Central Place Indexing:

Optimal Location
Representation for
Digital Earth
Kevin M. Sahr

Department of Computer Science

Southern Oregon University

1 Kevin Sahr - October 6, 2014

 The Situation

• Geospatial computing has achieved

many impressive results
• But it now faces unprecedented
challenges

2 Kevin Sahr - October 6, 2014

Advent of “Digital Earth”
• exemplifies the challenges facing geospatial
computing, combining in one platform:
✦ “mother of all (geospatial) databases”
✦ simulation, interactive 3D visualization, &
analysis of vast quantities of diverse
distributed global geospatial “big data”
✦ integrates real-time location update and
manipulation

3 Kevin Sahr - October 6, 2014

 The Key: Location
Representation
• to implement this vision in totality a
revolution in our fundamental approaches
to geospatial computing is required
• at the core of all geospatial applications
are data structures that represent location
✦ even minor efficiency improvements in
location representation can lead to
substantial performance increases

4 Kevin Sahr - October 6, 2014

Fundamental Location
Representation Types
• digital earth systems must provide data
structures for representing:
✦ raster/pixels for
✤ imagery
✤ discrete simulation
✤ “gridded” data analysis
✦ vector/point locations
✦ spatial databases/spatial indexes

5 Kevin Sahr - October 6, 2014

 Traditional Raster
Location
Representation
• raster of square pixels
• addressed using 2-tuple of integers

6 Kevin Sahr - October 6, 2014

 Traditional Image
Processing Model
• traditional raster representation supports
image processing based on a conceptual
model of:
✦ input from square raster of sensors
✦ stored internally as matrix of pixels
✦ displayed one-to-one on a square raster of
display pixels

7 Kevin Sahr - October 6, 2014

Digital Earth Reality
• image processing in digital earth
systems breaks this mold
✦ processed satellite image pixels rarely
correspond to individual sensors
✦ must support whole-earth image sets
✤ spherical topology
✦ internal pixels mapped to virtual 3D
surface for display

8 Kevin Sahr - October 6, 2014

 A Superior Alternative
• numerous researchers have proposed using
hexagon-shaped pixels, arranged in a
hexagonal raster
• the human eye uses a hexagonal
arrangement of photoreceptors
• compared to square rasters, hexagon rasters
✦ are 13% more efficient at sampling
✦ 25% to 50% more efficient for common
image processing algorithms

9 Kevin Sahr - October 6, 2014

 Discrete Simulation
• hexagonal grids also have numerous
advantages over square grids for discrete
simulation
✦ superior angular resolution
✦ discrete distance metric better approximates
cartesian distance
✦ display uniform unambiguous adjacency

10 Kevin Sahr - October 6, 2014

 Traditional Vector
Location Representation
• 3- or 2-tuples of floating point values
• attempt to mimic the real number
coordinates used in pre-computer scientific
analysis and 2D mapping

11 Kevin Sahr - October 6, 2014

 But…

• vectors of real numbers

✦ are continuous and infinite
• tuples of floating point values
✦ are discrete and finite

12 Kevin Sahr - October 6, 2014

 Problems
• the simplest operations (e.g., equality
test) can result in profound semantic
errors
• bounding the rounding error on
individual operations can be difficult
✦ on complex systems can be
impossible

13 Kevin Sahr - October 6, 2014

 The Reality
• floating point values are no more “exact” than
integer values
• given n bits, we can distinguish 2n unique
values
✦ all other points must be quantized to these
• all computer representations of location —
both raster and vector — are necessarily
discrete

14 Kevin Sahr - October 6, 2014

 A Superior Alternative
• the human brain represents location using a
hexagonal arrangement of neurons
• quantization to the points of a hexagonal lattice
is optimal using multiple formulations
✦ least average quantization error
✦ covering problem
✦ packing problem

15 Kevin Sahr - October 6, 2014

Traditional Spatial DBs
• traditional raster and vector representations
are inefficient for many common spatial
operations
• spatial DBs add a linear spatial index
✦ correspond to buckets containing
locations, providing
✤ more efficient spatial queries
✤ coarse filter for specific algorithms

16 Kevin Sahr - October 6, 2014

Traditional Spatial DBs
• underlying vector/raster representation
retained for
✦ final stage of many algorithms
✦ arbitrary spatial operations
• form of spatial DBs based on traditional
vector/raster representational forms
✦ structured: square quad tree
✦ semi-structured: rectangular buckets (e.g.
R-tree)

