al. 1983; Hildebrand and Mueller 1985; Smull and Houze 1987; Bousquet et al. 2007; Wurman et al. 2010; Wienhoff et al. 2018; Oue et al. 2019; and many other studies).
The paper is arranged as follows. In section 2 we introduce the cost function whose minimization underpins the analysis. The Euler–Lagrange equations arising from the minimization of this cost function are derived in sections 3 and 4. The numerical procedures used to solve the Euler–Lagrange equations are described in section 5. In section 6, analytical data are used to verify that the computer code for the analysis procedure is error-free. Tests using data gathered by a mobile X-band Doppler radar, the Atmospheric Imaging Radar (AIR), of a band of heavy convective rainfall near El Reno, Oklahoma, on 4 September 2018, and a tornadic storm near Canadian, Texas, on 27 May 2015, are presented in section 7. Conclusions follow in section 8.
2. A cost function for advection correcting radial velocity data
As in SWP10a,b, we analyze data on a two-dimensional (2D) Cartesian grid on which gradually varying “large scale” advection fields U = U(x, y) and V = V(x, y) advect smaller-scale features in the data. The U and V fields are considered to be in a steady state. The analysis grid can be embedded on a surface of constant elevation angle [for data gathered in plan position indicator (PPI) mode] or on a surface of constant height [for data on constant-altitude PPI (CAPPI) surfaces]. The data on any analysis surface at any data input time are treated as if they were gathered instantaneously. A volume of data gathered as a sequence of surfaces can be analyzed sequentially, one surface at a time, with a single time stamp assigned for each surface. Some recently designed mobile Doppler radars can scan a single elevation angle in just a few seconds (e.g., French et al. 2014; Kurdzo et al. 2017; Wienhoff et al. 2018). However, operational radars such as those in the Weather Surveillance Radar-1988 Doppler (WSR-88D) network can take nearly 20 s to scan an elevation angle, even using a scan strategy designed for severe convective weather (Brown et al. 2005). In such cases, the peak displacement error on a PPI due to the assumed simultaneity of the data collection would be ~200 m in a storm translating at a speed of ~10 m s−1. The significance of such an error may vary with the application, but should be minor if the scales of interest are ~1 km or larger.
The analysis domain is a rectangle with sides of length L x parallel to the x axis, and L y parallel to the y axis. The lower left corner of the domain is at the point (x 0, y 0). The analysis window extends from an initial time t i to an end time t f . Radial velocity data are incorporated into the analysis at three data input times: the initial time t i , the end time t f , and an intermediate time 2 t m (t i < t m < t f ). The radial velocities are analyzed to all of the analysis (computational) times within the analysis window (Fig. 2). For notational simplicity, we omit the limits of integration on all integrals that span the full analysis domain and analysis window, as in the (upcoming) cost function.
Overview of the advection-correction procedure. During each outer iteration, the U and V fields are used to construct pattern-translation trajectories. A trajectory (example shown by red line) is launched forward and backward in time from every grid point at every analysis (computational) time. Radial velocity data are interpolated to each trajectory (blue arrows) at each of the three data input times. Between these times, the radial velocity is computed along each trajectory using (3.20a) and (3.20b). Since each trajectory originates from an analysis point at an analysis time, the radial velocities are now updated at every analysis point at every analysis time. The updated radial velocity (and updated DΓ/Dt and Ψ) are used to update the coefficients in (5.1) and (5.2), which are then solved via SOR to update U and V.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
The solution of (2.4) is α(x, y, t) = const (along trajectories). 3 We assign α = 0 to trajectories that enter or leave the domain at any time during the analysis window or are in a data void at any data input time, and α = 1 to trajectories on which data are available at all three data input times. Due to the “filling in” effect of the Cressman filter used to interpolate raw radar data to the Cartesian analysis grid, the data coverage at a data input time may be more extensive on the analysis grid than on the radar grid. However, since α(x, y, t) evolves as the numerical procedure iterates between evaluating Γ on trajectories, and solving the elliptic equations for U and V (section 5), the form of α(x, y, t) between data input times is not known until the procedure has converged. The κ|∇U|2 and κ|∇V|2 terms (∇ is the horizontal gradient operator, κ is a weight function) confer spatial smoothness on U and V. However, unlike the smoothness weight (β) in SWP10a,b, κ is not constant, but increases with radius as κ = Kr 2, where K (which has units of m2 s−4) is constant. The r 2 factor in κ allows the smoothness terms to keep pace with the radial growth of terms in the weak-constraint form of frozen turbulence associated with the factor r in Γ ≡ rυ r [an r 2 magnification factor for error in the weak-constraint form of frozen turbulence was noted by Matejka (2002)].
The Euler–Lagrange equations arising from (2.6)–(2.8) (sections 3 and 4) form a coupled set of differential equations for U, V, and Γ that will be solved iteratively (section 5).
In SWP10a, it was shown that many advection-correction procedures based on the frozen-turbulence constraint are subject to solution nonuniqueness associated with temporal aliasing [e.g., the reflectivity-based procedures in SWP10a,b, Gal-Chen (1982), Laroche and Zawadzki (1995), and Germann and Zawadzki (2002)]. A proof similar to that in SWP10a (not shown), establishes that the υ r -based procedure in the present study is also subject to solution nonuniqueness. If one suspects that multiple cost function minima are likely (e.g., if there are wavy features in the data), the prospect of multiple solutions can be explored by using multiple first guesses for U and V, as in Shapiro et al. (2015). Use of a low-resolution “scaling guess,” as in Laroche and Zawadzki (1995) and Germann and Zawadzki (2002), may also be desirable.
3. Euler–Lagrange equation for Γ
The integrand in the east/west boundary integral in (3.9) is proportional to α and to the boundary-normal velocity component (U). Clearly, the integrand is zero on any east/west boundary point on which U = 0. However, the integrand also vanishes on any east/west boundary point on which U ≠ 0, since any trajectory launched at any analysis time from such a point is entering or leaving the domain (so α = 0). Thus, this boundary integral is identically zero. Similarly, it can be shown that the north/south boundary integral is identically zero.
The evaluation of the triple integral in (3.9) depends on the number n of data input times. The procedure corresponding to n = 3 is described in detail in this section. In appendix A we show that an n = 2 procedure yields meaningless results. A procedure based on arbitrary n (≥3) is outlined in appendix B, but its development is beyond the scope of this study.
For each trajectory, we solve (3.12)–(3.14) on subwindows 1 and 2, with subscripts 1 and 2 affixed to the solutions on the respective subwindows.
