Computers & Geosciences 72 (2014) 84–93

Contents lists available at ScienceDirect

Computers & Geosciences

journal homepage: www.elsevier.com/locate/cageo

Optimisation of global grids for high-resolution remote sensing data

Bernhard Bauer-Marschallinger n, Daniel Sabel, Wolfgang Wagner
Vienna University of Technology, Department of Geodesy and Geoinformation, Gusshausstrasse 27-29, 1040 Vienna, Austria

art ic l e i nf o a b s t r a c t

Article history: Upcoming remote sensing systems onboard satellites will generate unprecedented volumes of spatial
Received 31 December 2013 data, hence challenging processing facilities in terms of storage and processing capacities. Thus, an
Received in revised form efficient handling of remote sensing data is of vital importance, demanding a well-suited definition of
7 June 2014
spatial grids for the data's storage and manipulation. For high-resolution image data, regular grids
Accepted 2 July 2014
Available online 11 July 2014
defined by map projections have been identified as practicable, cognisant of their drawbacks due to
geometric distortions. To this end, we defined a new metric named grid oversampling factor (GOF) that
Keywords: estimates local data oversampling appearing during projection of generic satellite images to a regular
Remote sensing raster grid. Based on common map projections, we defined sets of spatial grids optimised to minimise
High resolution
data oversampling. Moreover, they ensure that data undersampling cannot occur at any location. From
Big data
the resulting GOF-values we concluded that equidistant projections are most suitable, with a global
Global grid
Projection mean oversampling of 2% when using a system of seven continental grids (introduced under the name
Sampling Equi7 Grid). Opposed to previous studies that suggested equal-area projections, we recommend the
Equi7 Grid Plate Carrée, the Equidistant Conic and the Equidistant Azimuthal projection for global, hemispherical
and continental grids, respectively.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction generated by upcoming Earth observation missions, our research

aimed for identifying an optimal definition of global grids for high
High speed internet, enhanced sensor systems and cost- resolution satellite data over land.
efficient data storage allow vast amounts of data to be generated A key feature of satellite missions for Earth monitoring is the
in various fields of science and technology. Moore's law is no long-term record of data time series, built of alone-standing
longer valid – data volumes are scaling faster than computer images. Thus, in remote sensing and in the global scope, the
resources (Bierig et al., 2013). Accordingly, in the field of earth mosaicking of individual satellite images to a common data space
observation, a new generation of spaceborne high-resolution facilitates generation and analysis of temporal products.
imaging sensors is emerging. They will produce unprecedented Finally, the sought grid system should facilitate efficient data
data volumes and consequently challenge up-taking facilities in storage as well as accurate spatial and temporal manipulation of
terms of storage and processing capacities. For instance, the the data for processing higher level products. Ideally, archiving,
upcoming European satellite missions Sentinel-1 and Sentinel-2 processing and display of the data are carried out in the same grid
will generate radar and optical raw images, respectively, of 640 TB system, so that products of different levels are easily related.
(Hornacek et al., 2012) and 580 TB (Drusch et al., 2012) per annum.
Remote sensing observations deliver information on the Earth's 1.1. Raster grids
features and conditions in digital form and thus are often repre-
sented by discrete raster systems. These so-called grids reference Spatial grids define a set of geodetic locations and form a
the individual observations to geolocations. The efficient and reference for storage, manipulation and display of spatial data. In
accurate handling of remote sensing data is dependent on the general, there are two types of grids: Irregular grids define each
grid being used. Motivated by the vast amounts of data that will be location explicitly (through functions or lookup-tables), whereas
regular grids define each location implicitly with a fixed sampling
distance relative to a set of linear axis; commonly two orthogonal
n
Corresponding author. Department of Geodesy and Geoinformation, Remote ones. A detailed overview on grid definitions is given by Sahr et al.
Sensing Research Group, Vienna University of Technology, Gusshausstrasse 27-29,
1040 Vienna, Austria.
(2003), focusing on the potential of irregular grids.
E-mail address: [email protected] The studies of Hortal and Simmons (1991), Seong (2005b), Sun
(B. Bauer-Marschallinger). et al. (2007) and Birch et al. (2007) represent a rich variety of

http://dx.doi.org/10.1016/j.cageo.2014.07.005
0098-3004/& 2014 Elsevier Ltd. All rights reserved.
 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93 85

irregular grid definitions for spatial data. Data stored in such grids Usery (2001) came to the same conclusion, using a theoretical
suffers limited shape distortion, information redundancy and data approach based on vertical and horizontal scale factors. Further
loss since the grid locations are defined in a mesh on a three- recommendation of equal-area projections and in particular the
dimensional model of the Earth. However, global datasets with Sinusoidal for constructing global raster data grids is given by
high spatial resolution comprise an immense number of discrete Seong et al. (2002), Seong (2005a) and Usery et al. (2003).
locations. The individual identification of locations in an irregular
grid can be computationally demanding, requiring specific algo-
rithms and lookup-tables.
Alternatively, regular grids can be used, bringing some advan-
tages: (1) They address locations implicitly with a small set of
parameters. An orthogonal grid – a Cartesian coordinate system
grasped as an array – is directly indexed by computers and is thus
expected to be more suitable for handling datasets with high
resolution. (2) Neighbour relationships of measurements are
evident from the indices, whereas irregular grids necessitate
(costly) computation of those. (3) Image data are favourably
manipulated and exchanged in the form of arrays. (4) Remote
sensing software functions often use arrays as input and output,
requiring data stored in irregular grids to be transformed into a
two dimensional representation. Such transformation is not
required in the case of regular grids. (5) Similarly, also the display
of data on screen or in print impose a two dimensional
representation.
However, regular grids have a major drawback since they do
not allow a three-dimensional definition of locations in a con-
venient manner. In geometry, a regular grid is considered as a
plane to which geographic locations are transformed by a map
projection. The latter is a set of mathematical functions between
the Earth's curved surface and the plane of the grid. Due to the fact
that the Gaussian curvature of a plane is different from a sphere or
ellipsoid, map projections, and therefore also regular grids, cannot
represent the Earth's surface without distorting lengths, angles or
areas (Feeman, 2002).

