Boundary-Layer Meteorol (2013) 149:455481

DOI 10.1007/s10546-013-9848-4
ARTICLE

A Scale-Adaptive Approach for Spatially-Varying Urban

Morphology Characterization in Boundary Layer
Parametrization Using Multi-Resolution Analysis
P. Mouzourides A. Kyprianou M. K.-A. Neophytou

Received: 4 December 2012 / Accepted: 22 July 2013 / Published online: 15 August 2013
Springer Science+Business Media Dordrecht 2013

Abstract Urban morphology characterization is crucial for the parametrization of boundarylayer development over urban areas. One complexity in such a characterization is the threedimensional variation of the urban canopies and textures, which are customarily reduced
to and represented by one-dimensional varying parametrization such as the aerodynamic
roughness length z 0 and zero-plane displacement d. The scope of the paper is to provide
novel means for a scale-adaptive spatially-varying parametrization of the boundary layer
by addressing this 3-D variation. Specifically, the 3-D variation of urban geometries often
poses questions in the multi-scale modelling of air pollution dispersion and other climate or
weather-related modelling applications that have not been addressed yet, such as: (a) how
we represent urban attributes (parameters) appropriately for the multi-scale nature and multiresolution basis of weather numerical models, (b) how we quantify the uniqueness of an urban
database in the context of modelling urban effects in large-scale weather numerical models,
and (c) how we derive the impact and influence of a particular building in pre-specified
sub-domain areas of the urban database. We illustrate how multi-resolution analysis (MRA)
addresses and answers the afore-mentioned questions by taking as an example the Central
Business District of Oklahoma City. The selection of MRA is motivated by its capacity
for multi-scale sampling; in the MRA the urban signal depicting a city is decomposed
into an approximation, a representation at a higher scale, and a detail, the part removed at
lower scales to yield the approximation. Different levels of approximations were deduced for
the building height H and planar packing density p . A spatially-varying characterization
with a scale-adaptive capacity is obtained for the boundary-layer parameters (aerodynamic

Electronic supplementary material The online version of this article (doi:10.1007/s10546-013-9848-4)

contains supplementary material, which is available to authorized users.
P. Mouzourides M. K.-A. Neophytou (B)
Environmental Fluid Mechanics Laboratory, Department of Civil and Environmental Engineering,
University of Cyprus, Nicosia, Cyprus
e-mail: neophytou@ucy.ac.cy
A. Kyprianou
Department of Mechanical and Manufacturing Engineering, University of Cyprus, Nicosia, Cyprus

123

456

P. Mouzourides et al.

roughness length z 0 and zero-plane displacement d) using the MRA-deduced results for the
building height and the planar packing density with a morphometric model; an attribute that
is shown to be of great advantage to multi-scale and multi-resolution numerical weather
prediction models.
Keywords Building morphology Multi-resolution analysis Multi-scale sampling
Multi-scale/multi-resolution numerical weather prediction models Subgrid parametrizations
1 Introduction
The impact of urbanization has become increasingly greater in the last few decades and has
been made realizable in a variety of aspects of life, ranging from the urban microclimate and
more broadly atmospheric research, to the energy demands of pedestrian thermal comfort in
a city. A focus in atmospheric research has been the appropriate representation of the urban
effects in various numerical models, and particularly of their aerodynamic characteristics,
used to describe the vertical variation of the horizontal mean wind speed (U ). Urban morphology characterization is crucial for the parametrization of boundary- layer development
over urban areas; one of the main complications is the three-dimensional (3-D) variation of
the urban canopies and textures, which are often reduced to and represented by a 1-D varying
parametrization.
Based on the MoninObukhov similarity theory, the vertical profile of the mean horizontal
wind speed can be described by


u
zd
U (z) =
ln
(1)
k
z0
where k = 0.4 is the von Karman constant, u is the friction velocity, and z 0 and d are the
aerodynamic roughness length and the zero-plane displacement, respectively. It is important
to note that Eq. 1 is valid within the constant-stress layer (the inertial sublayer) and in
conditions of neutral stability. In addition, it is noted that the aerodynamic parameters z 0 and d
are physically realistic only if Eq. 1 is applied over a statistically homogeneous fetch (Hamlyn
et al. 2007). The advantage of using Eq. 1 within numerical weather prediction models resides
in the continuity with the surrounding rural areas, where a similar profile to that depicted
in Eq. 1 is adopted but with different z 0 and d parameters; in addition, the use of Eq. 1 has
a negligible impact on the computational costs (Martilli and Santiago 2007). Unfortunately
however, due to the inherent complexities of urban geometries there are great difficulties in
correctly estimating the z 0 and d parameters for urban surfaces. Two classes of methods have
been developed in order to estimate z 0 and d: (i) the morphometric methods that use the
building elevation data to calculate the set of geometric input parameters that the forecasting
model requires (e.g. Ratti et al. 2002; Di Sabatino et al. 2010), and (ii) the micrometeorological
methods that use field observations (e.g. Grimmond 1998; Hagishima et al. 2009; Zaki et al.
2011). In the frequent absence however of complete sets of wind-speed measurements, z 0
and d are estimated via the use of morphometric models, which relate these aerodynamic
parameters to various geometric properties of the urban surface (Grimmond and Oke 1999).
These models become increasingly attractive as substantial and detailed urban morphological
data become available (e.g. Cities Revealed The GeoInformation Group 2008).
Moreover, it is well-established (e.g. Britter and Hanna 2003) that the airflow and associated dispersion processes within and above urban areas occur over a wide range of spatial
scales. Four horizontal spatial scales are introduced: (i) The regional scale, that is the larger
surrounding area that is mutually influenced by the city area, extending up to several hundred

123

456

P. Mouzourides et al.

roughness length z 0 and zero-plane displacement d) using the MRA-deduced results for the
building height and the planar packing density with a morphometric model; an attribute that
is shown to be of great advantage to multi-scale and multi-resolution numerical weather
prediction models.
Keywords Building morphology Multi-resolution analysis Multi-scale sampling
Multi-scale/multi-resolution numerical weather prediction models Subgrid parametrizations
1 Introduction
The impact of urbanization has become increasingly greater in the last few decades and has
been made realizable in a variety of aspects of life, ranging from the urban microclimate and
more broadly atmospheric research, to the energy demands of pedestrian thermal comfort in
a city. A focus in atmospheric research has been the appropriate representation of the urban
effects in various numerical models, and particularly of their aerodynamic characteristics,
used to describe the vertical variation of the horizontal mean wind speed (U ). Urban morphology characterization is crucial for the parametrization of boundary- layer development
over urban areas; one of the main complications is the three-dimensional (3-D) variation of
the urban canopies and textures, which are often reduced to and represented by a 1-D varying
parametrization.
Based on the MoninObukhov similarity theory, the vertical profile of the mean horizontal
wind speed can be described by


u
zd
U (z) =
ln
(1)
k
z0
where k = 0.4 is the von Karman constant, u is the friction velocity, and z 0 and d are the
aerodynamic roughness length and the zero-plane displacement, respectively. It is important
to note that Eq. 1 is valid within the constant-stress layer (the inertial sublayer) and in
conditions of neutral stability. In addition, it is noted that the aerodynamic parameters z 0 and d
are physically realistic only if Eq. 1 is applied over a statistically homogeneous fetch (Hamlyn
et al. 2007). The advantage of using Eq. 1 within numerical weather prediction models resides
in the continuity with the surrounding rural areas, where a similar profile to that depicted
in Eq. 1 is adopted but with different z 0 and d parameters; in addition, the use of Eq. 1 has
a negligible impact on the computational costs (Martilli and Santiago 2007). Unfortunately
however, due to the inherent complexities of urban geometries there are great difficulties in
correctly estimating the z 0 and d parameters for urban surfaces. Two classes of methods have
been developed in order to estimate z 0 and d: (i) the morphometric methods that use the
building elevation data to calculate the set of geometric input parameters that the forecasting
model requires (e.g. Ratti et al. 2002; Di Sabatino et al. 2010), and (ii) the micrometeorological
methods that use field observations (e.g. Grimmond 1998; Hagishima et al. 2009; Zaki et al.
2011). In the frequent absence however of complete sets of wind-speed measurements, z 0
and d are estimated via the use of morphometric models, which relate these aerodynamic
parameters to various geometric properties of the urban surface (Grimmond and Oke 1999).
These models become increasingly attractive as substantial and detailed urban morphological
data become available (e.g. Cities Revealed The GeoInformation Group 2008).
Moreover, it is well-established (e.g. Britter and Hanna 2003) that the airflow and associated dispersion processes within and above urban areas occur over a wide range of spatial
scales. Four horizontal spatial scales are introduced: (i) The regional scale, that is the larger
surrounding area that is mutually influenced by the city area, extending up to several hundred