Although it should be clear from the context whether a time integral is evaluated at a fixed point or along a trajectory, we facilitate the distinction by using the symbol τ (with primes) to represent the dummy integration variable in the time integrals that are evaluated along trajectories, as in (3.16a) and (3.16b).
In arriving at (3.21a), (3.21b) and (3.22a), (3.22b), Cauchy’s formula for repeated integration (Herrmann 2011) was used to express the double integrals as single integrals.
One then gets C 1 from (3.35), C 2 from (3.37), A 2 from (3.38), D 1, D 2, B 1, and B 2 from (3.31)–(3.34), respectively, and Γ from (3.20a) and (3.20b).
4. Euler–Lagrange equations for U and V
The integrals in (4.8) and (4.9) are evaluated at fixed locations, not along trajectories.
5. Numerical procedures
As shown in appendix D, the solution of (5.1) and (5.2) subject to the Neumann conditions (4.2a), (4.2b) and (4.10a), (4.10b) for any outer iteration level (with Γ, DΓ/Dt, and Ψ known from the previous iteration) is unique. This simplifies the implementation of the SOR solver. However, this uniqueness result only pertains to the solution of (5.1) and (5.2) at any one outer iteration level. As discussed in section 2, uniqueness of the overall converged solution for U, V, and Γ cannot be ensured.
The analysis procedure is schematized in Fig. 2. During the (N + 1)th outer iteration, a fourth-order Runge–Kutta method (Press et al. 1992) is used to construct the pattern-translation trajectories x(t), y(t) from the U, V fields at the Nth iteration. 7 Trajectories are launched forward and backward in time from every analysis point at every analysis time. At each data input time, υ r data (really Γ) and divU are bilinearly interpolated to each trajectory. If data are missing on a trajectory at any data input time, we set α = 0 for that trajectory. On trajectories for which data are available at all data input times (α = 1), we calculate Γ from (3.20a) and (3.20b), DΓ/Dt from (3.19a) and (3.19b), and Ψ from (3.15a) and (3.15b). Since a trajectory passes through each analysis point at each analysis time, this calculation also updates Γ, DΓ/Dt, and Ψ at all analysis points at all analysis times. The updated variables are used to update the coefficients in (5.1) and (5.2). These equations are solved for U and V subject to boundary conditions (4.2a), (4.2b) and (4.10a), (4.10b) using SOR with an overrelaxation coefficient set to 1.95. If the maximum change in U or V from the Nth to (N + 1)th outer iteration levels is less than a prescribed tolerance (TolOUT = 0.001 m s−1), the procedure is deemed to have converged; otherwise, the procedure returns to the trajectory calculation step. If the procedure fails to converge, the nonconvergence can be manifested as U, V values that oscillate (cycle) around unchanged values, or as continued growth of U, V to unrealistically large values, with concomitant development of large data voids as virtual particles quickly leave the domain. In the former case, if the oscillation is of sufficiently small amplitude, the procedure can still provide good estimates of the radial velocity field.
Although we have not developed a procedure to estimate appropriate values for K, an amplitude principle for K can be inferred from (2.3): if, for a dataset with radial velocity field υ r1(x, y, t), a smoothness constant K 1 has been identified for which the retrieved U and V fields are deemed physically reasonable, then for a dataset with radial velocity field υ r2(x, y, t) having the same pattern as υ r1 throughout the analysis window but an amplitude ε times that of υ r1 (υ r2 = ευ r1), the same U and V can be obtained using the smoothness constant K 2 = ε 2 K 1. Although the all-other-things-being-equal conditions for the validity of this principle can never be met in practice, the idea that larger values of K should be applied to datasets with larger radial velocity amplitudes may be useful.
6. Analytical data tests
The radial velocities are obtained from (6.5), with u and υ given by (6.7)–(6.9).
It can be shown that u and υ in the uniform translation test [i.e., given by (6.3) and (6.4) with constant U and V] satisfy the inviscid equations of motion exactly, but the u and υ fields in the solid body vortex test [i.e., given by (6.7)–(6.9) with U and V satisfying (6.6)] do not. However, since the analyses are not constrained by the dynamical underpinnings of the flow (other than the assumption that the frozen-turbulence hypothesis is at least qualitatively valid), the failure of (6.7)–(6.9) to satisfy the equations of motion should not preclude their use in our tests. Of more concern, however, is that U, V fields satisfying (6.6) cannot satisfy all of the zero-normal gradient conditions (4.2a), (4.2b) and (4.10a), (4.10b). Thus, it will not be possible to retrieve U and V exactly in the solid body vortex test case. However, as we will see, we can retrieve U and V fields in good agreement with (6.6) away from computational boundaries.
In both sets of tests, we take A = 20 m s−1, k = l = 2π/λ, and λ = 10 km. The analysis domain is a square with sides of length 30 km, and a gridpoint spacing of Δx = Δy = 250 m (121 × 121 points). The computational time step is Δt = 4 s. The analysis window is 6 min, with data supplied at 3 min intervals. In the uniform translation tests, data are generated from (6.3) and (6.4) with U = 5 m s−1 and V = 15 m s−1. In the solid body vortex tests, data are generated using (6.7)–(6.9) with Ω = 3 × 10−4s−1. The velocity fields for these cases are shown in Fig. 3.
Velocity vectors in the analytical data tests at the initial time (t = 0 s). (a) An array of vortices is in uniform translation with U = 5 m s−1 and V = 15 m s−1. (b) The array of vortices is itself rotating as a solid-body vortex. This solid-body vortex has angular velocity Ω = 3 × 10−4 s−1, and is centered on a virtual radar at the lower-left corner of the domain (x = 0 km, y = 0 km). The velocity field is constructed from (6.3) and (6.4) for (a), and (6.7)–(6.9) for (b).
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
The U, V fields retrieved in the uniform translation tests were nearly constant and nearly equal to their true values for a range of values of the smoothness parameter. In tests run with K = 0.5, 1, 5, 10, 50, 100 m2 s−4, the maximum error in U or V at any location was ≈0.5% of the true values, and the maximum error in υ r at any location or time was ≈0.12 m s−1, about 0.34% of the true peak value of υ r (≈35 m s−1).
We also ran a uniform translation test to explore the possibility of solution nonuniqueness. Following SWP10a and Shapiro et al. (2015), we note that if a tracer varies with wavelength λ in the y direction, and U true and V true are the true pattern-translation components, there are an infinite number of solutions U = U true, V = V true + Mλ/T (M is an integer, T is data input time interval) that satisfy the analysis constraints exactly. Associated with the first spurious modes M = ±1 (with T = 180 s, λ = 10 km, V true = 15 m s−1) are the aliased solutions V ≈ −40.56 m s−1 and V ≈ 70.56 m s−1. To retrieve the −40.56 m s−1 aliased solution, we left the first guess for U at 0 m s−1, but changed the first guess for V to −50 m s−1. In a test with K = 1 m2 s−4, the procedure converged with good precision to the aliased solution: U was nearly constant throughout the domain and close to its true value of 5 m s−1 (4.97 < U < 5.00 m s−1), while V was nearly constant and close to the −40.56 m s−1 aliased solution (−40.45 < V < −40.42 m s−1).