1.2. Regular grids – state of the art

Recognising the necessity of regular grids for the practicable

and efficient handling of large remote sensing datasets, much
research on shortcomings of map projections in respect to ras-
terised data has been carried out. It is almost consistently reported
that (a) raster data should be treated differently than vector data,
(b) a grid's accuracy and efficiency is bound to the distortions of
the underlying map projection and (c) projections featuring true
areal scale yield more favourable conditions than those with true
angular directions.
Map distortions reduce accuracy and efficiency of a regular grid
in form of pixel loss and duplication (Mulcahy, 2000; White,
2006). Those effects appear inevitably when transferring input
datasets to raster databases, which is nothing other than resam-
pling (Seong and Usery, 2001; Kimerling, 2002). In the literature,
resampling is identified consensually as a source of errors and
cause of unrecoverable loss of information and thus should be
used as little as possible. Finn et al. (2012) discussed errors caused
by spatial resampling and presented enhanced reprojection meth-
ods that pay particular attention to data at a global scale.
Steinwand et al. (1995) pioneered on this topic, introducing
pixel distortion effects to the remote sensing community. They Fig. 1. (a) Illustration of mapping process using projections; on the left situation on
performed a window-based counting of resampling errors when the globe with the input image comprising individual measurements and on the
right the projected image overlaid on a regular grid. (b)–(d) Effects resulting from
transforming between different map projections. Similarly,
the Sinusoidal's skewing and the orientation of input images. On the left input
Kimerling (2002) examined the loss and duplication of data when raster images and on the right their projected output raster representations
transforming from the conversant Plate Carrée (alias Equirectan- overlaid on a regular grid with grid sampling distance d equal to the sampling
gular projection) to equal-area projections. Mulcahy (2000) devel- distance of the input raster. Orange colours indicate data density for each output
oped two metrics called pixel loss and pixel duplication. She raster pixel. (b) Case of input image raster aligned with output grid raster. (c) Case
of input image rotated by local skewing angle. (d) Case of input image aligned with
determined those for various global equal-area map projections directions of local extreme length distortions a and b. (For interpretation of the
and identified solely the Sinusoidal projection as optimal for raster references to color in this figure caption, the reader is referred to the web version of
images, effecting zero loss and duplication of pixels. Seong and this article.)
86 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93

However, Luo et al. (2008) and Khlopenkov and Trishchenko distortion as the variation of directions, meaning that local linear
(2008) reported degraded spatial quality of MODIS data in zones scale varies with the direction and straight lines and shapes are
of high latitude and longitude, stemming from significant distor- not preserved. (c) Areal distortion as the variation of area, meaning
tions of the underlying Sinusoidal projection. The angular defor- that measures of area are dependent on the location in the
mation in these areas is serious, evident by large values of projected map. Map projections can have true linear scale only
stretching, shortening and skewing. at particular points or/and in particular directions. Furthermore,
Fig. 1 depicts the principles of the mapping process (Fig. 1a) and projections can be either area-preserving (called equivalent or
local image distortions of the Sinusoidal in central Asia around 401 equal-area) or angular-preserving (called conformal), no projec-
North and 1101 East (Fig. 1b–d). The distortions stemming from the tion is both, and some are neither. Among the latter, some
projection are visible as skewing of shapes and length scales that projections show true linear scale between one or two points
vary with the direction. If the input image raster is aligned with and every other point on the map, or along every meridian
the output raster, no information is lost or duplicated, although (Snyder, 1987). They are called equidistant and show moderate
the shape is distorted severely (Fig. 1b). However, this does not distortion of angles and area.
apply when the orientation of the input image is changed: The The Tissot's Indicatrix (Tissot, 1881) integrates the most basic
rotated input image in Fig. 1c suffers both data loss and duplica- differential metrics for map projections and is a convenient tool
tion. An image aligned with the directions of local extreme length for displaying abovementioned distortions. It is the representation
distortions (a and b) suffers even more, with a minimum sample of an infinitesimal circle on the sphere projected onto the map. At
distance reduced to local maximum shortening b and an irregular each map location i, this ellipse's form, orientation and size
pattern of pixel loss and duplication (Fig. 1d). It becomes also indicate the length, angular and areal distortion, respectively.
evident that relations of neighbouring pixels are highly disturbed To give an example, Fig. 2b shows local distortions of the
and thus appropriate application of local filters and operators is Sinusoidal at 401N and 1101E using Tissot's Indicatrix, including
impeded. To recap, an arbitrary image projected by the Sinusoidal the horizontal and vertical scale, orthogonal directions of extreme
suffers from (1) pixel loss, (2) pixel duplication and (3) disordered scales a and b as well as the maximum angular deformation ω.
pixel neighbour relationships. These impairments reduce both In conformal maps, all indicatrices are circles ðωi ¼ 0-ai ¼ bi Þ
storage efficiency and accuracy during manipulation. and in equal-area maps they all have an area of one ðai bi ¼ 1Þ. The
These effects were not discussed in antecedent studies on the vectors of extreme scale factors correspond to the ellipse's major
usability of map projections for remote sensing data. Therefore, we and minor axes, which are orthogonal on the sphere as well as on
propose a generalised evaluation of map projections by the values !
the map. The vectors of the maximum scale factor, a i , and the
of extreme length distortions, apparent as local stretching and !
shortening that are commonly changing with location and direc- minimum scale factor, b i , form the basis of our evaluation.
Formulae for their calculation can be found e.g. in Snyder (1987),
tion. This approach considers that generic images observed by
remote sensing systems can have any location and orientation. To Kuntz (1990), Maling (1992) or Bugayevskiy and Snyder (1995).
this end, we defined a new metric named grid oversampling factor
(GOF) and computed it for candidate map projections. The exam-
ination comprises analysis of map distortions and the required
sampling distance of the overlaid grid. Motivated by user require-
ments of our research field in land application of satellite remote
sensing, we focus on global land masses.