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

457

km in the horizontal, (ii) The city scale, over which the urban area varies, i.e. up to 50 km, (iii)
The neighbourhood scale, which bridges the range of scales between street and city scales,
i.e. from 0.2 to 10 km in the horizontal, and (iv) the street scale, up to 200 m in the horizontal.
Although these spatial scales can be unambiguously defined from the physical/geographical
identification point of view, they are somewhat arbitrarily specified in terms of fluid dynamics
since the scale demarcation does not derive from the dynamics but from the geographical
identification perspective; this point becomes particularly important when model nesting and
chemical reactions (and their corresponding time scales) need to be taken into account in
association with the corresponding spatial and therefore time scale of mixing of pollutants
to calculate the pollutant concentration field (Neophytou et al. 2004, 2005).
Currently atmospheric numerical models addressing airflow and pollution dispersion are
tailored to see at a particular scale of observation where the urban scale rather lies at one of
the extremes of the range of scales present in the problem: large-scale weather models address
regional- or global-scale variations, in which urban areas (and their associated effects) are
represented as a parametrized effect, often occurring at a subgrid scale of the model; on the
other hand, engineers most often addressing local-scale models (often of the building/street
scale or neighbourhood scale) acknowledge the need to use appropriate boundary conditions
to reflect the city or the surrounding urban area that the buildings or neighbourhood comprize,
which in turn derive from a more macroscopic characterization of the city (e.g. Neophytou et
al. 2011; Panagiotou et al. 2013). Some examples where mesoscale numerical models have
thereby been adapted using numerous approaches to incorporate urban effects into numerical
weather prediction simulations are Brown (2000), Solazzo et al. (2010), Chen et al. (2011),
Salamanca et al. (2011).
In deriving macroscopic characterizations of such spatially heterogeneous urban areas,
it is important to note that the airflow in such urban areas is rarely in equilibrium with the
underlying surface. Therefore, when describing surface-layer wind profiles it is normally
necessary to calculate aggregated values of z 0 and d that account for the heterogeneity of
the surface. This could be achieved so far, via source area modelling (Schmid and Oke 1990;
Schmid 1994) or blending methods (Bou-Zeid et al. 2007). Other studies have attempted to
characterize such spatially heterogeneous geometries and roughness with an agglomerated
z 0 and d (e.g. Kastner-Klein and Rotach 2004) despite the fact that direct validation becomes
difficult to fully address (Cheng and Castro 2002). As a result of the emerging need for
a spatially-varying characterization of urban morphologies, simple uniform square grids,
neighbourhood regions, have been recently used, of horizontal resolutions ranging from
150 m to 1 km (by e.g. Bottema 1997; Ratti et al. 2002; Holland et al. 2008). Morphometric
models are then applied to the individual neighbourhoods and the deduced parametrization
schemes most frequently involve a statistical characterization of the urban building geometry,
e.g. with the derivation of an average building height, a standard deviation in building height
and the packing densities, all derived from a pre-specified area size (e.g. Oke 1988).
In a recent study by Millward-Hopkins et al. (2012) a methodology was developed using
an adaptive grid for providing a spatially-varying characterization of z 0 and d, estimated
using the morphological data within each grid cell and as a function of the local planar
packing density, p , and local average building height, H . A convergence criterion (admitted to be arbitrarily specified) was also used in order to specify small-enough normalized
changes (of either p or H ) for statistical homogeneity and therefore for estimate of relevant
aerodynamic parameters. It was accepted that, if more stringent values than those specified
for the convergence criterion in the study were used, the methodology failed to detect the
citys boundaries, while if higher, more lenient, values for the convergence criterion were
used, the results appeared to be unsatisfactory resulting into excessively small sub-areas.

123

458

P. Mouzourides et al.

Finally, Millward-Hopkins et al. (2012) conclude with the need for a more formal mathematical process to determine the most appropriate values for the convergence criterion and
the specification of corresponding aerodynamic parameters.
There are two important implications arising from the existence of various scale-referenced
models and the need for derivation of relevant parametrizations from various urban (or other)
datasets of given pre-specified size: (a) the use of appropriate representation of urban (or other)
datasets in the corresponding scale or resolution of a numerical model requires a coherent
framework to associate the appropriate sets of data to be represented both in the model and
subgrid parametrizations, and (b) a spatially-varying characterization of any parametrization
e.g. the roughness length or zero-plane displacement should take into account the spatial
heterogeneity in a structured and consistent way for any agglomeration of data.
Using the statistics of an urban morphology in the conventional way that is currently
used is useful, as illustrated in the range of applications reviewed. However, such use of the
statistics is unable to provide a distinctive characterization of a city or urban area; for example
two urban sub-areas may yield the same statistics quantitatively, while being qualitatively
different in morphology. This makes the need for a formal framework for a scale-adaptive
and spatially-varying characterization of the urban boundary-layer parametrizations more
pressing, as well as the need for multi-scale and multi-resolution models (with nested subgrid
parametrizations) to overlap with each other in a consistent and coherent way, more important.
In this paper, the multi-resolution analysis (MRA) is proposed as a formal framework to
represent urban morphologies consistently across different scales and resolutions of models
and datasets for the derivation of associated boundary-layer parametrizations; this is mainly
due to the capacity of MRA for multi-scale sampling. By being able to represent the urban
information (at each selected level-resolution) as an approximation and a detail, it can provide
appropriate and consistent representations of urban data across different scales and resolutions
of models. Due to this capacity, MRA has been attractive for applications in several fields, e.g.
in soil science for surface roughness characterization (Josso et al. 2001; Vannier et al. 2006),
in the detection of street features in urban fabric (Couloigner and Ranchin 2000; Benediktsson
et al. 2003; Horner 2007) as well as in environmental fluid dynamics (Torrence and Compo
1998) and turbulence characterization through coherent structure identification (Farge and
Rabreau 1988; Farge 1992; Hudgins et al. 1993; Domingues et al. 2005). Recently it has
been reported that the MRA provides the ability to retain and quantify the unique character
of each and every city, and as a function of grid size. (Ching 2012; Mouzourides et al. 2012).
Below we show how MRA can provide a formal scale-adaptive spatially-varying characterization of an urban morphology and thereby used to deduce associated aerodynamic (or
other) parameters. Moreover, through an illustrated example of a real city, the central business
district area of Oklahoma City, we reveal how MRA can provide a unique characterization
i.e., identification of a city and therefore exploit further the capacities of MRA in atmospheric
modelling.
The structure of the paper is as follows: the theoretical background of MRA as a formal framework for multi-scale or multi-resolution representation (sampling) is presented in
Sect. 2. In Sect. 3, the methodology is described both with respect to the urban data used to
illustrate the MRA application as well as the specific test cases constructed and investigated
in order to illustrate the advantages for multi-scale urban wind modelling. The results are
presented and discussed in Sect. 4, especially with regard to the unique identification capacity of an urban area or city as well as to the capacity to deduce the impact or influence of a
particular building in a pre-specified area. Finally, Sect. 5 concludes with the overall findings
and suggestions for future work. Details of the mathematics underpinning the MRA theory
are also included in the Appendices.

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

459

2 Multi-Resolution Analysis (MRA): Theory with a View Toward Multi-Scale

Sampling
The choice of MRA as a tool to analyze the urban spatial heterogeneity was motivated by
its capacity for multi-scale sampling and its ability of not losing the details of smaller scales
when used to sample signals over larger scales. In the following sub-sections we briefly
review the sampling process (Bracewell 2000) in order to develop and explain the concept
of a structured multi-scalesampling as a novel interpretation of MRA, which in turn can
prove a powerful tool in atmospheric modelling. Even though the presentation of MRA that
follows may be considered more mathematical than may be conventionally expected, it is
considered to be more beneficial since it addresses: (a) the formal background of the sampling
process itself through which an initial intuitive notion of scale can be obtained; this avoids
the possible source of confusion, that of misinterpreting frequency as scale, (b) how this
intuitive notion of scale becomes formal in the MRA context, and last, but not least, (c) the
facilitation of the interpretation of the results, enabling new connections and potentials to be
explored in applications for atmospheric modelling.
2.1 The MRA Framework as a Structured Multi-Scale Sampling Process
2.1.1 Sampling and Multi-Scale Samples
The concept and the underpinning mathematics of a sampling process for a 1-D signal (such
as, for example, a time series of pollutant concentration data in atmospheric modelling) is
explained in this sub-section so as to develop then the idea for a 2-D signal such as an urban
building database. Continuous 1-D signals, be they in time or space, f (t) or f (x), are often
sampled, with a characteristic sampling frequency at discrete instants of time, ti , or discrete
points of space, xi . Similarly, digital images (considered as 2-D signals) are composed of
colour values over rectangular pixels of size dx 1 dx2 . In the following discussion the
sampling process is assumed to occur with uniform sampling period, . The series of instantaneous samples of a continuous signal e.g. f (t) at every sec gives the discrete function (or
discrete signal), f s . This discrete signal f s can be expressed in terms of the corresponding
continuous signal, f (t), using the sampling property of the Dirac (impulse) function (t) as

(t a) f (t)dt = f (a) =  f (t), (t a) ,

(2)

thus expressing the instantaneous value of f (t) at a as a convolution of the translated impulse
function (t a) and f (t)Eq. 2. Hence the discrete function f s can be expressed as the
superposition of such convolutions produced by appropriately translated impulse functions
(t n ) as

 f, (t n ) (t n ),
fs =
(3)
n

where the coefficients,  f, (t n ), are given by the inner product,


 f, (t n ) =

(t n ) f (t)dt = f (n ) ,

(4)

and f (n ) are the values (samples) of f (t) at each sampling instance.