Results from a solid body vortex test in which K = 0.0002 m2 s−4 are shown in Fig. 4. As anticipated, the retrieved U, V fields are in qualitatively good agreement with the exact U, V fields in the interior of the domain, with larger errors—on the order of 1 m s−1—in the vicinity of boundaries on which the true normal gradients are inconsistent with zero-normal gradient conditions (southern/northern boundaries for U, western/eastern boundaries for V). Despite ~1 m s−1 peak errors in U and V, the peak error in υ r at any point or time was only ≈0.15 m s−1, or ≈0.75% of the true peak value of υ r which was ≈20 m s−1. An additional experiment confirmed the amplitude principle for K described in section 5: doubling the amplitude of the radial velocity (taking A = 40 m s−1) while increasing K by a factor of 4 (taking K = 0.0008 m2 s−4) produced results that differed imperceptibly from those in the above experiment.
(left) Exact analytical and (right) retrieved (top) U and (bottom) V pattern-translation fields from a test in which an array of vortices rotates as a solid body (see Fig. 3b). The smoothing parameter is K = 0.0002 m2 s−4.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
7. Tests using AIR data
The AIR is a mobile rapid-scan X-band phased array radar, designed, maintained, and operated by the Advanced Radar Research Center (ARRC) at the University of Oklahoma (Isom et al. 2013; Kurdzo et al. 2017; Mahre et al. 2018; Griffin et al. 2019). We conducted advection-correction tests using AIR data of heavy convective rainfall in a band of thunderstorms during the afternoon of 4 September 2018 near El Reno, Oklahoma, and of a decaying EF-2 rated tornado and its parent storm on 27 May 2015 near Canadian, Texas.
Here, X and Y are the analyzed and observed υ
r
fields, respectively, at the verification time, the summations are over the P analysis points on which both analyzed and observed υ
r
fields are available, and an overbar denotes an average over those P points; that is, for a variable ϕ,
To put the analysis results in context, we also ran a suite of control experiments:
-
Forward Persistence: The υ r field analyzed at the verification time t υ is the υ r field observed at the initial time t i . The velocity field is treated as stationary—no attempt is made to account for advection.
-
Backward Persistence: This run is similar to the Forward Persistence run, but with the υ r field analyzed at the verification time t υ taken as the υ r field observed at the intermediate time t m . Again, there is no advection-correction step.
-
Linear Time Interpolation: The υ r field is analyzed by linearly interpolating υ r data from the initial and intermediate times to the verification time. There is no advection-correction step.
-
Specified Constant U, V: The υ r field is advection corrected, but with U, V imposed as specified constants (U and V varied in 1 m s−1 increments from −15 to 15 m s−1) rather than derived from (5.1) and (5.2). Of these 312 = 961 experiments, the one yielding the smallest RMSE is referred to as the Best Constant U, V experiment.
-
Large K: The υ r field at the verification time is obtained from the full advection correction procedure, but with the smoothness constant K set to a value so large that the retrieved U, V fields are nearly constant. This run mimics the υ r -based Gal-Chen (1982) procedure.
In both the 4 September 2018 and 27 May 2015 deployments, the AIR was steered mechanically in azimuth, while transmitting a fan beam 20° wide in elevation and 1° wide in azimuth. For each pulse, data received by the AIR’s multielement antenna array were postprocessed using digital beamforming (Kurdzo et al. 2014). For both cases, the temporal resolution of the data was 10 s. The scan mode employed pulse compression with a 5.25-μs pulse, yielding a native range resolution of 37.5 m (Kurdzo et al. 2014, 2017). AIR data were oversampled to 30 m in range and 0.5° resolution in both azimuth and elevation (factor-of-2 oversampling). Radial velocities from 27 May 2015 were dealiased using Solo II software (Oye et al. 1995), with questionable values removed manually. A ground-clutter filter (Siggia and Passarelli 2004; Cho and Chornoboy 2005) was applied to the 4 September 2018 data. No dealiasing was needed for the 4 September 2018 case, and no ground-clutter filter was needed for the 27 May 2015 case. More details about the 27 May 2015 dataset can be found in Griffin et al. (2019).
AIR data were interpolated to a 2D Cartesian analysis grid on a constant elevation angle surface with a uniform grid spacing (Δx = Δy = 250 m) using a Cressman filter (Cressman 1959) with 400 m radius of influence. Experiments with 4 September data were conducted on the 5.5° elevation surface because that surface had the most extensive data coverage. Experiments with 27 May data were conducted on the 4° elevation surface because that level was within a layer of good data coverage, tornado strength decreased with height above that surface, and tornado size decreased (albeit slightly) beneath that surface. We report on experiments for which data are input at ~2 min intervals, and the verification time is halfway between the initial and intermediate data input times. The radial velocity fields at the three data input times are shown for both cases in Fig. 5. The tornado appears in the 27 May data as a dipole-like feature in the radial velocity field centered on a point near x = −5 km, y = 3 km. It is strongest at the initial time. Even away from the tornado, the winds in the 27 May case are generally stronger and more variable than in the 4 September case.
The AIR υ r in (left) a convective rainband on 4 Sep 2018 and (right) a decaying tornado and its parent storm on 27 May 2015 at the three data input times: (top) initial time t i , (middle) intermediate time t m , and (bottom) end time t f . The AIR is at x = 0 km, y = 0 km.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
For each dataset, we determined a threshold value of K at and beneath which the procedure did not converge, but subsequently increasing K by 0.001 m2 s−4 did yield convergence. The nonconvergence in these threshold experiments was associated with a small-amplitude “cycling” of U, V iterates around unchanging values, with only slight changes to υ r from one iteration level to the next.