2. Grid sampling analysis

2.1. Map distortion measures

Academic approaches on map distortion analysis use methods of

differential geometry to classify and measure the transformation
properties of sphere-to-plane-projections. We also adopt this
approach since our goal is to find optimal solutions for generic
spatial remote sensing observations. This is opposed to several
studies mentioned above, which imply already projected search-
and-count windows (Steinwand et al., 1995; Kimerling, 2002;
White, 2006) or specific raster orientations (Mulcahy, 2000; Seong
and Usery, 2001; Seong et al., 2002). We use a spherical Earth model
with the assumption that the differences to distortion measures of
an elliptical model are insignificant (ellipsoid flattening
1=300).
A central concept of classic map distortion analysis is local linear Fig. 2. (a) As Fig. 1(a); On the left the sphere with geographic grid lines and the unit
scale. It is the ratio of an infinitesimal segment on the map to the circle on a map in yellow. On the right, the map with local Tissot's Indicatrix,
corresponding infinitesimal segment on the sphere in a given geographic grid lines and grid raster. (b) The relationship between unity circle and
Tissot's Indicatrix: On the left the (true) circle as on the sphere; on the right the
direction from a given point (Bugayevskiy and Snyder, 1995). When
corresponding Indicatrix for the Sinusoidal at 401N and 1101E as in Fig. 1. Directions
divided by the nominal scale (between the Earth and the map), it of extreme scale factors are shown in red and directions to which angles are most
becomes the local linear scale factor and quantifies local length distorted in blue. The projection yields a skewing angle of 501, a maximum scale
distortion as a function of location and direction. It is 1 for true factor a ¼1.76, a minimum scale factor b ¼0.57 and a maximum angle distortion
scale, less than 1 for shortening and greater than 1 for stretching. ω¼ 611. Input and output grids with sampling distance d ¼ 1 and directions of
longitude λ and latitude ϕ in grey. Remote sensing observations are depicted by
This leads to three types of map distortions: (a) Length distor- circles and corresponding ellipses in light grey. (For interpretation of the references
tion as the variation of scale, meaning that the local linear scale is to color in this figure caption, the reader is referred to the web version of this
not constant and lengths vary over the projected map. (b) Angular article.)
 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93 87

2.2. The grid oversampling factor (GOF) We justify describing changes of counts of discrete and exten-
sive pixels by infinitesimal small Tissot's Indicatrices with two
Following Nyquist's sampling criterion (Shannon, 1949), a grid's applicable assumptions: (1) The variation of a and b in the
sampling distance must be smaller than half the spatial resolution neighbourhood of a location i is assumed to be negligible in the
of the data throughout the grid in order not to lose any informa- case of a high resolution grid. (2) GOFi is understood as estimate of
tion. In remote sensing, the image resolution is commonly given local oversampling and should accurately sum up in an extensive
by twice the image sampling. The sampling distance between grid grid – meaning that although GOFi may not quantify pixel loss/
locations may therefore not be greater than half the spatial duplication correctly locally (e.g. in a 3  3-pixel-matrix), it still
resolution of the image data that should be gridded. While the gives a good estimate of overall pixel loss/duplication in a
sampling distance of a regular grids is constant in map projection grid area.
space, it varies on the sphere. Hence, the design of an optimal grid
must take length distortion measures of the map projection into
account. 2.3. Grid evaluation method
Oversampling appears when the grid spacing is smaller than the
necessary Nyquist sampling distance in the satellite data. In such a We investigated the oversampling characteristics of various
case, the gridded data contains information redundancy in the form common map projections (Table 1 and Fig. 3) by calculating their
of dependent samples, thus requiring more storage space than average grid oversampling factors ðGOF i Þ.
necessary. Undersampling appears where the grid spacing is larger A 360  180 matrix was set up that represents the Earth's
than Nyquist sampling distance and causes loss of information. The surface, subdivided into 11-tiles. The free World Borders Dataset
latter shall not be allowed and thus yields this study's major criterion. shapefiles (TM_WORLD_BORDERS-0.3) provided on http://blog.
It demands that this minimal sampling must be achieved throughout thematicmapping.org/ was rasterised to 11-pixels and used as land
the grid. As a second criterion, oversampling should be held minimal mask. All geospatial operations were done using the open source
in order to save storage space and reduce unnecessary computational Geospatial Data Abstraction Library (GDAL, version 1.9.2) released
overhead whenever the data is processed. by the Open Source Geospatial Foundation (OSGeo). For each
To be able to characterise the gridding of generic satellite candidate projection and for each 11-land-tile the GOF at the tile's
images we have to make certain assumptions: We disregard the centre point was calculated and weighted by the cosine of latitude
digitising processes (during sensor acquisition) that transform to account for the cell's area on the Earth (in order not to give
continuous physical variables into discrete raster images. An image disproportional influence of distortions at higher latitudes when
is assumed to have the same (and invariant) sampling in the comparing mean GOF-values). Finally, the mean GOF value of all
horizontal and vertical direction of the sensor's orthogonal raster candidate projections is compared and ranked. Projections that
image geometry. To account for varying satellite orbit azimuths feature lower average GOF are regarded as more efficient since
relative to a fixed grid, the image's orientation is regarded to be they produce less data redundancy when gridding arbitrary
arbitrary. Hence, in this study, a satellite image is modelled by an satellite image datasets onto it.
equidistant orthogonal lattice with an arbitrary orientation, hold- If several grids are combined to achieve global coverage, the
ing the image values on the lattice points. total data oversampling can be further reduced. In theory, over-
!
For each location i in the grid, the vector b i gives the direction sampling can be nearly completely avoided by using an arbitrary
and value of the local minimum scale factor bi. The optimal grid high number of grids, each of them optimised to its local
sampling distance is equal to the product of the Nyquist sampling neighbourhood. For example, the Universal Transverse Mercator
distance of the data dim and the local scale factor. It is therefore the coordinate system (the UTM grid, Defense Mapping Agency, 1989)
minimum scale factor bi that dictates the grid's sampling distance covers the Earth by separate conformal projections in 60 long-
dg in order to avoid undersampling anywhere in the grid. This itudinal stripes and 2 polar caps. Thanks to this setup, it achieves
consideration is done locally. So, to achieve this for the entire grid, highest geometric accuracy and is widely used by national and
the minimum value of bi across the entire grid, bmin, needs to be administrative services. In practice, such a solution would make it
determined. The projection-specific optimal grid sampling dis- necessary to re-project data to a common map projection when-
tance is thus defined as ever spatial manipulation or visualisation of data that covers more
than one grid zone is required. This would result in increased
dg ¼ dim bmin ð1Þ
computation cost and reduced data quality due to the required
To quantify grid oversampling caused by map projections, we
now introduce a new metric. In accordance with the above, it Table 1
analytically quantifies the local value of oversampling on the Candidate map projections for a global grid system, as shown in Fig. 3; with
condition so that undersampling is not allowed anywhere in the common applications and undistorted metrics. Scale ϕ is the local scale along
meridians; Scale λalong parallels; Centric Scales emanate from the projection
grid. We call it the grid oversampling factor (GOF) at location i:
centre. Plate Carrée# is also known as Equirectangular. Parameters λ0 and ϕ0 define
ai bi the projection centre. ϕ1 is latitude of true scale for azimuthal projections and
GOF i ¼ 2
: ð2Þ standard parallel for conic projections (ϕ2 the optional second standard parallel).
bmin
Fig. 3 Projection Typical zones Preserves Parameters
It is the ratio between the local area distortion ðai bi Þ and the
square of the global minimum scale factor (b2min). It is always larger a Plate Carrée# Global Scale ϕ λ0
than 1, since ai Zbi Z bmin and ai 4 bmin are valid for every i in an b Mercator Global, Regional Angles λ0
extensive grid. The GOF relates the data's inflation geometrically to c Lambert Cylindrical Global Area λ0 ϕ1
d Sinusoidal Global Area, Scale λ λ0
the imposed sampling distance and thus measures (over-) sam-
e Hammer-Aitoff Global Area –
pling. In other words, it estimates how much more pixels are f Equidistant Conic Mid-latitudes Scale ϕ λ0 ϕ0 ϕ1 ϕ2
required in the output grid than in the input grid to represent the g Lambert Conic Mid-latitudes Angles λ0 ϕ0 ϕ1 ϕ2
same information. Here, the pixels are regarded as entities of h Albers Conic Mid-latitudes Area λ0 ϕ0 ϕ1 ϕ2
individual locations in the (regular) in- and output grids. There- i Equidistant Azimuthal Poles, Continents Centric Scales λ0 ϕ0
j Stereographic Poles, Continents Angles λ0 ϕ0 ϕ1
fore, the GOF is independent from the actual resampling proce- k Lambert Azimuthal Poles, Continents Area λ0 ϕ0 ϕ1
dure since only pixel counts but not pixel values are considered.
88 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93