17 Kevin Sahr - October 6, 2014

 Spatial Queries
• traditional primary spatial query type:
window/axes-aligned rectangle
• but primary query type in modern
geospatial systems is proximity
✦ recall that hexagonal discrete distance
metric better approximates cartesian
distance
✤ hexagon buckets provide more
efficient proximity coarse filter

18 Kevin Sahr - October 6, 2014

 The Task
• design a hierarchical integer index for
hexagon lattices that can be used for:
✦ multi-precision vector location
✦ multi-resolution raster location
✦ structured spatial index
• must be explicitly spherical
• Digital Earth Primary Spatial Index:
✦ One Index to Rule Them All

19 Kevin Sahr - October 6, 2014

Hexagon Coordinate
Systems
• single resolution hexagon grids have
three natural axes spaced at 120°
angles j axis

i axis

k axis

20 Kevin Sahr - October 6, 2014

 Hexagonal Multi-Res
• regular multi-precision/resolution hexagon
lattices can be created with an infinite number
of apertures
✦ aperture: ratio of cell areas between
resolutions
• research has focused on the Central Place
(Christaller, 1966) apertures 3, 4, and 7

21 Kevin Sahr - October 6, 2014

Central Place Apertures
aperture 3 aperture 4 aperture 7

22 Kevin Sahr - October 6, 2014

 Prefix Codes
• hierarchical prefix codes have many
advantages for hierarchical spatial indexes
✦ each digit in index corresponds to a single
precision in the representation
✤ provides locality preserving total ordering
✤ implicitly encodes precision
✤ provides efficient generalization via
truncation

23 Kevin Sahr - October 6, 2014

 Aperture 7 Case
• note that each hexagon is naturally associated
with 7 hexagons at the next finer resolution

24 Kevin Sahr - October 6, 2014

 GBT
• Generalized Balanced Ternary (GBT)
(Gibson & Lucas, 1982) is a hierarchical
prefix code system for aperture 7 grids
j

2 6
• single digits correspond
3
0 4 i
to each possible
hexagonal direction
1 5

• each indexing child

adds the appropriate
digit to their parent’s
index
25 Kevin Sahr - October 6, 2014
 Apertures 3 and 4

• note that in
apertures 3 and 4
each cell also
naturally has 7
finer precision
potential indexing
children

26 Kevin Sahr - October 6, 2014

Central Place Indexing
• we can apply the GBT arrangement to the
aperture 3 and 4 cases
• we call the result Central Place Indexing
(CPI)
✦ provides uniform indexing for all 3
apertures
✦ allows for indexing mixed-aperture
resolution sequences

27 Kevin Sahr - October 6, 2014

Pixel/Bucket Indexing
• cells in aperture 3 and 4 resolutions can
have multiple parents cells and
therefore multiple valid CPI indexes
✦ aperture 3 example:

A4 B2
A B

28 Kevin Sahr - October 6, 2014

Pixel/Bucket Indexing
• if the cells represent pixels or DB buckets,
then a single unique index must be chosen
for each cell
✦ a consistent choice of child assignment
must be made
✦ example aperture 3 solutions:
ib+1 ib+1 ib+1

ib ib ib

G123G5G G1G3G56 G12GG56

29 Kevin Sahr - October 6, 2014

 Vector Indexing
• in apertures 3 and 4 point quantization can
be performed at each successive resolution

A A4 B
B2

30 Kevin Sahr - October 6, 2014

 Vector Indexing
• thus aperture 3 and 4 grids effectively
address cell sub-regions
✦ provides true multi-precision point
quantization
✦ truncation and rounding are
equivalent
✦ indexes can be progressively
transmitted

31 Kevin Sahr - October 6, 2014

 CPI Algorithms
• we have implemented planar CPI algorithms
for
✦ forward & inverse quantization
✦ addition/translation
✦ subtraction
✦ metric distance
• implemented using efficient per-digit table
lookups

32 Kevin Sahr - October 6, 2014

 The Sphere
• we can apply any CPI system
to a spherical icosahedron to
index a hexagonal Discrete
Global Grid System (DGGS)

• note that cells centered on the icosahedral

vertices are pentagons
✦ we can apply CPI indexing to them by
deleting one of the non-centroid indexing
sub-sequences

33 Kevin Sahr - October 6, 2014

 Conclusions
• multi-resolution hexagonal DGGSs provide
the best known basis for raster, vector, and
spatial DB location representation for digital
earth systems
• CPI provides a unified efficient hierarchical
indexing for all types of location
representation on hexagonal DGGSs

34 Kevin Sahr - October 6, 2014

www.discreteglobalgrids.org

35 Kevin Sahr - October 6, 2014