Statistics from tests with the 4 September data are summarized in Table 1. The best of the advection-correction experiments had lower RMSE and higher CC than any of the non-advection-correction experiments, with the RMSE in the two persistence experiments being more than twice as large as the RMSE in any of the advection-correction experiments. The RMSE generally decreased with decreasing K, and was lowest in the K threshold run (K = 0.001 m2 s−4). Statistics from the Specified Constant U, V runs and a Large-K run are shown in Fig. 6. The RMSE in the Large-K run was similar to but somewhat larger than in the Best Constant U, V experiment, while the RMSE in the runs with the smaller values of K were less than the RMSE in the Best Constant U, V runs. The analyzed υ r , AIR υ r , and error-in-υ r fields in the K = 0.002 m2 s−4 run are presented in Fig. 7. The corresponding U, V fields are shown in Fig. 8. The dominant northward translation evident in Fig. 8 is consistent with the motion seen in animations of the reflectivity field and our perception of cloud motions during the field deployment. The error-in-υ r field (lower panel of Fig. 7) appears to contain both amplitude and phase errors. A peak error of almost 3.7 m s−1 is found near the point x = −1.0 km, y = −2.5 km, in a lobe of relatively weak radial velocities.
Experiments using AIR data from a band of heavy convective rainfall on 4 Sep 2018. Radial velocity data from the 5.5° elevation angle are supplied at ~2 min intervals (t i = 2202:53 UTC, t m = 2204:58 UTC, t f = 2207:03 UTC). The analyzed radial velocity is compared to the AIR radial velocity at t υ = 2203:55 UTC. An asterisk indicates a “K threshold” experiment in which the procedure did not converge but subsequently increasing K by 0.001 m2 s−4 did yield convergence. The Best Constant U, V run is the Specified Constant U, V run with the lowest RMSE (see Fig. 6).
(left) RMSE and (right) CC at the verification time t υ = 2203:55 UTC 4 Sep 2018 from the Specified Constant U, V experiments. The Best Constant U, V run yields RMSE ≈ 0.521 m s−1 and CC ≈ 0.985. The red dot marks results from the Large-K run (K = 1 m2 s−4), where U ≈ 0.8 m s−1, V ≈ 10.2 m s−1.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
(top left) Analyzed υ r , (top right) AIR υ r , and (bottom) error in analyzed υ r (analyzed υ r minus AIR υ r ) at the verification time t υ = 2203:55 UTC 4 Sep 2018 from the advection correction experiment in which K = 0.002 m2 s−4. The AIR is at x = 0 km, y = 0 km.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
Pattern-translation components (left) U and (right) V from a 4 Sep 2018 experiment in which K = 0.002 m2 s−4.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
Results from the 27 May case are summarized in Table 2. As in the tests with 4 September data, the best of the 27 May advection-correction experiments had lower RMSE and higher CC than any of the non-advection-correction experiments. Again, the RMSE decreased with decreasing K, and was lowest in the K threshold run (K = 0.002 m2 s−4). Statistics from the Specified Constant U, V runs and the Large-K run are shown in Fig. 9. As in the 4 September case, errors in the Large-K run were similar to but somewhat larger than in the Best Constant U, V run, while the RMSE in the runs with the smaller values of K were less than the RMSE in the Best Constant U, V run. The analyzed υ r , AIR υ r , and error-in-υ r fields in the K = 0.005 m2 s−4 run are shown in Fig. 10. The corresponding U, V fields are shown in Fig. 11. Although the errors in υ r tend to be somewhat larger near the tornado than in many other locations, the largest error (≈8.7 m s−1) is far from the tornado, near a point (x = −17.5 km, y = 4.5 km) on the northern flank of a lobe of negative radial velocities (blue blob). At this location, the AIR υ r is negative, with a magnitude of a few meters per second, while the analyzed υ r is positive and of similar magnitude. This error might be a phase error enhanced by evolution effects not properly accounted for by the procedure (size of blue blob is too small in the analysis).
Experiments using AIR data from a decaying tornado and its parent storm on 27 May 2015. Radial velocity data from the 4.0° elevation angle are supplied at ~2 min intervals (t i = 2203:52 UTC, t m = 2206:01 UTC, t f = 2207:59 UTC). The analyzed radial velocity is compared to the AIR radial velocity at t υ = 2204:56 UTC. An asterisk indicates a “K threshold” experiment in which the procedure did not converge but subsequently increasing K by 0.001 m2 s−4 did yield convergence. The Best Constant U, V run is the Specified Constant U, V run with the lowest RMSE (see Fig. 9).
(left) RMSE and (right) CC at the verification time t υ = 2204:56 UTC 27 May 2015 from the Specified Constant U, V runs. The Best Constant U, V run yields RMSE ≈ 1.696 m s−1 and CC ≈ 0.948. The red dot marks results from the Large-K run (K = 10 m2 s−4), where U ≈ −3.2 m s−1, V ≈ 5.7 m s−1.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
(top left) Analyzed υ r , (top right) AIR υ r , and (bottom) error in analyzed υ r (analyzed υ r minus AIR υ r ) at the verification time t υ = 2204:56 UTC 27 May 2015 from the advection correction experiment in which K = 0.005 m2 s−4. The AIR is at x = 0 km, y = 0 km.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
Pattern-translation components (left) U and (right) V from a 27 May 2015 experiment in which K = 0.005 m2 s−4.
Citation: Journal of the Atmospheric Sciences 78, 1; 10.1175/JAS-D-20-0048.1
Tables 1 and 2 show that the value of K used in the 27 May Large-K run is 10 times larger than in the 4 September Large-K run. We settled on these K values because they produced similar (small) spatial changes to U and V in both datasets, with peak domain-wide variations of ≈0.05 m s−1. Additionally, the threshold for K at and beneath which the procedure did not converge was about twice as large in the 27 May case than in the 4 September case. Despite their failure to converge, these threshold tests produced the lowest RMSE and largest CC for both datasets. The relatively larger values of K in the Large-K and threshold runs in the 27 May case compared to the 4 September case are qualitatively consistent with the amplitude principle for K discussed in section 5. In additional tests in which K was beneath its threshold value and further decreased, the amplitude of the U, V–iteration cycle increased, and the procedure eventually became unstable.
The analytical and real data tests described herein were obtained with parameter settings that had not been optimized for computational efficiency. Additional tests using AIR data (not shown) showed that a significant speedup could be achieved simply by increasing the tolerance (TolOUT) for convergence of U and V in the outer loop. However, to prevent the procedure from converging prematurely in that loop—which could substantially increase analysis errors—the blending parameter wgt in (5.4) should be increased in tandem with the increase in TolOUT. When the original 27 May analysis using K = 0.005 m2 s−4 was rerun with TolOUT increased a hundredfold to 0.1 m s−1 and wgt increased tenfold to 0.25, the execution time decreased from 112 to 3 min. Similarly, when the original 4 September analysis using K = 0.002 m2 s−4 was rerun using these same increased parameters, the execution time decreased from 87 to 5 min. In both sped-up runs, the increase in RMSE (ΔRMSE) over the RMSE in the corresponding original run was negligible (ΔRMSE ≈ 0.007 m s−1 for the 27 May analysis and ≈0.002 m s−1 for the 4 September analysis). Similar speedups were achieved by setting TolOUT to 0.1 m s−1 without changing wgt, but at the expense of a much larger ΔRMSE (ΔRMSE ≈ 0.174 m s−1 for the 27 May analysis and ≈0.324 m s−1 for the 4 September analysis).