Fig. 3. Candidate map projections for a global grid system, as listed in Table 1. (a)–(e) (Pseudo-) cylindrical projections, (f)–(h) conic p., (i)–(k) azimuthal p., Tissot's
Indicatrices in purple. Letters (a)–(k) correspond to Table 1. Colour of landmasses indicate projections that are equidistant (blue), conformal (red) and equal-area (yellow).
(For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

additional manipulations. While such a solution may be consid-

ered optimal for archiving, it is not optimal for data processing.
Here, the efficiency of using more than one grid for global
coverage was evaluated with different combinations of hemisphe-
rical and continental grids. The same analysis as for the global
grids was carried out, followed by computation of the mean GOF of
the combined grid systems, weighted by the covered land areas.
Fig. 4 depicts the experiments' grid zones.

2.4. Validation

To validate the grid's GOF averages, we computed actual grid

pixel counts and related them to the corresponding area extent on
a spherical Earth. For each grid zone, and for each projection, we
re-projected and rasterised the World Border Database shapefiles Fig. 4. Zones to which oversampling experiments are conducted. Letters A–H
to a raster image, respecting the required grid sampling distance correspond to Table 2. Continental zones to Table 3 and Fig. 6.
 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93 89

as in Eq. (1). As image sampling distance we chose 1 km ðdim ¼ 1Þ northern hemisphere. Typical for equidistant projections, the Plate
to approximate high resolution conditions (as against 11). The Carrée features bi ¼ 1 at every location, yielding bmin ¼ 1 and
resulting test grid carries a simple land-sea-mask with a pixel further ai ¼ GOF i . The distortion pattern of the Sinusoidal is star-
spacing (and extent) of dg ¼ 1  bmin . The ratio of the land pixel shaped: Regions around the equator and the central meridian are
count to the zone's land area extent on Earth in km2 forms our undistorted and regions apart these lines are moderately distorted.
validation quantity, named here pixel area factor (pxAF): However, as persevering areas, bi equals 1=ai and thus bmin 5 1, it
will produce higher GOFi. The high degree of shortening and
land pixels
pxAF ¼ 2
ð3Þ stretching is the major drawback of equal-area projections, what
land area ½km  also implies a high degree of angular deformation (Fig. 5d and e).
In a grid designed to sample 1 km images, it measures how The pattern of the Equidistant Azimuthal with projection centre
many more pixels represent the zone's land area than a spherical over Mongolia features an almost undistorted Asian continent but
model would do and thus it naturally describes oversampling. very high distortions towards the opposite pole of the projection
The pxAF analysis was integrated in this study to confirm results of centre south-west of South America. Analogous to the Plate Carrée,
the GOF analysis and one should be aware of the much higher all bi are equal to 1 and thus ai ¼ GOF i .
computational costs spent for pxAF than for GOF results. Table 2 lists the average distortions for global and hemisphe-
rical zones for various projections, as well as minimum scale
factors, including land area extent and GOF averages.
3. Results and discussion
3.1. Global grids
Introductorily, maximum scale factors (ai) and maximum
angular deformations (ωi) of the Plate Carrée, the Sinusoidal and The first rows in Table 2 list distortion values for global zone
the Equidistant Azimuthal (the latter centred over Asia) are plotted comprising major land masses in zones A and B. The different
in Fig. 5. For the Plate Carrée, one can see the latitudinal increase projections feature remarkable difference in mean GOF
in distortion from low to high latitudes. Africa and great parts of values, with a six times higher mean for the Sinusoidal than for
South America, southern Asia and Oceania do not suffer much the Plate Carrée. The Lambert Cylindrical even exceeds the
from distortions, as opposed to Antarctica and large parts of the measurable range.