123

460

P. Mouzourides et al.

It should also be noted at this point that Eq. 4, i.e. the instantaneously sampled values
of a signal f (t) (as convolutions of the impulse function) can be seen as the limit of the
convolution of a pulse function of a finite support T , as the support tends to zero, expressed
below as

 f, (t n ) = lim

T0

1
P
T

t n
T


f (t)dt = f (n ),

(5)

where P(t) is a square pulse of unit length and height. It is this insight that enables the
interpretation of MRA as a multi-scale sampling process. Samples at different scales are
produced as averages over the compact support (time duration) of the square pulse, i.e.
samples derived over a finite time window yielding one approximated value of the function
f (t) over this time window. The value of the integral in the argument of the limit,


1
P
T

t n
T


f (t)dt = f s (n ) ,

(6)

is interpreted to be an approximation to the actual value, f (n ), obtained by averaging of

the initial f s over a larger time scale T (than the initial sampling period ); this is defined
here as the scaled-sample. Since such a process can be recursively applied over different time
window sizes, multi-scale samples can be formed. In the following sub-section, it will be
shown how MRA can be interpreted as a structured multi-scale sampling process.
2.1.2 The MRA Framework
MRA introduced by Meyer (1992) and Mallat (1999), formalizes, through a set of axioms,
the notion of scale in much the same way as Euclidian axioms formalize the intuitive notion
of space as well as axiomatic set theory formalizing the notion of set. Specifically, the
axioms of MRA provide the mathematical notions that stimulate the development of scaling
functions, which can play the role of the pulse-like function explained earlier, and that hold
the information at large-scales, also called the approximation (corresponding to the scaledsamples) and the wavelets functions, that hold the details at smaller scales, removed during
larger scale sampling; furthermore, under certain circumstances the scaling functions and
the wavelets can be orthogonal, yielding a particularly attractive property, that the process of
deconstruction into approximation and details is simply achieved by taking inner products
and implementing the appropriate superpositions. In addition the original signal is recovered
by simple addition of the details on the appropriate approximation.
Specifically, MRA axioms postulate the existence of nested subspaces, V j , such that
. . . V j V j+1 V j+2 . . ., where j denotes an integer, while the original signal f can be
projected on to any of the subspaces, V j , as f j . In addition they postulate the existence of
a scaling function that can be used to generate these spaces. Both f j and V j belong to the
Hilbert function space of finite energy signals, denoted as L 2 (R) where R denotes the set of
real numbers.
Moreover, it is important to point out that the space V j is considered to contain information
up to resolution of 2 j . This means that V j can be deconstructed into an approximation
subspace V j+1 that contains information up to a resolution of 2 j+1 . Furthermore the detail
subspace W j+1 contains the details removed from resolution 2 j in order to obtain the
next level approximation with a resolution of 2 j+1 expressed as W j+1 is the orthogonal

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

461

complement of V j+1 in V j expressed as

V j = V j+1 W j+1 .

(7)

The most important mathematical consequence of MRA axioms is the existence of a

function (t), the so-called scaling function, of compact support t (bounded interval over
which (t) is non-zero) whose translates (n = (t n)) generate an orthonormal basis of
Vo (the space on which the original/initial f s is projected). By mere application of the axiom
concerning the dilations,
 j it can be shown that the translates of the dilated scaling function,
n
j,n = 1 j t2
constitute an orthonormal basis of V j .
2j
2

Therefore, a signal f L 2 (R) can be projected onto an f j in V j by,




f, j,n j,n (t),
fj =

(8)

as the approximation
where .., .. denotes the inner product of L 2 (R). Here, f
j is interpreted

of f at scale 2 j t, which is constructed from its samples f, j,n taken over a neighbourhood
at scale 2 j t. The details removed from lower scales are replaced by the functions j,n (t),
where their smoothness features are transferred to the reconstructed

 approximation f j . The
mathematical properties of (t) given in Appendix 2 allow f, j,n to afford interpretation
as multi-scale samples.
In view of Eq. 7 f j can be decomposed into
f j = f j+1 + d j+1 ,

(9)

where f j is the approximation of the original f s at scale 2 j , while f j+1 is correspondingly

an approximation at scale 2 j+1 and d j+1 is the detail that is removed form f j in order to
yield the coarser-scale approximation f j+1 .
Moreover, it can be deduced that, given an integer k > j and by recursively substituting
f  where j + 1 <  k in Eq. 9, f j can be written as,
f j = fk +

k


d ,

(10)

= j+1

which is the sum of its approximation at scale 2k with the sum of all its details lost from the
intermediate scales between 2 j and 2k . A more complete description of MRA that includes
the mathematical properties of the scaling function and the wavelets generated from MRA
is given in Appendix 1.
2.2 The MRA Framework Applied to an Urban Morphology as a Multi-Scale Sampling of
a 2-D Signal
An image such as the digitized image depicting the building or floor elevation height of
Oklahoma City in Fig. 1 can be considered as a 2-D signal in space, which can be expressed
as a function f (x1 , x2 ), where x1 and x2 are coordinates in the horizontal plane, to yield
the corresponding building height. MRA can thereby
 be defined on such a 2-D signal that
belongs to the space of 2-D signals f (x1 , x2 ), L 2 R 2 , by taking the tensor product (Mallat
1989),
V j2 = V j V j ,

(11)

of the approximation spaces V j of the 1-D MRA.

123

462

P. Mouzourides et al.

Fig. 1 The Oklahoma City (OKC) Central Business District (CBD) depicted as a Google image (a), and as
a digitized building elevation image (b)

As a consequence of the tensor product, a scaling function, which belongs to the original
space V0 , (x1 , x2 ) V02 , and is associated with the 2-D MRA, is obtained by the product
of the scaling function (x) with itself,
2 (x1 , x2 ) = (x1 )(x2 ).

(12)

Therefore, the MRA for the 2-D finite energy signals can similarly be defined by a sequence
2
2 . . .. The basis of each V 2 is given by the translated
V j+2
of nested subspaces . . . V j2 V j+1
j
2
and dilated versions j,(n 1 ,n 2 ) = j,n 1 (x1 ) j,n 2 (x2 ) j, n 1 , n 2 Z of the scaling function
2
2 (x1 , x2 ). The basis of the orthogonal complements, W 2j+1 , of V j+1
in V j2 constitutes the
2-D wavelets. The mathematical derivation of the bases and therefore the corresponding
scaling and wavelet functions is provided in Appendix 2.
So, for example, an image, as a 2-D signal, f (x 1 , x2 ), can be projected on an f j (x1 , x2 )
in V j2 = V j V j by,
f j (x1 , x2 ) =


n 1 ,n 2 =


f (x1 , x2 ), 2j,(n 1 ,n 2 ) 2j,(n 1 ,n 2 ) .

(13)

Similar to the 1-D case, f j (x1 , x2 ) is interpreted as the approximation of f at scale 2 j dx1
2
2 , which is reconstructed from its samples  f (x 1 , x 2 ), j,(n 1 ,n 2 )  taken at the same scale.
In addition, similar to 1-D case, approximations at higher scales, (lower resolutions) are
obtained by removing details along the horizontal, vertical and diagonal directions of the
image. The corresponding mathematical expressions (of the inner products) for the vertical,
horizontal and diagonal directions are provided in Appendix 2.
Figure 1b depicts a digitized building elevation image of Oklahoma City whose Google
image is shown in Fig. 1a. The image is represented at an original resolution of 1195 1185
pixels, as in Fig. 1b, where each pixel corresponds to 1 m2 ; the colour of each pixel maps the
building height. This information represents an original signal f (x 1 , x2 ) originally sampled
at a resolution dx1 dx2 of 1 1 m2 . Applying the MRA on such a 2-D signal, it can yield
different levels of approximation (corresponding to resolutions of 2 2 m2 , 4 4 m2 , etc.)
2 j dx

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

463

with their associated details at each level. Further details on the urban image that is analyzed
are provided in Sect. 3.