8. Conclusions
A procedure to advection correct (analyze) the Doppler radial velocity field using spatially variable pattern-translation components (U and V) is developed, with encouraging results obtained in proof-of-concept tests using data from the Atmospheric Imaging Radar (AIR). The advection-correction procedure is based on the minimization of a cost function that accumulates errors in a second-derivative frozen-turbulence constraint on the radial velocity field, while penalizing spatial gradients in U and V. The radial velocity field and associated U, V fields are analyzed together as the iterative solution of the Euler–Lagrange equations arising from the minimization of the cost function. Radial velocities from three data input times are incorporated into the analysis. As in many advection-correction techniques that use the reflectivity field as a tracer, the new radial-velocity-based technique may be subject to solution nonuniqueness associated with temporal aliasing. Solution nonuniqueness was demonstrated in an analytical test in which the first guesses for U and V were set close to values associated with a spurious solution, but did not appear to be an issue in the AIR data tests.
After performing verification tests using analytical data, we conducted experiments using AIR data from a convective rainband on 4 September 2018 and a decaying tornado and its parent storm on 27 May 2015. In these AIR data tests, the procedure yielded radial velocities with lower RMSE and higher CC than radial velocities analyzed in any of the control runs (Forward Persistence, Backward Persistence, Linear Time Interpolation, Specified Constant U, V, and Large K).
The procedure is sensitive to the value of a smoothness parameter (K). In AIR data tests spanning three orders of magnitude of K, the best results were obtained with relatively small values of K, though for values not much smaller than these, the procedure became unstable. For relatively large K, the retrieved U, V fields were nearly constant, and the errors in the analyzed radial velocity field were similar to (though slightly worse than) those from the Best Constant U, V experiment. As the procedure lacks an objective method to determine appropriate values of K, we recommend that such values be explored on a case-by-case basis, using trial and error, perhaps guided by visual estimates of physically reasonable U, V fields from animations of the data.
The stability and execution time of the procedure are sensitive to computational parameters such as the convergence thresholds (tolerances) for U and V on the inner and outer iteration loops and an underrelaxation (blending) parameter used to weight the updated and previous iterates on the outer loop. Experimentation may also help determine appropriate values of these parameters on a case-by-case basis.
The primary utility of the new procedure is to the analysis of nonsimultaneously collected radial velocity data from two or more radars to common analysis times for use in multiple-Doppler wind syntheses of severe convective phenomena. Since some experimentation is required to establish appropriate smoothing and computational parameter values or to explore the threat of solution nonuniqueness, the procedure is currently more suitable for non-real-time research applications than use in an operational analysis or data assimilation setting. Future work will focus on determining procedures to estimate some of these parameters objectively.
Acknowledgments
This research was supported by the National Science Foundation (NSF) under Grant AGS-1623626. Data collection and analysis of the 27 May 2015 case were supported by NSF AGS-1303685 and AGS-1823478; Jim Kurdzo led the deployment, Javier Lujan assisted with the data collection, and Casey Griffin dealiased and quality controlled the data. The AIR is maintained and operated by the Advanced Radar Research Center (ARRC) of the University of Oklahoma. We are grateful to the ARRC engineering staff for their support of our radar experiments. Shawn Riley provided computing assistance. The authors thank Lou Wicker, Qin Xu, Nikola Petrov, and the anonymous referees for their helpful comments.
Data availability statement
The AIR data and Fortran code used to generate the analyses reported on herein are available from the corresponding author upon request.
APPENDIX A
Analysis Using Two Time Levels of Data
If υ r data are only available at two data input times, then (3.11) and the equivalent system (3.12)–(3.14) apply throughout the analysis window, and (3.18a) and (3.18b) are replaced by the single equation Θ = (A 1 t + B 1)E 1(t). In view of (3.10a), (3.10b), and (3.12), A 1 = B 1 = 0, so D 2Γ/Dt 2 is identically zero. Equations (3.12) and (3.13) then show that Θ and Ψ are identically zero, and (4.8) and (4.9) reduce to Laplace’s equations for U and V, the solutions of which [subject to the zero-normal-gradient conditions (4.2a), (4.2b) and (4.10a), (4.10b)] are arbitrary constants, with no links to the υ r field. Such an analysis has no practical value.
APPENDIX B
Analysis Using an Arbitrary Number of Time Levels of Data
To incorporate data from an arbitrary number n of input times, the analysis window should include n − 2 intermediate data input times and n − 1 analysis subwindows. The solution of (3.11) [analog of (3.20a) and (3.20b)] then contains 4(n − 1) integration constants, which satisfy 4(n − 1) linear algebraic equations: 1 equation from imposing data at the initial time, 1 equation from imposing data at the end time, 2(n − 2) equations from imposing data at the intermediate data input times (as in the n = 3 case, data at each intermediate data input time are used twice), 2 equations from (3.10a), (3.10b), and 2(n − 2) equations expressing continuity of DΓ/Dt and D 2Γ/Dt 2 at the intermediate data input times. The integration constants can thus be obtained by solving 4(n − 1) linear algebraic equations. U and V then follow as solutions of (4.8) and (4.9) subject to boundary conditions (4.2a), (4.2b) and (4.10a), (4.10b).
APPENDIX C
Classification of (4.8) and (4.9) for U and V
The second-order quasi-linear operator (C.1) is classified using the same procedure as for linear operators (Garabedian 1998). Since κ, α, t f − t i , and (Γ x )2 are nonnegative, (C.3) shows that γ is nonnegative. Equations (C.2) then show that A, C, and AC − B 2 are strictly positive for all U and V, so (C.1) is an elliptic operator. Similarly, it can be shown that the operator analogous to (C.1) arising from (4.9) is elliptic.
APPENDIX D
Comment on the Forcing Terms in (4.8) and (4.9)
As there are no restrictions on Q N that would make its integral vanish, (D.4) will generally be violated, and no solution for U N+1 exists.
As the discriminant c 2 − bg in the quadratic formula solution of (D.15) for U d as a function of V d is negative (c 2 ≤ bg), U d is imaginary if V d is real. However, since U d and V d should be real, we must have U d = V d = 0 (U 2 = U 1, V 2 = V 1), and the solution is unique.