Fig. 5. Distortion values over global land masses for selected projections (Plate Carrée, Sinusoidal, Equidistant Azimuthal (centred over Asia)). On the left, maximum scale
factor a (a–c) and on the right, maximum angular deformation ω (d–e). All plots projected as Plate Carrée for easier comparison.
90 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93

Table 2 orientated images can be distorted unfavourably due to skewing

Average values ðÞ over land locations of maximum- ai and minimum scale factors caused by severe angular deformation (as showed in Fig. 1c and d).
bi, maximum angular deformations wi and grid oversampling factors GOF as well as
This finding is remarkable since it does not support recommenda-
global minimum scale factors bmin for different map projections and zones A–G.
Names of conic projections include optimal standard parallel(s); names of azi- tions of previous studies to use the Sinusoidal for raster databases.
muthal projections include centre latitude. Values indicated by c are constant On the contrary, non-equal-area projections show lower over-
throughout the map. For reasons of clarity, results displayed only with two sampling factors. In zones A and B the equidistant Plate Carrée,
decimals places. preserving neither angles nor area but featuring bi ¼ bmin ¼ 1,
Zone/Projection Area ai bmin ωi
performs best with a global mean oversampling of 36% in zone B
bi GOF i
(Mio km2) 1 1 1 1 1
ðGOF i ¼ 1:36Þ. Using parameters different to λ0 ¼ 01 or ρ1 ¼ 01 does
not reduce mean GOF, however. When compared with equal-area
A  901o ϕo 901 146.9 maps, its mean angular deformation is low. For the Mercator, the
Plate Carrée 1.81 1c 1 21.2 1.81 angular deformation is everywhere 01, but because bi is as large as
(Mercator) 1.81 1.81 1 0c 10.39
ai, the mean oversampling is higher than for the Plate Carrée.
Hammer-Aitoff 1.41 0.76 0.35 33.8 8.00c
Sinusoidal 1.50 0.73 0.29 38.3 11.73c
Lambert Cylindrical 1.81 0.74 0 38.2 1c
3.2. Hemispherical grids
B  561 o ϕ o 721 131.9
Plate Carrée 1.35 1c 1 14.6 1.36 Significant reduction of oversampling can be achieved when
Hammer-Aitoff 1.35 0.78 0.37 30.1 7.41c the idea of using only one grid for the entire globe is abandoned,
Sinusoidal 1.44 0.75 0.31 35.3 10.73c
as demonstrated by the results of using hemispheric grid layouts
Lambert Cylindrical 1.35 0.80 0.32 28.0 9.93c
for zones C–H in Table 2. Zone C's and D's limits of 721 North and
C131 o ϕ o 721 84.3 561 South frame all six non-polar continents (except a small
Lambert Conic 44.01 1.04 1.04 1 0c 1.09
Albers Conic 25.51 67.01 1.04 0.96 0.94 4.7 1.14 Table 3
Equidistant Conic 45.01 1.04 1c 1 2.7 1.04 Average values ðÞ of local grid oversampling factors GOF and angular deformations
Plate Carrée 1.52 1c 1 21.2 1.52 ωi over land locations for continents and polar zones, using suitable map projec-
tions. For azimuthal projections, optimal projection centres λ0 and ϕ0 are given; for
D  561 o ϕ o 131 47.6 conic projections, optimal standard parallel ϕ1 are given. Values indicated by c are
Lambert Conic  12.01 1.04 1.04 1 0c 1.07 constant throughout the map. For Antarctica, projections marked with n are centred
Albers Conic  47.51 1.01 1.06 0.94 0.91 6.9 1.20 at the pole and GOF-accuracy is reduced to two digits, accounting for the seasonal
Equidistant Conic  13.01 1.03 1c 1 2.0 1.04 variation of the ice cap extent.
Plate Carrée 1.06 1c 1 3.2 1.06
Continent / Projection Area/bmin λ0 ϕ0 =ϕ1 ωi GOF i
E01 o ϕ o 901 99.7
Mio km2/1 1E 1N 1 1
Lambert Azimuthal 901 1.14 0.89 0.71 14.1 1.98c
Stereographic 901 1.30 1.30 1 0c 1.75
Europe 9.7
Equidistant Azim. 901 1.19 1c 1 9.3 1.19
Lambert Azimuthal 0.967 9.5 51.0 1.2 1.069c
Stereographic 1 24.0 53.0 0c 1.029
F  901 o ϕ o 01 47.2
Equidistant Azimuthal 1 24.0 53.0 0.5 1.009
Lambert Azimuthal 901 1.19 0.85 0.71 18.6 1.98c
Equidistant Conic 1 54.0 0.8 1.014
Stereographic 901 1.42 1.42 1 0c 2.12
Equidistant Azim. 901 1.25 1c 1 12.3 1.25
Africa 33.6
Lambert Azimuthal 0.934 22.5 3.5 2.6 1.147c
G721o ϕo 901 2.7
c Stereographic 1 21.5 8.5 0c 1.094
Lambert Azimuthal 901 1.01 0.99 0.99 0.6 1.02
Equidistant Azimuthal 1 21.5 8.5 1.7 1.030
Stereographic 901 1.02 1.02 1 0c 1.03
Equidistant Conic 1 8.5 2.7 1.049
Equidistant Azim. 901 1.01 1c 1 0.6 1.01
Asia 38.6
H  901 o ϕ o  561 12.3
c Lambert Azimuthal 0.922 104.0 47.5 2.8 1.178c
Lambert Azimuthal 901 1.01 0.99 0.97 1.0 1.06
Stereographic 1 94.0 47.0 0c 1.094
Stereographic 901 1.02 1.02 1 0c 1.03
Equidistant Azimuthal 1 94.0 47.0 1.7 1.030
Equidistant Azim. 901 1.01 1c 1 0.6 1.01
Equidistant Conic 1 49.5 2.3 1.043