3 Methodology
3.1 Urban Building Morphology as Data for Multi-Resolution Analysis
This section discusses the methodology followed in the application of MRA on an urban
building database in order to obtain multi-scale representations of such an urban information.
The illustrated example is taken from Oklahoma City (OKC) and a detailed building database
and associated statistics of the broader OKC can be found in Burian et al. (2003). The
investigated area specifically derives from the central business district (CBD) and has an
approximate size of 1200 1200 m2 . This urban area is depicted in Fig. 1 at a resolution
of 1195 1185 pixels, where each pixel corresponds to a 1 1 m2 area and the colour
of each pixel maps the corresponding building height. The accuracy in the mapping of the
actual geometries is within 3 m for the vertical direction and within 5 m for the horizontal
directions. This urban database information as represented in Fig. 1 will be treated as our
2-D signal to be analyzed using MRA.
3.1.1 The OKC Urban Dataset
A number of commonly used statistical measures can be derived in order to characterize urban
morphologies and to model their effects on the wind field, such as the mean building height
H , the plan and frontal area densities, p and f respectively, the distribution of heights, their
standard deviation , based on which aerodynamic parametrizations can thereby be derived.
For the dataset (size and resolution) used in this analysis, a number of statistical parameters
can be computed, such as the mean building height H , the average building height weighted
over its planar area H AW weighted over its planar area Ap , the planar packing density p and
the standard deviation in building height, from
m
Hi
(14)
H = i=1 ,
mm
i=1 Hi Ap,i
H AW = 
,
(15)
m
i=1 Ap,i
m
Ap,i
,
(16)
p = i=1
AT


m 
2
i=1 Hi H
=
,
(17)
m1
where m is the total number of buildings in the urban dataset and AT is the total planar area
of the domain.
The mean building height (i.e. without considering any in-between empty space) is H =
25.17 m, the building-area-weighted average building height, H AW = 22.77 m, the plan area
packing density p = 0.221 and the standard deviation in the building heights = 27.35 m
(all listed in Table 1). The dataset contains a number of relatively tall buildings, such as the
Bank One Building (with a mapped height at 150 m), the FNC Building (123 m) and the
Oklahoma Tower (117 m).

123

464

P. Mouzourides et al.

3.1.2 Deriving Aerodynamic Parameters from Urban Morphology Data

A number of models have been developed that attempt to relate the aerodynamic roughness
length (normalized over the mean building height), z 0 / H , and the zero-plane displacement
(normalized over the mean building height), d/ H , to the simplified geometrical parameters
of the underlying surface; the plan and frontal area densities, p and f respectively (e.g.
Raupach 1992, 1994, 1995; MacDonald et al. 1998; Jia et al. 1998; Kastner-Klein and Rotach
2004; Shao and Yang 2005). An important issue is with respect to model validation and
specifically with the difficulty of accurately estimating z 0 and d from measurements, and
the associated inconsistencies between experiments and the methodologies used to estimate
these parameters. Consequently, there is large scatter in reported values of z 0 and d, even
from wind-tunnel experiments over identical uniform arrays (Grimmond and Oke 1999). As
a consequence, a comprehensive review of many of these models concluded that, in real
urban areas, estimates of both z 0 and d parameters can be highly uncertain (Grimmond and
Oke 1999). Despite this uncertainty there have been a few methods based on morphometric
concepts proposed to determine both z 0 and d. The methods are divided into three sets: (i)
the simplest and most frequently used set is the height-based rule-of-thumb, that z 0 and d
are directly proportional to the mean building height, (ii) the second set utilizes the fraction
of the plan surface area covered by roughness elements p , and (iii) the third set uses the
frontal area index of the elements as a function of wind direction, f . Here, we use as an
example the morphometric relations proposed by Kastner-Klein and Rotach (2004), which
were derived based on wind-tunnel measurements over a scaled model of a real city area in
order to deduce z0 and d over the p packing density parameter, as
z0
= 0.072p [exp{2.2(p 1)} 1],
H
d
= 0.4p exp{2.2(p 1)} + 0.6p .
H

(18)
(19)

The methodology used is not sensitive to the exact morphometric model used for the
derivation of the aerodynamic parameters. Herein it is illustrated how MRA-deduced results
for the building/floor height and planar packing density can be used together with models
such as that of Kastner-Klein and Rotach (2004) in order to derive coherent, spatially-varying
descriptions of z 0 and d. The discrepancies noted in the recent literature (e.g. by MillwardHopkins et al. 2012), as arising from the planar packing densities estimated over different
domain area sizes and the associated parametrization of z 0 and d are attributed to the weakness
to relate and reference morphology parameters at different scales, that the MRA methodology
formally addresses and therefore overcomes. This will be discussed further in the light of the
results in Sect. 4.
3.2 MRA Methodology Description
The selection of the proper scaling and wavelet functions in the MRA is dictated by the nature
of the application. For urban datasets the square shape of the Haar function is proposed, since
it enables the demarcation of the characteristics of an urban area domain, where buildings have
similar square- or rectangular-like shapes. An additional advantage of the Haar function is that
its operation corresponds to simple averaging and differencing of 2 2 cell neighbourhoods
(Nievergelt 1999). Since this is the first time that MRA is applied in the field of atmospheric
modelling and particularly in relation to urban effects, we also illustrate the selection rationale
of the analyzing function, by means of a comparison between the Haar and Symlet 4 scaling

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

465

Fig. 2 The Haar and Symlet 4 scaling and wavelet functions illustrated in 1-D (top row) and 2-D (bottom
row) forms

and wavelet functions. Figure 2 shows the 1-D and 2-D Haar and Symlet 4 scaling and
wavelet functions. For instance, at the first level of decomposition, the Haar wavelet transform
performs sequential averaging and differencing of 2 2 adjacent pixels; conceptually, this
process can be viewed as a structural averaging process in which the height difference between
adjacent buildings in the horizontal (x1 ), vertical (x2 ) and diagonal directions is tracked. Due
to the shape of the Symlet 4 function, it would be used for example in detecting or capturing
phenomena dominated by peaks or sudden occurrences. Thus for the MRA analysis of a
building database information it is recommended that the Haar analyzing function is used.
The MRA analysis was implemented in Matlab using its wavelet toolbox and it is important
to note that in MRA analysis, the number of computations required to analyze a signal
composed of N samples is proportional to N .
The above mentioned scaling and wavelet functions were applied to the Oklahoma City
building database and the results are depicted in Fig. 3. Both sets of functions identify the
sub-area where the citys tallest buildings are located; however, the two scaling and wavelet
functions give completely different qualitative and quantitative results with regard to the rest
of the building information analysis. The Symlet 4 scaling function, as opposed to the Haar
function, is not constant over its support. This means that the approximation values at the
scale 28 m for the Symlet 4 analysis are not the same over its support and hence the 28 28
cells are not demarcated as in the case of the Haar scaling function.

123

466

P. Mouzourides et al.

Fig. 3 Example of MRA results using the Haar (left) and Symlet 4 (right) functions to the OKC-CBD area

3.3 Test Cases

A number of test cases have been constructed and investigated in order to illustrate the
capacity and potential of the MRA application in the context of multi-scale numerical weather
modelling: (i) in the first, reference case, the MRA of the actual OKC-CBD area is performed;
this case is also used as the benchmark case for the comparisons with the subsequent test cases;
(ii) a second test case is constructed in order to manifest the influence of different distributions
of building height in a specified area with the same statistical urban morphology parameters
as those in the benchmark case; and (iii) a third test case is performed in order to illustrate
the extent of the influence of a particular building in pre-specified areas.
The specific settings in the test cases are summarized below.
Test case 1 (Benchmark case) This case analyzes the actual OKC-CBD area as a digitized
signal. This case is considered the reference benchmark case.
Test case 2 (Sibling city) This case investigates a very similar, hypothetically reconstructed
OKC-CBD area, where the distribution of heights of the buildings is now changed while the
original built plan area of the actual OKC-CBD area is retained. The new resulting values
of the statistical parameters i.e. the average building height of the area, the area-weighted
average building height and the building-height standard deviation, are kept the same (within
1 % accuracy) to those of the benchmark case (Test case 1). Specifically, the height of 28
buildings in the OKC-CBD area was changed, corresponding to the 9.7 % of the overall 290
buildings. Figure 7 depicts the modified OKC-CBD urban area analyzed in Test case 2. The
aim behind this investigation is to reveal the MRA capacity of identifying or distinguishing
apparently similar urban cases (e.g. with the same building-height averages and buildingheight standard deviations).
Test case 3 In this case the range of impact (or influence) of a building in pre-specified surrounding neighbourhoods of different sizes or scales is investigated. Specifically, this is done
by constructing a modified urban area in which the particular building under investigation
is removed from the original-reference OKC-CBD urban database. Here we remove one of
the tallest buildings. The building whose impact is examined in Test case 3 (and therefore
removed from the original database) is the Bank One Building with a mapped height of 150

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

467

Table 1 Basic urban building morphology statistics for Oklahoma City (OKC) and test cases
Test case 1
(Reference)
Real OKC

Test case 2

Test case 3

Average building height H (m)

25.17

25.17

24.65

Standard deviation (m)

27.35

27.16

26.38

Area-weighted average building height H AW (m)

22.77

22.14

22.27

Planar packing density p

0.221

0.221

0.220

Total number of volumes

290

290

289

Number of changed buildings

28

% Change of buildings

9.7

0.3

m and a 2,250 m2 plan area; as a consequence of this removal, the resulting average building
height is 24.7 m corresponding to a 2.1 % change of the original-reference and the resulting
standard deviation is 26.4 m corresponding to 3.5 %. All test cases performed are summarized
in Table 1.