REFERENCES
Anagnostou, E. N. , and W. F. Krajewski , 1999: Real-time radar rainfall estimation. Part I: Algorithm formulation. J. Atmos. Oceanic Technol., 16, 189–197, https://doi.org/10.1175/1520-0426(1999)016<0189:RTRREP>2.0.CO;2.
Austin, P. M. , 1987: Relation between measured radar reflectivity and surface rainfall. Mon. Wea. Rev., 115, 1053–1070, https://doi.org/10.1175/1520-0493(1987)115<1053:RBMRRA>2.0.CO;2.
Bellon, A. , and I. Zawadzki , 1994: Forecasting of hourly accumulations of precipitation by optimal extrapolation of radar maps. J. Hydrol., 157, 211–233, https://doi.org/10.1016/0022-1694(94)90106-6.
Bousquet, O. , P. Tabary , and J. Parent du Châtelet , 2007: On the value of operationally synthesized multiple-Doppler wind fields. Geophys. Res. Lett., 34, L22813, https://doi.org/10.1029/2007GL030464.
Brown, R. A. , R. M. Steadham , B. A. Flickinger , R. R. Lee , D. Sirmans , and V. T. Wood , 2005: New WSR-88D volume coverage pattern 12: Results of field tests. Wea. Forecasting, 20, 385–393, https://doi.org/10.1175/WAF848.1.
Burghelea, T. , E. Segre , and V. Steinberg , 2005: Validity of the Taylor hypothesis in a random spatially smooth flow. Phys. Fluids, 17, 103101, https://doi.org/10.1063/1.2077367.
Carbone, R. E. , 1982: A severe frontal rainband. Part I: Stormwide hydrodynamic structure. J. Atmos. Sci., 39, 258–279, https://doi.org/10.1175/1520-0469(1982)039<0258:ASFRPI>2.0.CO;2.
Castro, J. J. , A. A. Carsteanu , and J. D. Fuentes , 2011: On the phenomenology underlying Taylor’s hypothesis in atmospheric turbulence. Rev. Mex. Fis., 57, 60–64.
Cho, J. Y. N. , and E. S. Chornoboy , 2005: Multi-PRI signal processing for the terminal Doppler weather radar. Part I: Clutter filtering. J. Atmos. Oceanic Technol., 22, 575–582, https://doi.org/10.1175/JTECH1730.1.
Chong, M. , J. Testud , and F. Roux , 1983: Three-dimensional wind field analysis from dual-Doppler radar data. Part II: Minimizing the error due to temporal variation. J. Climate Appl. Meteor., 22, 1216–1226, https://doi.org/10.1175/1520-0450(1983)022<1216:TDWFAF>2.0.CO;2.
Clark, T. L. , F. I. Harris , and C. G. Mohr , 1980: Errors in wind fields derived from multiple-Doppler radars: Random errors and temporal errors associated with advection and evolution. J. Appl. Meteor., 19, 1273–1284, https://doi.org/10.1175/1520-0450(1980)019<1273:EIWFDF>2.0.CO;2.
Cressman, G. P. , 1959: An operational objective analysis system. Mon. Wea. Rev., 87, 367–374, https://doi.org/10.1175/1520-0493(1959)087<0367:AOOAS>2.0.CO;2.
Creutin, J.-D. , E. Leblois , and J.-M. Lepioufle , 2015: Unfreezing Taylor’s hypothesis for precipitation. J. Hydrometeor., 16, 2443–2462, https://doi.org/10.1175/JHM-D-14-0120.1.
Dennis, D. J. C. , and T. B. Nickels , 2008: On the limitations of Taylor’s hypothesis in constructing long structures in a turbulent boundary layer. J. Fluid Mech., 614, 197–206, https://doi.org/10.1017/S0022112008003352.
Dowell, D. C. , and H. B. Bluestein , 1997: The Arcadia, Oklahoma, storm of 17 May 1981: Analysis of a supercell during tornadogenesis. Mon. Wea. Rev., 125, 2562–2582, https://doi.org/10.1175/1520-0493(1997)125<2562:TAOSOM>2.0.CO;2.
Elsgolc, L. E. , 1961: Calculus of Variations. Pergamon Press, 178 pp.
Fabry, F. , A. Bellon , M. R. Duncan , and G. L. Austin , 1994: High resolution rainfall measurements by radar for very small basins: The sampling problem reexamined. J. Hydrol., 161, 415–428, https://doi.org/10.1016/0022-1694(94)90138-4.
French, M. M. , H. B. Bluestein , I. PopStefanija , C. A. Baldi , and R. T. Bluth , 2014: Mobile, phased-array, Doppler radar observations of tornadoes at X band. Mon. Wea. Rev., 142, 1010–1036, https://doi.org/10.1175/MWR-D-13-00101.1.
Gal-Chen, T. , 1982: Errors in fixed and moving frame of references: Applications for conventional and Doppler radar analysis. J. Atmos. Sci., 39, 2279–2300, https://doi.org/10.1175/1520-0469(1982)039<2279:EIFAMF>2.0.CO;2.
Garabedian, P. R. , 1998: Partial Differential Equations. 2nd ed. Chelsea Publishing Co., 672 pp.
Germann, U. , and I. Zawadzki , 2002: Scale-dependence of the predictability of precipitation from continental radar images. Part I: Description of the methodology. Mon. Wea. Rev., 130, 2859–2873, https://doi.org/10.1175/1520-0493(2002)130<2859:SDOTPO>2.0.CO;2.
Gerstner, E.-M. , and G. Heinemann , 2008: Real-time areal precipitation determination from radar by means of statistical objective analysis. J. Hydrol., 352, 296–308, https://doi.org/10.1016/j.jhydrol.2008.01.016.
Griffin, C. B. , D. J. Bodine , J. M. Kurdzo , A. Mahre , and R. D. Palmer , 2019: High temporal resolution observations of the 27 May 2015 Canadian, Texas, tornado using the Atmospheric Imaging Radar. Mon. Wea. Rev., 147, 873–891, https://doi.org/10.1175/MWR-D-18-0297.1.
Gustafson, K. E. , 1987: Introduction of Partial Differential Equations and Hilbert Space Methods. 2nd ed. John Wiley and Sons, 409 pp.
Harrison, D. L. , R. W. Scovell , and M. Kitchen , 2009: High-resolution precipitation estimates for hydrological uses. Proc. Inst. Civ. Eng. Water Manage., 162, 125–135, https://doi.org/10.1680/wama.2009.162.2.125.
Herrmann, R. , 2011: Fractional Calculus: An Introduction for Physicists. World Scientific Publishing, 276 pp.