North America 24.2

Considering Eq. (2), one can see that in order to keep over- Lambert Azimuthal 0.922  111.0 47.5 2.5 1.176c
sampling GOFi small, a small value of maximum scale factor ai is Stereographic 1  97.5 52.0 0c 1.075
Equidistant Azimuthal 1  97.5 52.0 1.3 1.024
demanded. Furthermore, the ratio between the individual ai and
Equidistant Conic 1 57.0 2.5 1.045
bmin should be small, whereas the latter is dependent on the bi. Per
definition, conformal projections (Mercator) fulfil those require- South America 17.9
ments since ai ¼ bi for every i. In contrast, equal-area projections Lambert Azimuthal 0.956  67.0  20.5 1.7 1.095c
feature an inverse relationship of ai ¼ 1=bi . Indeed, our results for Stereographic 1  60.5  14.0 0c 1.049
Equidistant Azimuthal 1  60.5  14.0 0.9 1.016
landmasses in global (A, B) in Table 2) conclude poor performance Equidistant Conic 1  15.0 1.9 1.035
of equal-area projections such as the Hammer-Aitoff, the Sinusoi-
dal and the Lambert Cylindrical. Oceania 11.2
For equal-area projections employed in the (globally extended) Lambert Azimuthal 0.874 149.0  19.0 2.8 1.309c
Stereographic 1 131.5  19.5 0c 1.060
zones A and B, the high values of mean GOF are to a lesser extent
Equidistant Azimuthal 1 131.5  19.5 1.1 1.019
due to the difference between ai and bi, but mainly due to Equidistant Conic 1  20.5 1.4 1.026
relatively low values of bmin. In particular, the Lambert Cylindrical
performs inadmissibly in zone A because of bmin ¼ 0. The Sinusoi- Antarctica 12.3
dal shows highest mean angular deformation wi among all tested Lambert Azimuthal 0.962 28.0  76.0 1.4 1.08c
Stereographicn 1 0.0 -90.0 0c 1.03
projections. This is linked to largest deviations from unity of local Equidistant Azimuthaln 1 0.0 -90.0 0.6 1.01
extreme scales and consequently high GOFi values. Arbitrarily
 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93 91

portion of Siberia), where the partition at 131 North was found to For each continent, the three most suitable azimuthal projec-
be optimal for a setup using two conic projections. We tested the tions were identified and tested for oversampling against the
conformal Lambert Conic, the equal-area Albers Conic and the Equidistant Conic, which has performed best in the hemispheric
Equidistant Conic and compared them with the Plate Carrée cases (not for Antarctica).
projection. For both hemispheres, the Equidistant Conic excels Not unexpectedly, the projection's ranking is the same for each
the others with an oversampling of 4%. The reason lies in the continental section: The Equidistant Azimuthal performs best,
nature of equidistant projections that balance the virtues of followed by the Equidistant Conic, the Stereographic and the
conformity and area truth, though not reach them. In spite of Lambert Azimuthal. Concerning oversampling, the conic projec-
some limited angular deformation, the Equidistant Conic retains tion reaches performance of the azimuthal ones only over Asia
bmin ¼ 1 and oversample less than the angle-preserving Lambert and Oceania, where the area is spread in the east-west direction
Conic. The latter performs better than the area-preserving Albers and located in mid-latitudes. Regarding the values separately
Conic (which distorts angles more than the Equidistant Conic). for each (non-conformal) projection type, a direct relationship
While the (cylindrical) Plate Carrée can compete with the conic between mean angular deformation and mean grid oversampling
projections in the southern zone D, its mean GOF is not compe- factor is evident, proving mean GOF being a function of skewing,
titive in northern zone C. quantified by wi . Nevertheless, grids are extensive (as opposed to
For the pole-centred zones E–H, we tested three common local) and need to be optimised for all locations. Thus, projections
azimuthal projections, namely the conformal Stereographic, the featuring also a small range of local scale factors throughout the
equal-area Lambert Azimuthal and the Equidistant Azimuthal map as the Equidistant Azimuthal are good choices for efficient
projection. Again, a clear superiority of the Equidistant Azimuthal and accurate grid systems. In the end, also the land compactness
can be observed, with a mean oversampling that is roughly five (E governs the oversampling, especially for equal-area projection,
and F) and three (G and H) times smaller. Nevertheless, the average where remote islands erratically reduce bmin as it is the case for
values of GOF for pole-centred setups (E and F) are not competitive Oceania.
with those of conic projections covering comparable areas (C and
D). Different to the conic case, the equal-area Lambert Azimuthal
and the angle-preserving Stereographic perform similarly. 3.4. Global grid systems

We seek a spatial reference for remote sensing data covering the

entire globe. Therefore, we compared the results of grid systems
3.3. Continental grids using a varying number of sub-grids for two global domains: A first
comparison examines setups of one, two and six sub-grids over
Azimuthal projections attain the performance of conic projec- major landmasses (561 o ϕ o 721 ¼ B ¼ C þ D
6 non-polar con-
tions, when considering continental-scale grid extents (Table 3). tinents) and a second examination considers all landmasses from
For grids adapted to the seven continents separately, the mean pole to pole ( 901 o ϕ o901 ¼ A ¼ B þ G þ H ¼ 7 continents).
oversampling is again lower (down to 1%) and obviously also a Table 4 lists the global average oversampling factors of the individual
function of the spatial extent. This was already exemplified in setups, as well as the summarised pixel area factors (pxAF) and the
Fig. 5(c), where entire Asia is projected almost without distortions. preserved geometry features.

Table 4
Average values of grid oversampling factors ðGOF Þ and pixel area factors (pxAF) for combined global grid systems.