4 Results and Discussion

In this section, the MRA results obtained in the various test cases are presented and discussed. Before proceeding, however, we demonstrate first how some of the theory and the
methodology ideas are interpreted. Figure 4 demonstrates four out of the 10 reconstructed
approximations and details obtained from the MRA of the OKC-CBD urban 2-D signal, with
the single plot at the top of the figure being the digitized image that is analyzed. The plots
in the two following rows depict the approximation and detail reconstructions of the urban
signal at levels 1 and 2 respectively. There are 10 levels overall resulting from MRA since the
size and resolution of the original 2-D urban signal allows up to 210 210 (1,024 1,024) m2
to be covered/completed. The number denoting each level, corresponds to the exponents e.g.
1, 2 of the scales (as area sizes) 21 21 and 22 22 m2 respectively, that are used to label the
scales. The fourth row (level 8) of Fig. 4 illustrates schematically the outcome of the same
process at the scale 28 28 m2 .
More specifically, the reconstructed approximation at level 1 in the first row of Fig. 4
is obtained by removing the reconstructed details shown in the three remaining plots of
the first row entitled Horizontal Detail 1, HD1, Vertical Detail 1, VD1, and Diagonal
Detail 1, DD1, from the original, digitized image at the top of the figure (as per Eq. 13 in
Sect. 2). The plot denoted as HD1 depicts the reconstructed detail at level 1 that simulates
changes in the vertical direction within the scale 21 21 m2 . It is the function d V1 W1
obtained using Eq. 36 in Appendix 2. The inner products in Eq. 36 measure the variation
within the scale 21 21 m2 in the vertical direction, and thereby where these variations are
high it appears as high values in the plot HD1. Changes occurring along the horizontals and
diagonals of the image at 21 21 m2 are depicted in the plots denoted as VD1 and DD1
respectively. What exactly happens in the plots at level 1 becomes clearer if the encircled
area, for example appearing in the web-link contained in Fig. 4 (named as web-link Fig.
12) is zoomed and the corresponding results observed. In Approximation 1, A1, the edges
and corners of the buildings are smoothed, simply because they are adjacent to non-built,
empty space. The removed details that cause the edges and corners to smooth out during the

123

468

P. Mouzourides et al.

Fig. 4 Illustrations of the reconstructed approximations and details of the MRA of the digitized image of the
OKC-CBD at different MRA levels using the Haar scaling and wavelet function. The last row illustrates the
plots of the approximation and details obtained at level 8 from the approximation at level 7 of the OKC-CBD
area. The encircled part in red of the digitized image at the top of the Figure contains a weblink that shows
zoomed plots of the reconstructed approximation and details of the MRA level 1 (i.e. at scale 21 21 m2 ) of
the digitized OKC-CBD image

approximation process are reconstructed in the detail part as HD1, VD1, DD1 in the same row
(level 1).
4.1 The MRA Results for Unique Characterization of an Urban Database and Deductions
for Spatially-Varying Aerodynamic Parameters
4.1.1 The Actual OKC-CBD Area (Test Case 1Benchmark)
The results obtained from the application of the MRA to the original OKC-CBD urban
dataset (Test case 1benchmark) using the Haar analyzing function are presented in Fig. 5;
specifically, the resulting approximation and details that the MRA yields at 10 successive
levels of the analysis are shown for this urban dataset. Approximation at level 1 is obtained by
an averaging process over cells of size of 2 2 pixels (of the original resolution); similarly
approximation at level 2 is obtained by an averaging process over cells of size of 2 2 cells
of the previous resolution (i.e. of 4 4 pixels of the original resolution) and similarly for all
the subsequent levels, with approximation at the tenth level being the result of the averaging

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

469

Fig. 5 The complete MRA results of the OKC-CBD area depicting the resulting 10 levels of analysis, with
an approximation and three types of details (horizontal, vertical and diagonal) at each level (Test case 1:
Referencebenchmark case)

process over a cell of 210 210 (1,024 1,024) pixels of the original resolution. Because
of the use of the Haar analyzing function and its particular characteristics (as discussed in
Sect. 3.2), the value of each of the resulting cells in the yielded approximation at each level
of the analysis corresponds to the weighted average of the height over the entire cell area.
As is also explained earlier in Sect. 3.3, the details H, V, D at the same level of analysis
are directional averaging and differencing in the corresponding directions. This is illustrated
more clearly in the zoomed result of Fig. 4 that depicts the approximation and details obtained
at level 8 from the approximation at level 7: Fig. 4i depicts the approximation of OKC at
level 8 obtained from the MRA of the approximation at level 7, while the remaining three
plots (Fig. 4iiiv) show the details removed from the approximation at level 7 in order to
obtain the approximation at level 8. The size of the cells in the images of the approximations
is determined by the size of the support of the scaling function at the corresponding scale.
The values of the four cells from where the arrow in Fig. 4ii originates are firstly averaged
across the rows, and subsequently, the difference of the values resulting from the averaging
process and appearing in the same column are computed to give the results showed in the
highlighted cells in the Horizontal Detail image. This reflects the Haar analyzing function
part depicted under the title wavelet function (1) in Fig. 2a, where over its first half support
takes the value 1 and over its first other half support the 1 value. As a consequence, in the
plot of Fig. 4ii entitled Horizontal Detail 8, adjacent cells, above or beneath each other, have
opposite values. This indicates the changes at the scale 27 dx1 27 dx2 along the vertical
direction.
A similar operation of that discussed above, but performed in the other direction, namely
row differencing first and then column averaging, gives the plot entitled Vertical Detail 8. It
is evident that the outcome is the average of the differences and again, due to the support
shape of the Haar analyzing function, horizontal neighbouring cells have opposite values.

123

470

P. Mouzourides et al.

In Fig. 4iii such values are denoted by the corresponding colours, which appear together in
horizontal neighbouring cells. This indicates the changes of the 27 dx1 27 dx2 scale along
the horizontal. Finally, in a similar way, Fig. 4iv represents the actual computations that
generate the Diagonal Detail 8 plot at level 8 as also indicated by Eq. 34. Similarly, as it can
be seen from the plots the values in the cells situated in diagonal positions in every square
neighbourhood of four cells are the same. This is obvious in Fig. 4 where such cells are
painted by the same colour.
At low levels of the MRA (e.g. level 2 or 3) where the size of the averaging cell is
still relatively small compared to the urban database information resolution, the obtained
approximations as seen in Fig. 5, may not yield discernible differences. However as the level
of analysis increases, and the cell size of the averaging (and differencing) process increases,
the differences in the distribution of the averaged urban height become more evident. For
example, the cell at the OKC-CBD centre, i.e. at the point (600, 600) in (x 1 , x2 ) coordinates,
has a value at the Approximation level 8 of 19.61 m, which is the sum of the values of the
cells containing the point in the approximation at level 9 (i.e. 10.37 m), in the Horizontal, the
Vertical as well as the Diagonal Details of the same cell at level 9, i.e. 1.77, 6.17, and 1.31
m respectively. Similarly, at the level 10 for example, the yielded approximation provides an
average height of 5.94 m, which corresponds to the average area-weighted building height of
that area, taking into consideration the unbuilt-empty spaces. This stratified or structural
averaging process leads to an average of the area 1,024 1,024 m2 , which corresponds
to p H AW . In terms of physical interpretation, this value is in agreement with the average
building height of the area (22.8 m) and the planar packing density (the built fraction area,
0.22); if an area-weighted average is calculated, taking into account the fraction of the unbuilt
area, this yields 5.0 m.
A particular advantage of this stratified-structural averaging process is that the removed
details are tracked at each level of analysis, and therefore sub-cell information can also be
provided. Moreover, the removed horizontal, vertical and diagonal details provide directional
differences in the heights of the cells areas and can be possibly associated with other currently
used statistical measures in the urban morphology characterization. It is also noted that the
different values of approximations obtained at the different levels of MRA are the samples
of the urban database at the corresponding scales/grid sizes associated with model resolution
(also mentioned in Sect. 2.1.2).
From the above, MRA is illustrated to be a powerful methodology for multi-scale numerical weather prediction models; it provides gridded and scaled attributes of cities as well as
subgrid information for a hierarchy of scales of such numerical models. Moreover, the ability
to track the details at every stage of averaging, and thus scale, enables a unique identification
or rather a quantification of the uniqueness of a city. This will be further discussed in the
light of the results for the next test case presented below.
The MRA can also be used to provide a spatially-varying, consistent-across-the-scales
description both of the urban morphology (building height and planar packing density) and of
the boundary-layer aerodynamic parameters (roughness length and zero-plane displacement).
Figure 6 shows the scale-adaptive values of the packing density p at the various levels
(cell resolutions) of the OKC-CBD area obtained in a similar way as described for the
average building height. It is noted that the initial digitized image for p was obtained by
assigning a value of 1 to a pixel of 1 1 m2 that corresponds to built plan area and a value
of zero to the unbuilt plan area. It is observed that the last (tenth) level of MRA for p is
consistent with the p listed in Table 1, which was calculated as usual based on the original
domain size.