Heymsfield, G. M. , 1978: Kinematic and dynamic aspects of the Harrah tornadic storm analyzed from dual-Doppler radar data. Mon. Wea. Rev., 106, 233–254, https://doi.org/10.1175/1520-0493(1978)106<0233:KADAOT>2.0.CO;2.
Higgins, C. W. , M. Froidevaux , V. Simeonov , N. Vercauteren , C. Barry , and M. B. Parlange , 2012: The effect of scale on the applicability of Taylor’s frozen turbulence hypothesis in the atmospheric boundary layer. Bound.-Layer Meteor., 143, 379–391, https://doi.org/10.1007/s10546-012-9701-1.
Hildebrand, P. H. , and C. K. Mueller , 1985: Evaluation of meteorological airborne Doppler radar. Part I: Dual-Doppler analyses of air motions. J. Atmos. Oceanic Technol., 2, 362–380, https://doi.org/10.1175/1520-0426(1985)002<0362:EOMADR>2.0.CO;2.
Isom, B. , and Coauthors, 2013: The Atmospheric Imaging Radar: Simultaneous volumetric observations using a phased array weather radar. J. Atmos. Oceanic Technol., 30, 655–675, https://doi.org/10.1175/JTECH-D-12-00063.1.
Kurdzo, J. M. , B. L. Cheong , R. D. Palmer , G. Zhang , and J. B. Meier , 2014: A pulse compression waveform for improved-sensitivity weather radar observations. J. Atmos. Oceanic Technol., 31, 2713–2731, https://doi.org/10.1175/JTECH-D-13-00021.1.
Kurdzo, J. M. , and Coauthors, 2017: Observations of severe local storms and tornadoes with the Atmospheric Imaging Radar. Bull. Amer. Meteor. Soc., 98, 915–935, https://doi.org/10.1175/BAMS-D-15-00266.1.
Lanczos, C. , 1986: The Variational Principles of Mechanics. Dover, 418 pp.
Laroche, S. , and I. Zawadzki , 1995: Retrievals of horizontal winds from single-Doppler clear-air data by methods of cross correlation and variational analysis. J. Atmos. Oceanic Technol., 12, 721–738, https://doi.org/10.1175/1520-0426(1995)012<0721:ROHWFS>2.0.CO;2.
Lazarus, S. , A. Shapiro , and K. Droegemeier , 2001: Application of the Zhang–Gal-Chen single-Doppler velocity retrieval to a deep convective storm. J. Atmos. Sci., 58, 998–1016, https://doi.org/10.1175/1520-0469(2001)058<0998:AOTZGC>2.0.CO;2.
Li, B. , A. Murthi , K. P. Bowman , G. R. North , M. G. Genton , and M. Sherman , 2009: Statistical tests of Taylor’s hypothesis: An application to precipitation fields. J. Hydrometeor., 10, 254–265, https://doi.org/10.1175/2008JHM1009.1.
Liou, Y.-C. , 1999: Single radar recovery of cross-beam wind components using a modified moving frame of reference technique. J. Atmos. Oceanic Technol., 16, 1003–1016, https://doi.org/10.1175/1520-0426(1999)016<1003:SRROCB>2.0.CO;2.
Liou, Y.-C. , 2007: Single Doppler retrieval of the three-dimensional wind in a deep convective system based on an optimal moving frame of reference. J. Meteor. Soc. Japan, 85, 559–582, https://doi.org/10.2151/jmsj.85.559.
Liou, Y.-C. , and I.-S. Luo , 2001: An investigation of the moving-frame single-Doppler wind retrieval technique using Taiwan Area Mesoscale Experiment low-level data. J. Appl. Meteor., 40, 1900–1917, https://doi.org/10.1175/1520-0450(2001)040<1900:AIOTMF>2.0.CO;2.
Liu, C. , and W. F. Krajewski , 1996: A comparison of methods for calculation of radar-rainfall hourly accumulations. J. Amer. Water Resour. Assoc., 32, 305–315, https://doi.org/10.1111/j.1752-1688.1996.tb03453.x.
Liu, S. , C. Qiu , Q. Xu , and P. Zhang , 2004: An improved time interpolation for three-dimensional Doppler wind analysis. J. Appl. Meteor., 43, 1379–1391, https://doi.org/10.1175/JAM2150.1.
Mahre, A. , J. M. Kurdzo , D. J. Bodine , C. B. Griffin , R. D. Palmer , and T.-Y. Yu , 2018: Analysis of the 16 May 2015 Tipton, Oklahoma, EF-3 tornado at high spatiotemporal resolution using the Atmospheric Imaging Radar. Mon. Wea. Rev., 146, 2103–2124, https://doi.org/10.1175/MWR-D-17-0256.1.
Matejka, T. , 2002: Estimating the most steady frame of reference from Doppler radar data. J. Atmos. Oceanic Technol., 19, 1035–1048, https://doi.org/10.1175/1520-0426(2002)019<1035:ETMSFO>2.0.CO;2.
Mittermaier, M. P. , R. J. Hogan , and A. J. Illingworth , 2004: Using mesoscale model winds for correcting wind-drift errors in radar estimates of surface rainfall. Quart. J. Roy. Meteor. Soc., 130, 2105–2123, https://doi.org/10.1256/qj.03.156.
Nielsen, J. E. , S. Thorndahl , and M. R. Rasmussen , 2014: A numerical method to generate high temporal resolution precipitation time series by combining weather radar measurements with a nowcast model. Atmos. Res., 138, 1–12, https://doi.org/10.1016/j.atmosres.2013.10.015.
Oue, M. , P. Kollias , A. Shapiro , A. Tatarevic , and T. Matsui , 2019: Investigation of observational error sources in multi-Doppler-radar three-dimensional variational vertical air motion retrievals. Atmos. Meas. Tech., 12, 1999–2018, https://doi.org/10.5194/amt-12-1999-2019.
Oye, R. , C. Mueller , and S. Smith , 1995: Software for radar translation, visualization, editing, and interpolation. Preprints, 27th Conf. on Radar Meteorology, Vail, CO, Amer. Meteor. Soc., 359–361.
Press, W. H. , S. A. Teukolsky , W. T. Vetterling , and B. P. Flannery , 1992: Numerical Recipes in FORTRAN: The Art of Scientific Computing. 2nd ed. Cambridge University Press, 963 pp.
Qiu, C.-J. , and Q. Xu , 1992: A simple adjoint method of wind analysis for single-Doppler data. J. Atmos. Oceanic Technol., 9, 588–598, https://doi.org/10.1175/1520-0426(1992)009<0588:ASAMOW>2.0.CO;2.
Rosenhead, L. , Ed., 1963: Laminar Boundary Layers. Oxford University Press, 688 pp.