Grid system—non-polar global

 561 o ϕ o 721 Subset zones Preserves GOF pxAF

Set of 1 projections
Plate Carrée B Scale ϕ 1.355 1.353
Sinusoidal B Area 10.736 10.720
Set of 2 projections
Equidistant Conic C, D Scale ϕ 1.041 1.039
Lambert Conic C, D Angles 1.083 1.082
Albers Conic C, D Area 1.165 1.164
Set of 6 projections
Equidistant Azimuthal Continents Centric Scale 1.025 1.022
Stereographic Continents Angles 1.077 1.075
Lambert Azimuthal Continents Area 1.162 1.158

Grid system—Global
 901 o ϕ o 901 Subset zones Preserves GOF pxAF

Set of 1 projections
Plate Carrée A Scale ϕ 1.807 1.797
Set of 3 projections
Plate Carrée B Scale ϕ
& Equidistant Azimuthal G, H Centric Scale 1.320 1.317
EASE 2.0 Cylindrical B Area
& EASE 2.0 Azimuthal G, H Area 9.025 9.011
Set of 4 projections
Equidistant Conic C, D Scale ϕ
& Equidistant Azimuthal G, H Centric Scale 1.037 1.035
Set of 7 projections
Equidistant Azimuthal Continents Centric Scale 1.023 1.021
92 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93

The first analysis over the non-polar landmasses allows follow- of this study's GOF analysis at a 11 resolution, considering the 1 km
ing conclusions: Using only one grid leads to an average over- resolution of the pxAF analysis.
sampling of at least 36% when projecting to the Plate Carrée. Using The second analysis over the entire Earth surface further
the Sinusoidal is not recommended since it oversamples by the confirms above findings that both, small grid extents and usage
factor 10. A grid setup of two conic projections decreases over- of equidistant projections, facilitate minimal oversampling. Four
sampling to 4–16% and one of six azimuthal projections to 2–16%. sub-grids have an average GOF of 3.7% and an optimised setup of
As expected, equidistant systems perform best, followed by con- seven continental grids of 2.3%. A three-grid system following the
formal and equal-area ones. EASE-Grid 2.0 approach (Brodzik et al., 2012) is far off those
We gained strong confidence in our oversampling estimates results, due to severe distortions on the sub-grid's peripheries
determined by GOF averages when we compared them to the caused by the underlying Lambert Cylindrical projection.
actual pixel counts in gridded test images, as determined by the The sets using equidistant projections outperform the confor-
pxAF ratios. As visible in Table 4, the pxAF matches the GOF mal and equal-area ones considerably and are therefore recom-
sufficient enough for each grid system. From this experiment, mended for use in global raster databases for high resolution
we learned also that the pxAF never exceeds the GOF and there- satellite data. Moreover, equidistant projections possess another
fore the latter can be regarded as reliable estimator of grid advantage as they feature a constant value of bi ¼ 1. They thereby
oversampling. This even holds true for the relatively coarse setup allow to set the grid spacing of all the sub-grids to one and the
same length, independent of the sub-grids’ extents. Otherwise
optimisation would entail different grid spacings among the sub-
grids and thus vitiating the data handling.
Finally, an optimised grid system definition was elaborated and
proposed for storage and processing of global high resolution
satellite data. Described already as the lowermost set in Table 4, it
consists of seven projected continental regular sub-grids using the
Equidistant Azimuthal with projection centres as in Table 3. We
named it the Equi7 Grid. Considering this study's findings, and
with processor- and user requirements in mind, the subset borders
were optimised by following requirements: (1) Landmasses should
form compact entities. (2) Borders should be preferably over
oceans. (3) Countries should preferably be not split. Schematic
illustrations of the (already projected) continental grids are dis-
played in Fig. 6. These seven grids cover the Earth entirely without
gap and with 50 km overlap over land borders.

4. Conclusion

In order to define efficient and accurate grids for satellite data,

we performed a distortions analysis of various map projections.
Resting on relevant literature and practical experience, we focused
on regular grids spawned by map projections listed in Table 1. The
local grid oversampling factor GOF developed during this research
can be regarded as a generalisation of metrics proposed in
previous studies that investigate scale distortions along meridio-
nal and latitudinal or vertical and horizontal directions. Here, the
orientation of the input image's rows and columns has no
influence on the metric. Taking the global minimum scale factor
bmin in the definition of the optimal grid sampling distance into
account ensures that the projection from image to grid is lossless
everywhere in the grid.
Compared to other projections tested in this study, and con-
trary to expectations, equal-area projections do not provide con-
ditions for efficient grids – although one might intentionally think
that equal-area grids are best suited for retaining the satellite
sensor footprint. However, raster images projected to conformal
grids do not suffer from skewing and disintegrated neighbour
relations. Hence, oversampling is lower than for equal-area grids
and correct application of filters and local operators is not
impeded. Equidistant projections feature moderate deformation
of angles (as well as areas) and are suitable alternatives if
preservation of angles is not rigorously required. When using
equidistant projections for grid definitions, the discarding of strict
conformity and area truth is rewarded with smallest possible
oversampling. Best cases of average oversampling over the Earth's
major landmasses are 36% for a single grid based on the Plate
Fig. 6. Schematic images of the proposed grid system's zonal projection of the Carrée, 4% for a grid system of two grids based on the Equidistant
Earth. Colours as in Fig. 4. Conic and 2% for seven grids bases on the Equidistant Azimuthal
 B. Bauer-Marschallinger et al. / Computers & Geosciences 72 (2014) 84–93 93