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

471

Fig. 6 The complete MRA results of the planar packing density p of the OKC-CBD area as example,
resulting into 10 levels of analysis, with an approximation and three types of details (horizontal, vertical and
diagonal) at each level

4.1.2 The Sibling OKC-CBD Case (Test case 2)

A hypothetical, sibling OKC-CBD case has been constructed such that the built plan area as
well as the average building height and standard deviation of the building heights were the
same as the real, reference OKC-CBD, but only the distribution of building heights across
the domain was differentiated; Table 1 lists the number of buildings that were modified.
Figure 7 depicts the modified database as digitized images to be analyzed as well as the
associated results obtained from the application of the MRA at all levels. In addition, a
zoomed view of the level 8 analysis for both siblings, the real and hypothetical OKC-CBD,
is presented in Fig. 8 depicting the approximation as well as the horizontal, vertical and
diagonal details for both cases; the colour bar scales in the plots were kept the same in both
cases to facilitate direct comparison. By direct inspection, it becomes obvious that MRA
yields a different multi-resolution representation of the area city with the same H , p and
H AW . At the level 10 (last) of the analysis (e.g. in Fig. 7) where the averaging cell is nearly
the entire investigated area, both approximations yield the same average (as one would
expect). However because of this directional differencing that the details provide through
their tracking, a distinction between the two cases is made possible through the horizontal,
vertical and diagonal details. In fact, the addition of all the details (obtained at each level
of analysis) would provide all the subtracted detail from the original urban database in
obtaining the (final) approximation, i.e. the overall area-weighted building-height average.
In addition, this cell-like or rather matrix-like tracking of the details can be used to assess
the impact of the removed details at different resolutions of mapping (visualization), of
the city, which in turn can be linked with topological information and therefore dynamics
of an urban problem. In contrast, conventional statistics currently used, such as the standard

123

472

P. Mouzourides et al.

Fig. 7 The complete MRA results of a slightly modified (hypothetically-reconstructed) OKC-CBD area in
which the plan area remains the same but the distribution of the heights of the buildings is modified, such that
the new, yielded average building height and the standard deviation of the building height remain the same as
the original case (hypothetical sibling OKC-CBD area/Test case 2)

Fig. 8 Direct comparison between the reference and sibling OKC-CBD areas (Test cases 1 and 2) with zoomed
plots of the approximation and details obtained at level 8 from the MRA results; the colour-scale used in the
two cases is the same to facilitate direct comparison

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

I. MRA Deduced Results for the original OKC-CBD (Test case 1)

473

I. MRA Deduced Results for the Sibling (hypothetical) OKC-CBD (Test case 2)

Building/ Floor height, H (m)

Level 8

Level 9

Planar packing density,

Level 8

Building/ Floor height, H (m)

Level 10

Level 8

Level 10

Level 8

Planar packing density,

Level 9

II. Calculated parameters based on Kastner-Klein and Rotach (2004) model

Level 9

Level 9

(a)

Level 9

Level 10

Roughness length, z0 (m) - Eq.18

Level 10

Level 8

Level 10

Level 8

Level 9

Level 10

Zero-plane displacement, d (m) - Eq.19

Zero-plane displacement, d (m) - Eq.19

Level 8

Level 10

II. Calculated parameters based on Kastner-Klein and Rotach (2004) model

Roughness length, z0 (m) - Eq.18

Level 8

Level 9

Level 9

Level 10

(b)

Fig. 9 The representation of the aerodynamic description of OKC-CBD area at a scale > 250 m. I MRA
deduced results; II Calculated parameters based on Kastner-Klein and Rotach (2004) model. a Results refer
to original/benchmark case; b Results refer to hypothetical Sibling OKC-CBD area/Test case 2

deviation of the building heights, cannot be insightful enough in relation to the dynamics
since the spatial variability collapses to a single value. Due to its capacity, MRA encodes the
unique information that an urban building database embodies. By having the approximation
of the last level and the details of all the MRA levels, we can reconstruct the original urban
building database signal.
In addition to the scale-adaptive description of the urban morphology (building height and
planar packing density), the boundary-layer aerodynamic parameters have also been derived.
As it was shown in Sect. 3, the surface roughness length z 0 can be expressed as a function of
the packing density p , or f , and the average building height H , based on various morphometric and other models. In order to determine z 0 and d, the multi-resolution approximations
of the building height and the corresponding p (of area sizes of 250 250 m2 and above)
are used in conjunction with the morphometric model by Kastner-Klein and Rotach (2004).
Using Eqs. 18 and 19, scale-adaptive values of z 0 and d were obtained at scales of 250 m
(that is a representative of a neighborhood scale) and above using corresponding values of H
and p at the corresponding levels of MRA, i.e. at cell-sizes of 256 256 m2 , 512 512 m2
and 1, 024 1, 024 m2 corresponding to levels 8, 9 and 10 of the MRA respectively. Figure 9
depicts the scale-adaptive results of z 0 and d for the cases of the actual Oklahoma City (Test
case 1) and the hypothetical sibling Oklahoma City (Test case 2). Although in both cases the
average building/floor height and packing density over the entire domain (1,310 1,280 m2 )
are the same, the MRA deduced results depicted in Fig. 9 for example, shows the spatial differentiation of the aerodynamic parameters of the two sibling cases enabled exactly because
of the structural averaging process.

123

474

P. Mouzourides et al.

Fig. 10 The depiction of the impact of the tallest building of OKC-CBD area (in different sub-domains at
different scales) as a result of the subtraction between the approximations and the details of benchmark case
(Test case 1) and the corresponding approximations and the details of Test case 3. This image contains a
weblink which depicts the direct MRA results of the Test case 3

Such capacity of representation corresponds to potential subgrid-scale information that

can be held in climate models with a coherent scale-adaptive capacity. This also resolves
a weakness that has been identified in recent literature in specifying sub-domain areas of
variable sizes in a rather arbitrary way in order to derive the corresponding aerodynamic
parametrization for boundary-layer flow (e.g. Millward-Hopkins et al. 2012).

4.2 The MRA Results: Impact Assessment of Individual Buildings in an Urban Database
Test case 3 (OKC-CBD area with removal of the tallest building): Building on the MRA
capacity for a unique encoding/representation of a city and scaled attributes of it, this Test
case 3 illustrates how the impact of a particular building can be deducedboth in the entire
domain as well as in different pre-specified sub-domainse.g. of different sizes. Test case
3 (as listed in Table 1) was constructed by removing the tallest building (The Bank One
Building) in the actual OKC-CBD area which was examined as the benchmark case (Test
case 1). The selected building which is depicted encircled in Fig. 10 has a height of 150
m and a plan area of 2,250 m2 . The values of the new average building height and the
new standard deviation are reduced by 2.0 and 3.5 % respectively. This confirms that the
final level of analysis yields approximation and details that are consistent with the variation
in the conventional average building height: the cell value in the approximation obtained
in the last analysis level is 2 % smaller than that in Fig. 5. Although these observations
confirm the expected output of the MRA and its association and consistency with conventional statistical measures, the most attractive feature from the MRA analysis in this
case is that the differences observed in the approximations yielded at different levels (or

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

475

scales) of the MRA reveal and represent the impact of the removed building in the scaled
attributes.
Figure 10 shows directly for the sake of ease and clarity of observation the difference between the MRA results obtain for the original OKC-CBD domain i.e. including the tallest building (Fig. 5) and the MRA results obtained for the same domain but
excluding the tallest building. For those interested in the direct MRA results of Test
case 3 are included in the weblink (as Fig. 13). In fact, the results in Fig. 10 show that
the removal of the tallest building becomes evident right from the first level of analysis (i.e. the denser resolution) while at higher levels of analysis (coarser resolutions), the
impact of the building is manifested differently in different adjacent cells. This capacity
enables useful deductions relevant to urban planning deriving from the quantified evaluation of the impact of individual (or complex of) buildings within a domain or sub-domain
area. This illustrates in another way how MRA can quantify the uniqueness of an urban
database.
5 Concluding Remarks
In this paper we have illustrated by using an urban building database, the capacity of
MRA for multi-scale representation (sampling) and its advantages for multi-scale and
multi-resolution mesoscale numerical modelling. By taking as an example the urban building database from Oklahoma City central business district (OKC-CBD) it was illustrated
that MRA, by decomposing the urban information to an approximation, a representation at a larger scale, and a detail, the part removed (from the original information) at
smaller scales, yields appropriately the approximations that embody the different representations (samplings) of the urban domain/information at the model scales or model
resolutions. Specifically the MRA application over the 1.2 1.2 km2 area depicted at
1 m2 /1 pixel resolution has produced 10 levels of approximations, with their associated
details at each level; the last, level 10, corresponds to a scale representation/sampling of
1,024 1,024 m2 . The numerical value of the approximation at the last level is the areaweighted built height average (including the empty space in between the built area), i.e.
Happrox = p H AW .
The different levels of representation/sampling enable the distinction between urban databases recognized as identical or similar based on conventional statistics. This distinction is
made possible both through the different levels of approximation-averaging and the details.
The spatial representation of the MRA results as a square matrix enables a representation of spatial heterogeneity of urban attributes that can be directly used by multi-scale
or multi-resolution models. Furthermore, it was shown that MRA enables useful deductions with respect to the assessment of the impact of individual (or complex of) buildings
within a domain or a sub-domain area; furthermore, it can quantify the impact at different
scales.
In addition it was illustrated how MRA-deduced results for the building/floor height and
planar packing density can be used in conjunction with suggested morphometric models in
the literature such as that by Kastner-Klein and Rotach (2004) in order to derive associated
coherent, scale-adaptive boundary-layer parametrizations of z 0 and d over different domain
area sizes/cells.
Acknowledgments The authors wish to acknowledge Dr Michael J. Brown from Los Alamos National
Laboratory for providing access to the urban building data of Oklahoma City. M. Neophytou also acknowledges Dr Jason Ching (formerly at U.S.-EPA now at NCAR) for stimulating discussions as well as the

123

476

P. Mouzourides et al.

financial support by the Cyprus Research Promotion Foundation through the research project contract
ANABAMI H/
AIO/0308/33.