Rudin, W. , 1976: Principles of Mathematical Analysis. 3rd ed. McGraw-Hill, 342 pp.
Sasaki, Y. , 1970: Some basic formalisms in numerical variational analysis. Mon. Wea. Rev., 98, 875–883, https://doi.org/10.1175/1520-0493(1970)098<0875:SBFINV>2.3.CO;2.
Seo, B.-C. , and W. F. Krajewski , 2015: Correcting temporal sampling error in radar-rainfall: Effect of advection parameters and storm characteristics on the correction accuracy. J. Hydrol., 531, 272–283, https://doi.org/10.1016/j.jhydrol.2015.04.018.
Shapiro, A. , S. Ellis , and J. Shaw , 1995: Single-Doppler velocity retrievals with Phoenix II data: Clear air and microburst wind retrievals in the planetary boundary layer. J. Atmos. Sci., 52, 1265–1287, https://doi.org/10.1175/1520-0469(1995)052<1265:SDVRWP>2.0.CO;2.
Shapiro, A. , K. M. Willingham , and C. K. Potvin , 2010a: Spatially variable advection correction of radar data. Part I: Theoretical considerations. J. Atmos. Sci., 67, 3445–3456, https://doi.org/10.1175/2010JAS3465.1.
Shapiro, A. , K. M. Willingham , and C. K. Potvin , 2010b: Spatially variable advection correction of radar data. Part II: Test results. J. Atmos. Sci., 67, 3457–3470, https://doi.org/10.1175/2010JAS3466.1.
Shapiro, A. , S. Rahimi , C. K. Potvin , and L. Orf , 2015: On the use of advection correction in trajectory calculations. J. Atmos. Sci., 72, 4261–4280, https://doi.org/10.1175/JAS-D-15-0095.1.
Siggia, A. D. , and R. E. Passarelli , 2004: Gaussian model adaptive processing (GMAP) for improved ground clutter cancellation and moment calculation. Proc. Third European Conf. on Radar Meteorology, Visby, Sweden, ERAD, 67–73.
Smull, B. F. , and R. A. Houze , 1987: Dual-Doppler radar analysis of a midlatitude squall line with a trailing region of stratiform rain. J. Atmos. Sci., 44, 2128–2149, https://doi.org/10.1175/1520-0469(1987)044<2128:DDRAOA>2.0.CO;2.
Taylor, G. I. , 1938: The spectrum of turbulence. Proc. Roy. Soc. London, 164A, 476–490, https://doi.org/10.1098/rspa.1938.0032.
Thorndahl, S. , J. E. Nielsen , and M. R. Rasmussen , 2014: Bias adjustment and advection interpolation of long-term high resolution radar rainfall series. J. Hydrol., 508, 214–226, https://doi.org/10.1016/j.jhydrol.2013.10.056.
Tuttle, J. D. , and G. B. Foote , 1990: Determination of the boundary layer airflow from a single Doppler radar. J. Atmos. Oceanic Technol., 7, 218–232, https://doi.org/10.1175/1520-0426(1990)007<0218:DOTBLA>2.0.CO;2.
Uddin, A. K. M. , A. E. Perry , and I. Marusic , 1997: On the validity of Taylor’s hypothesis in wall turbulence. J. Mech. Eng. Res. Dev., 19–20, 57–66.
Wienhoff, Z. B. , H. B. Bluestein , L. J. Wicker , J. C. Snyder , A. Shapiro , C. K. Potvin , J. B. Houser , and D. W. Reif , 2018: Applications of a spatially variable advection correction technique for temporal correction of dual-Doppler analyses of tornadic supercells. Mon. Wea. Rev., 146, 2949–2971, https://doi.org/10.1175/MWR-D-17-0360.1.
Wilczek, M. , H. Xu , and Y. Narita , 2014: A note on Taylor’s hypothesis under large-scale flow variation. Nonlinear Processes Geophys., 21, 645–649, https://doi.org/10.5194/npg-21-645-2014.
Wurman, J. , Y. Richardson , C. Alexander , S. Weygandt , and P. F. Zhang , 2007: Dual-Doppler analysis of winds and vorticity budget terms near a tornado. Mon. Wea. Rev., 135, 2392–2405, https://doi.org/10.1175/MWR3404.1.
Wurman, J. , K. A. Kosiba , P. Markowski , Y. Richardson , D. Dowell , and P. Robinson , 2010: Finescale single- and dual-Doppler analysis of tornado intensification, maintenance, and dissipation in the Orleans, Nebraska, supercell. Mon. Wea. Rev., 138, 4439–4455, https://doi.org/10.1175/2010MWR3330.1.
Zawadzki, I. I. , 1973: Statistical properties of precipitation patterns. J. Appl. Meteor., 12, 459–472, https://doi.org/10.1175/1520-0450(1973)012<0459:SPOPP>2.0.CO;2.
Zhang, J. , and T. Gal-Chen , 1996: Single-Doppler wind retrieval in the moving frame of reference. J. Atmos. Sci., 53, 2609–2623, https://doi.org/10.1175/1520-0469(1996)053<2609:SDWRIT>2.0.CO;2.
We use the strong- and weak-constraint terminology of Sasaki (1970). A constraint that is imposed exactly is a strong constraint. A constraint that is imposed approximately (i.e., in a least squares error sense) is a weak constraint.
At least three time levels of input data were also required in the υ r -based procedure of Gal-Chen (1982), in which a ∂2 υ r /∂t 2 term was discretized. As shown in appendix A, minimizing (2.3) with υ r data incorporated at only two time levels yields meaningless results.
In the numerical solution procedure (section 5), trajectories are launched forward and backward in time from every analysis point at every analysis/computational time.
Treated as field variables, U, V, and divU are in a steady state. However, due to their spatial variations, these same variables vary with time along trajectories. We obtain U, V, and divU along trajectories by interpolating their values from neighboring grid points to the location of the virtual particle tracing out the trajectory.
To ensure continuity of Γ at time t m , the input data are used twice: for Γ1 at the end of subwindow 1, and for Γ2 at the beginning of subwindow 2.
Second derivative terms in U and V which appear in the expansion of Ψ must be taken into account in the classification of (4.8) and (4.9) (appendix C).
Unless stated otherwise, the first guesses for U and V in our experiments (i.e., at the first iteration) are 0 m s−1.
We refer to “large” or “small” values of K in a relative sense, that is, with respect to a given dataset. A value of K large enough to produce nearly constant U and V in one dataset may be too small to produce nearly constant U and V in another dataset, even if the two datasets have the same spatial and temporal resolution (as in the two AIR datasets considered in section 7).