(the Equi7 Grid). In the end, these analytical measurements were Defense Mapping Agency, 1989. The Universal Grids—Universal Transverse Merca-
successfully validated by a pixel count experiment. tor (UTM) and Universal Polar Stereographic (UPS). Defense Mapping Agency.
DMA Technical Manual edition. DMATM 8358.2.
Spatial division of data in case of a system with several grids Drusch, M., Bello, U.D., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B.,
may demand a higher level of data handling during processing and Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F.,
archiving. Moreover, the display of composited global images Bargellini, P., 2012. Sentinel-2: ESA's optical high-resolution mission for GMES
operational services. Remote Sens. Env. 120, 25–36, The Sentinel Missions -
requires a single projection of the (output) data at some point. New Opportunities for Science.
However, the separation into continental grids provides natural Feeman, T., 2002. Portraits of the Earth: A Mathematician Looks at Maps.
and seamless spatial references for users interested in regional or Mathematical world, American Mathematical Society, Providence, Rhode
Island.
local applications. The drawbacks of additional processing with Finn, M.P., Steinwand, D.R., Trent, J.R., Buehler, R.A., Mattli, D.M., Yamamoto, K.H.,
such a combination of grids appear small compared to the costs of 2012. A program for handling map projections of small scale geospatial raster
storing and processing an unnecessary data overhead of at least data. Cartogr. Perspect., 53–67.
Hornacek, M., Wagner, W., Sabel, D., Truong, H.L., Snoeij, P., Hahmann, T., Diedrich,
36% as would be required in the case of a single global grid.
E., Doubková, M., 2012. Potential for high resolution systematic global surface
Independent from the choice of the underlying projection, the soil moisture retrieval via change detection using Sentinel-1. IEEE J. Sel. Top.
discussed regular grids allow easy and fast data access due to their Appl. Earth Obs. Remote Sensing 5, 1303–1311.
linearity and orthogonal structure. This facilitates also cross- Hortal, M., Simmons, A., 1991. Use of reduced Gaussian grids in spectral models.
Mon. Weather Rev. 119, 1057–1074.
referencing between layers of different resolution simply by Khlopenkov, K.V., Trishchenko, A.P., 2008. Implementation and evaluation of
application of linear functions. The use of map projections in concurrent gradient search method for reprojection of MODIS Level-1B
many remote sensing software packages and the imperative of imagery. IEEE Trans. Geosci. Remote Sens. 46, 2016–2027.
Kimerling, A.J., 2002. Predicting data loss and duplication when resampling from
two-dimensionality for some processing and most display inter- equal-angle grids. Cartogr. Geogr. Inf. Sci. 29, 111–126.
faces of spatial data provides additional motivation for the use of Kuntz, E., 1990. Kartennetzentwurfslehre: Grundlagen und Anwendungen. Wichmann
appropriate regular grids, rather than irregular grids, for satellite Herbert, Karlsruhe, Germany.
Luo, Y., Trishchenko, A.P., Khlopenkov, K.V., 2008. Developing clear-sky, cloud and
imagery databases. cloud shadow mask for producing clear-sky composites at 250-meter spatial
In summary, the (technically imposed) use of orthogonal resolution for the seven MODIS land bands over Canada and North America.
regular grids for high resolution remote sensing data does not Remote Sens. Env. 112, 4167–4185.
Maling, D., 1992. Coordinate Systems and Map Projections. Pergamon Press, Oxford,
necessarily generate much data oversampling. The employment of United Kingdom.
equidistant projections allows the most efficient data storage Mulcahy, K.A., 2000. Two new metrics for evaluating pixel-based change in data
while suppressing inaccuracies during transformation. sets of global extent due to projection transformation. Cartographica: Int. J.
Geogr. Inf. Geovisualization 37, 1–12.
Sahr, K., White, D., Kimerling, A.J., 2003. Geodesic discrete global grid systems.
Cartogr. Geogr. Inf. Sci. 30, 121–134.
Acknowledgments Seong, J.C., 2005a. Assessing resampling accuracy of categorical data using random
points. Cartogr. Geogr. Inf. Sci. 32, 393–400.
Seong, J.C., 2005b. Implementation of an equal-area gridding method for global-
This research has received funding from the Austrian research scale image archiving. Photogramm. Eng. Remote Sens. 71, 623–627.
Seong, J.C., Mulcahy, K.A., Usery, E.L., 2002. The Sinusoidal projection: a new
funding association (FFG) under the scope of the ASAP 9 program
importance in relation to global image data. Prof. Geogr. 54, 218–225.
within the research project 840114, Soil Moisture Data Cubes. Seong, J.C., Usery, E.L., 2001. Assessing raster representation accuracy using a scale
factor model. Photogramm. Eng. Remote Sens. 67, 1185–1191.
Shannon, C., 1949. Communication in the presence of noise. Proc. IRE 37, 10–21.
References Snyder, J.P., 1987. Map projections—a working manual. 1395, USGPO.
Steinwand, D.R., Hutchinson, J.A., Snyder, J.P., 1995. Map projections for global and
continental data sets and an analysis of pixel distortion caused by reprojection.
Bierig, R., Piroi, F., Lupu, M., Hanbury, A., Berger, H., Dittenbach, M., Haas, M., 2013. Photogramm. Eng. Remote Sens. 61, 1487–1497.
Conquering data: the state of play in intelligent data analytics. Position Paper. Sun, W., Zhao, X., Chen, J., 2007. A method for global-scale archiving of imaging
Birch, C.P., Oom, S.P., Beecham, J.A., 2007. Rectangular and hexagonal grids used for data based on QTM pixels. Data Sci. J. 6, S301–S309.
observation, experiment and simulation in ecology. Ecol. Modell. 206, 347–359. Tissot, A., 1881. Mémoire sur la représentation des surfaces et les projections des
Brodzik, M.J., Billingsley, B., Haran, T., Raup, B., Savoie, M.H., 2012. Ease-grid 2.0: cartes géographiques. Gauthier-Villars.
incremental but significant improvements for earth-gridded data sets. ISPRS Usery, E.L., Finn, M.P., Cox, J.D., Beard, T., Ruhl, S., Bearden, M., 2003. Projecting
Int. J. Geo-Inf. 1, 32–45. global datasets to achieve equal areas. Cartogr. Geogr. Inf. Sci. 30, 69–79.
Bugayevskiy, L., Snyder, J., 1995. Map Projections: A Reference Manual. Taylor & White, D., 2006. Display of pixel loss and replication in reprojecting raster data
Francis, London, United Kingdom. from the Sinusoidal projection. Geocarto Int. 21, 19–22.