Appendix 1
The Definition of the Scaling Function in the MRA Framework (Carrier of Approximations)
The axioms of MRA postulate the existence of a sequence of nested subspaces, . . . V j
V j+1 . . . j Z , where Z denotes the set of integers, of the Hilbert function space L 2 (R)
of continuous time finite energy signals, R denotes the set of real numbers. In addition the
axioms state that the translates, f (t 2 j n), by 2 j n of a signal f (t) V j , belong to V j , and

its dilation 1 f 2t belongs to V j+1 . The complete list of MRA axioms that can be found
2
in many textbooks, as for example Mallat (1999), Kaiser (2010) and Daubechies (1992),
contains another three, rather technical, axioms. It follows a description of the mathematical
properties of the scaling function (t).
Through MRA axioms, it can be shown that a scaling function of support t, has a pulselike shape and can be normalized to unit area,

(t)dt = 1.

(20)

Therefore the dilations j,n of the scaling function are supported on intervals of
width 2 j ,
and by Eq. 20 can be used in place of P, which means that the inner products f, j,n are
interpreted as samples at scale 2 j .
Since any function that has a pulse-like shape and satisfies Eq. 20 can be used to approximate the impulse function it can furthermore be used to sample a function at a higher scale,
as opposed to the actual values sampled at sharp instants by the impulse function. This means
that in the integral Eq. 5 the pulse function P(t) can be substituted by a scaled and translated
scaling function, j,n , to give
1
2j

tn
+2 j

tn


f (t) j,n (t)dt = f, j,n .

(21)

2 j

Therefore, according to the discussion in the previous paragraph the inner product Eq. 21
corresponds to samples of the function f (t) at tn at scale 2 j or to an approximation of the
value of f (tn ) at scale 2 j .


Now f j is
the function
reconstructed from samples, f, j,n of f taken at scale 2 j .

The sample f, j,n approximates the value of f at the time tn . The actual basis functions
represent the actual shape that the reconstruction takes in the neighbourhood of f (tn ). This
shape is determined by the actual shape of the scaling function as it determines how the
neighbourhood of tn is weighted.
In addition since V1 V0 , the following functional equation of the scaling function ,
called the dilation equation, can be deduced,
(t) =


h n (2t n),
2
n

123

(22)

A Scale-Adaptive Approach Using Multi-Resolution Analysis

477

where the coefficients h n are given by h n = (t), (2t n) and they can be thought of
as a sequence of low-pass filter coefficients. According to the discussion in the previous
paragraph h n are the samples of the scaling function itself at scale 21 i.e. is a measure of self
similarity of . These coefficients determine the scaling function and they are very important
because, (a) they are used to construct the wavelet function, and (b) their convolution with
the analyzing signal yields computational analysis and synthesis procedures.
The Deduction of the Wavelet Function in the MRA Framework (Carrier of Details)
Wavelets in the MRA framework are created through the orthogonal complements of
V j+1 , W j+1 in V j , in other words, the subspaces of L 2 (R) containing signals f of V j ,
that are orthogonal to signals g of V j+1 , namely,


W j+1 f V j such that  f, g = 0g Vm+1 ,

(23)

which means that V j can be expressed as the direct product of V j+1 and W j+1 , written as
V j = V j+1 W j+1 .
As in the case of the approximation spaces V j , each spaceW j has
an orthonormal basis
constituted by translated and dilated versions, j,n =

1
 j
2

t2 j n
2j

of a function Wo ,

called the mother wavelet. A signal f L 2 (R) can be projected on a function d j W j ,

dj =


f, j,n (t) j,n (t).

(24)

Therefore the detail removed

from approximation
f j1 to obtain f j at a higher scale is

retained by the inner products f, j,n . Since (t/2) belongs to W1 and W1 is the orthogonal
complement to V1 in V0 then (t) can be written as,
(t) =


2
gn (2t n),

(25)

where gn are considered as coefficients of a differencing (high-pass) filter. These coefficients

can be obtained from the averaging coefficients h n by,
gn = (1)n h 1n .

(26)

In applications, analysis is performed on a sampled signal f s assumed to belong to V0

and hence it is set to f 0 . Then f 0 is decomposed up to a level k according to Eq. 10 and
reconstructed back according to
f0 = fk +

k


d .

(27)

=1

Figure 11 depicts diagrammatically this decomposition and reconstruction; the corresponding

levels and subspaces are identified. The maximum value of k is dictated by the number of
samples of f 0 and the actual wavelet used. The functional Eqs. 22 and 25 are satisfied by
many different sets of coefficients h n and gn , meaning that there is a multitude of different
mother wavelets and scaling functions.

123

478

P. Mouzourides et al.

Fig. 11 The decomposition and reconstruction of a sampled signal f 0

Appendix 2: The 2-D Formulation of the MRA

Multi-resolution analysis can be defined on L 2 (R 2 ), the space of 2-D finite energy signals
f (x1 , x2 ), by taking the tensor product (Mallat 1989),
V j2 = V j V j ,

(28)

of the approximation spaces V j of the 1-D MRA. As a consequence of the tensor product a
scaling function, (x1 , x2 ) V02 , associated with the 2-D MRA is obtained by the product
of the scaling function (x) with itself,
2 (x1 , x2 ) = (x1 )(x2 ).

(29)

Therefore, MRA for L 2 (R 2 ) can similarly be defined by a sequence of nested subspaces

2
2 . . .. The basis of each V 2 is given by the translated and dilated
. . . V j2 V j+1
V j+2
j
2
versions j,(n 1 ,n 2 ) = j,n 1 (x1 ) j,n 2 (x2 ) j, n 1 , n 2 Z of the scaling function 2 (x1 , x2 ).
2 in V 2 constitute the 2-D wavelets.
The basis of the orthogonal complements, W 2j+1 , of V j+1
j
The basis can be constructed from the 1-D basis of V0 and W0 by noting that,
V02 = V0 V0 = V12 W12 ,

(30)

and then by substituting V0 = V1 W1 and exploiting the distributive property of over

it can be shown that
W12 = (V1 W1 ) (W1 V1 ) (W1 W1 ) ,
and hence it can be deduced that the basis of
the following functions,

(31)

are the dilated and translated versions of

V0 W0 = (x1 )(x2 ),

(32)

W0 V0 = (x1 )(x2 ),

(33)

W0 W0

which play the role of the 2-D wavelets.

123

W 2j

= (x1 )(x2 ),

(34)

A Scale-Adaptive Approach Using Multi-Resolution Analysis

V j2

479

A 2-D signal, i.e. an image, f (x1 , x2 ) L 2 (R 2 ) can be projected into an f j (x1 , x2 ) in

= V j V j by,


f j (x1 , x2 ) =

n 1 ,n 2 =


f (x1 , x2 ), 2j,(n 1 ,n 2 ) 2j,(n 1 ,n 2 ) .

(35)

Similar to the 1-D case f j (x1 , x2 ) is interpreted as the approximation of f at scale 2 j dx

j
2 dy, which is reconstructed from its samples  f (x1 , x2 ), 2j,(n 1 ,n 2 )  taken at the same scale.

In addition, similar to the 1-D case, approximations at higher scales (lower resolutions)
are obtained by removing details along the horizontal, vertical and diagonal directions of
the image. The details removed across the vertical, horizontal and diagonal directions are
V W
W j V j
given by the inner products  f (x1 , x2 ), n 1j,n 2 j (x1 , x2 ),  f (x1 , x2 ), n 1 ,n
2 (x 1 , x 2 ) and
W W

j
j
(x1 , x2 ) respectively.
 f (x1 , x2 ), n 1 ,n
2
According to Eqs. 3234 three separate projections, d V j W j , d W j V j , d W j W j in the
corresponding orthogonal spaces V j W j , W j V j , W j W j can be obtained. The projection
in V j W j is obtained by,

d V j W j (x1 , x2 ) =

 V W
V W
f (x1 , x2 ), n 1j,n 2 j (x1 , x2 ) n 1j,n 2 j (x1 , x2 )

(36)

n 1 ,n 2 =

and the projections in W j V j andW j W j are obtained by using similar equations.

References
Benediktsson JA, Pesaresi M, Amason K (2003) Classification and feature extraction for remote sensing images
from urban areas based on morphological transformations. IEEE Trans Geosci Remote Sens 41(9):1940
1949
Bottema M (1997) Urban roughness modelling in relation to pollutant dispersion. Atmos Environ 31:3059
3075
Bou-Zeid E, Parlange MB, Meneveau C (2007) On the parameterization of surface roughness at regional
scales. J Atmos Sci 64:216227
Bracewell R (2000) The Fourier transform and its applications, 3rd edn. McGraw-Hill, New York, 616 pp
Britter RE, Hanna SR (2003) Flow and dispersion in urban areas. Annu Rev Fluid Mech 35(1):469496
Brown M (2000) Urban parameterizations for mesoscale models. In: Boybeyi Z (ed) Mesoscale atmospheric
dispersion. WIT Press, Southampton, pp 193255
Burian SJ, Han WS, Brown MJ (2003) Morphological analyses using 3D building databases: Oklahoma City,
Oklahoma. LA-UR, Los Alamos National Laboratory, Los Alamos, 63 pp
Chen F, Kusaka H, Bornstein R, Ching J, Grimmond CSB, Grossman-Clarke S, Loridan T, Manning KW,
Martilli A, Miao S, Sailor D, Salamanca FP, Taha H, Tewari M, Wang X, Wyszogrodzki AA, Zhang C
(2011) The integrated WRF/urban modelling system: development, evaluation, and applications to urban
environmental problems. Int J Climatol 31(2):273288
Cheng H, Castro IP (2002) Near-wall flow development after a step change in surface roughness. BoundaryLayer Meteorol 105:411432
Ching J (2012) WUDAPT: conceptual framework for an international community urban morphology database
to support meso-urban and climate models. Newsletters IAUC. http://urban-climate.com/
Couloigner I, Ranchin T (2000) Mapping of urban areas: a multiresolution modeling approach for semiautomatic extraction of streets. Photogramm Eng Remote Sens 66(7):867874
Daubechies I (1992) Ten lectures on wavelets. CBMS-NSF regional conference series in applied mathematics,
vol 61. SIAM, Philadelphia, 343 pp
Di Sabatino S, Leo LS, Cataldo R, Ratti C, Britter RE (2010) Construction of digital elevation models for a
southern European city and a comparative morphological analysis with respect to Northern European and
North American cities. J Appl Meteorol 49(7):13771396
Domingues MO, Mendes O, da Costa AM (2005) On wavelet techniques in atmospheric sciences. Adv Space
Res 35(5):831842

123

480

P. Mouzourides et al.

Farge M (1992) Wavelet transforms and their applications to turbulence. Annu Rev Fluid Mech 24(1):395458
Farge M, Rabreau G (1988) Wavelet transform to detect and analyze coherent structures in two-dimensional
turbulent flows. CR Acad Sci Paris II 307:14791486
Grimmond CSB (1998) Aerodynamic roughness of urban areas derived from wind observations. BoundaryLayer Meteorol 89(1):124
Grimmond CSB, Oke TR (1999) Aerodynamic properties of urban areas derived from analysis of surface
form. J Appl Meteorol 38(9):12621292
Hagishima A, Tanimoto J, Nagayama K, Meno S (2009) Aerodynamic parameters of regular arrays of rectangular blocks with various geometries. Boundary-Layer Meteorol 132(2):315337
Hamlyn D, Hilderman T, Britter R (2007) A simple network approach to modelling dispersion among large
groups of obstacles. Atmos Environ 41(28):58485862
Holland DE, Berglund JA, Spruce JP, McKellip RD (2008) Derivation of effective aerodynamic surface
roughness in urban areas from airborne lidar terrain data. J Appl Meteorol 47:26142626
Horner MW (2007) A multi-scale analysis of urban form and commuting change in a small metropolitan area
(19902000). Ann Reg Sci 41(2):315332
Hudgins L, Friehe CA, Mayer ME (1993) Wavelet transforms and atmospheric turbulence. Phys Rev Lett
71(20):32793282
Jia YQ, Sill BL, Reinhold TA (1998) Effects of surface roughness element spacing on boundary-layer velocity
profile parameters. J Wind Eng Ind Aerodyn 73:215230
Josso B, Burton DR, Lalor MJ (2001) Wavelet strategy for surface roughness analysis and characterisation.
Comput Methods Appl Mech Eng 191(8):829842
Kaiser G (2010) A friendly guide to wavelets. Birkhuser, Boston, 321 pp
Kastner-Klein P, Rotach MW (2004) Mean flow and turbulence characteristics in an urban roughness sublayer.
Boundary-Layer Meteorol 111:5584
MacDonald RW, Griffiths RF, Hall DJ (1998) An improved method for the estimation of surface roughness of
obstacle arrays. Atmos Environ 32:18571864
Mallat S (1989) Multiresolution approximations and wavelet orthonormal bases of L2 (R). Trans Am Math
Soc 315(1):6987
Mallat S (1999) A wavelet tour of signal processing. Academic Press, New York, 637 pp
Martilli A, Santiago JL (2007) CFD simulation of airflow over a regular array of cubes. Part II: Analysis of
spatial average properties. Boundary-Layer Meteorol 122:635654
Meyer Y (1992) Wavelets and operators, vol 2. Cambridge University Press, Cambridge, 244 pp
Millward-Hopkins JT, Tomlin AS, Ma L, Ingham DB, Pourkashanian M (2012) Aerodynamic parameters of
a UK city derived from morphological data. Boundary-Layer Meteorol 146:447468
Mouzourides P, Kyprianou A, Neophytou MK-A (2012) Searching for the distinctive signature of a city: could
the MRA be the DNA of a city? In: Proceedings of the 8th International Conference on Urban Climate,
610 August 2012, Dublin, Ireland
Neophytou MK-A, Goussis D, Van Loon M, Mastorakos E (2004) Reduced chemical mechanisms for
atmospheric pollution using computational singular perturbation analysis. Atmos Environ 38(22):3661
3673
Neophytou MK-A, Goussis D, Mastorakos E, Britter RE (2005) The development of a scale-adaptive reactive
pollutant dispersion model. Atmos Environ 39(15):27872794
Neophytou MK-A, Gowardan A, Brown MJ (2011) An inter-comparison of three urban wind models using the
Oklahoma City Joint Urban 2003 wind field measurements. Int J Wind Eng Ind Aerodyn 99(4):357368
Nievergelt Y (1999) Wavelets made easy. Birkhauser, Boston, 297 pp
Oke TR (1988) Street design and urban canopy layer climate. Energy Build 11(1):103113
Panagiotou I, Neophytou MK-A, Hamlyn D, Britter RE (2013) City breathability as quantified by the exchange
velocity and its spatial variation in real inhomogeneous urban geometries: an example from central London
area. Sci Total Environ 442:466477
Ratti C, Di Sabatino S, Britter R, Brown M, Caton F, Burian S (2002) Analysis of 3-D urban databases
with respect to pollution dispersion for a number of European and American cities. Water Air Soil Pollut
2(5):459469
Raupach MR (1992) Drag and drag partition on rough surfaces. Boundary-Layer Meteorol 60:375395
Raupach MR (1994) Simplified expressions for vegetation roughness length and zero-plane displacement as
functions of canopy height and area index. Boundary-Layer Meteorol 71:211216
Raupach MR (1995) Corrigenda. Boundary-Layer Meteorol 76:303304
Salamanca F, Martilli A, Tewari M, Chen F (2011) A study of the urban boundary layer using different urban
parameterizations and high-resolution urban canopy parameters with WRF. J Appl Meteorol 50(5):1107
1128
Schmid HP (1994) Source areas for scalars and scalar fluxes. Boundary-Layer Meteorol 67:293318

123

A Scale-Adaptive Approach Using Multi-Resolution Analysis

481

Schmid HP, Oke TR (1990) A model to estimate the source area contributing to turbulent exchange in the
surface-layer over patchy terrain. Q J R Meteorol Soc 116:965988
Shao Y, Yang Y (2005) A scheme for drag partition over rough surfaces. Atmos Environ 39(38):73517361
Solazzo E, Di Sabatino S, Aquilina N, Dudek A, Britter R (2010) Coupling mesoscale modelling with a simple
urban model: the Lisbon case study. Boundary-Layer Meteorol 137(3):441457
Torrence C, Compo GP (1998) A practical guide to wavelet analysis. Bull Am Meteorol Soc 79(1):6178
Vannier E, Gademer A, Ciarletti V (2006) A new approach for roughness analysis of soil surfaces. In: Proceedings of 14th European signal processing conference (EUSIPCO 2006), Florence, Italy
Zaki SA, Hagishima A, Tanimoto J, Ikegaya N (2011) Aerodynamic parameters of urban building arrays with
random geometries. Boundary-Layer Meteorol 138(1):99120

123