XXIX-B4, 2012
ore or have multiple
y important solution
large task into many
asks concurrently in
odes. As a result, the
e major benefits of
'up (more users) for
hould be an essential
e that concurrency is
ich can help improve
Cie, 2006).
erful SQL parallel
SQL-based operation
Database in parallel.
Oracle Database it is
-sources, which are
jlan (Dijcks, 2010).
ork, however, the
uch as mosaic and
lirectly parallelized
Ss is because each of
ipulation operations
oic in how the raster
ed.
le parallel execution
table function is an
ictions can be used
et of table functions
'eraging the parallel
racle 2008. Dijcks,
ncapsulate complex
an process different
ect in parallel. To
>gin with explicitly
1 and deciding what
Jbprocess. We used
nto subsets and the
egree of parallelism
ut. Then the Oracle
1c whole task into
mber of subsets and
sets independently.
cess is done.
tery finds all pixels
of the first band is
/alue of the second
s than 150, and the
200 and less than
meeting the query
the process will be
ich will process a
y, thus the overall
orid = 1;
eorid = 2 for update;
»200)&( (21-245),
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
‘parallel=4");
update georaster table set georaster — georl where georid = 2;
commit;
end;
We conducted some initial tests on the parallel implementation
of sdo geor ra.findcells as well as sdo geor ra.classify. The
sdo geor.findcells test uses the above script to query an image
based on all three bands while the sdo geor ra.classify test
segments the first band of the three-band image into 12 classes.
We used a x86. 64 Linux Server to do the tests. It has 4 Intel(R)
Xeon(R) X5670 CPUs and the CPU speed is 2.93GHz. The
total memory is 8GB. The operating system is Red Hat
Enterprise Linux Server 5.4. We used four 3-band images of
different sizes. The results are shown in table 1 and 2.
Table 1. Execution time in seconds of sdo geor.findcells with and without parallelism
image size and image dimension size in row x column x band
Degree of
Parallel ism 238.7MB 954.8MB 2.15GB 3.82GB
8930x8912x3 17860x17824x3 26790x26736x3 35720x35648x3
1 17.33 45.14 87.80 153.36
2 15.33 26.43 55.19 106.00
4 11.67 23:33 49.80 80.79
8 10.12 20.68 40.92 74.63
Table 2. Execution time in seconds of sdo_geor.classify with and without parallelism
image size and dimension size in row x column x band
Degree of
Parallelism 238.7MB 954.8MB 2.15GB 3.82GB
8930x8912x3 17860x17824x3 | 26790x26736x3 35720x35648x3
1 6.47 44.03 80.50 138.45
2 4.89 36.33 58.19 86.68
4 3.86 26.52 44.84 71.55
8 3.50 21.35 41.09 68.00
There is only 1 disk on this machine so the I/O contention
among parallelized subprocesses is very high. That has a big
impact on the performance numbers. However, even with only
one disk, table 1 and 2 show that when the raster algebra
operation is parallelized into 2 to 8 subprocesses, the processing
operation is significantly faster than the same operation without
parallelism. In addition, the overall performance improvement
scales very well with image size increasing as shown in table 1
and 2. There are still more room for improvement yet to be
done. However, we can reasonably assume the performance
improvement could be much better if the machine has more
CPU's and more memory, and particularly if a high-speed
storage cluster (using Oracle ASM technology) or a high-end
machine such as Oracle Exadata Database Machine is used.
Some database tuning techniques will help improve parallel
performance as well.
S. APPLICATIONS
Currently, the raster algebra engine implementation is focused
only on the “local” function type of map algebra and is
designed to work with the standard PL/SQL language and run
completely inside the database. Using the PL/SQL and the
raster algebra expressions and functions, users can implement a
wide range of applications, such as applying complex pixel
queries in the database, editing a raster based on raster cell
values and formulated query conditions, segmenting images or
classifying a thematic map, and conducting cartographic
91
modeling over a large number of rasters and images of
unlimited size. The engine runs these algebra expressions and
functions as single processes inside the database and each of
those processes can be parallelized, thus dramatically improves
the analytical capability and performance of the GeoRaster
database.
Map Algebra is mainly used in cartographic modeling and is
considered an essential component of any GIS systems. These
applications and the importance of the map algebra expressions
and functions are well known. Due to the lack of testing dataset
of thematic layers for a case study area and the easy access of
Landsat imagery, we only use Normalized Difference
Vegetation Index (NDVI) and Tasseled Cap Transformation
(TCT) as our application examples in this paper to demonstrate
the capability.
In remote sensing, NDVI was one of the most successfully and
widely used vegetation index (VI), which can quickly identify
vegetated areas and monitor plant growth or their "condition".
Using Landsat TM imagery, the standard NDVI computation
formula is (TM4 — TM3) / (TM4 + TM3). The following script
takes a Landsat 7 ETM+ image and compute the NDVI, which
is stored as another raster of floating number data type. Note, in
our algebra language, band number starts with 0, so the formula
translates into the expression '({3}-{2})/({3}+{2})’.
declare
georl MDSYS.SDO GEORASTER;
geor2 | MDSYS.SDO GEORASTER;
begin
-- source ETM+ image
select georaster into georl from georaster table where georid = 2;
-- to store NDVI
select georaster into geor2 from georaster table where georid — 3 for update;
mdsys.sdo geor ra.rasterMathOp(georl,
SDO STRING ARRAY(((3)-(2)/((3)-(21)),
'celldepth=32bit real',geor2);
update georaster_ table set georaster = geor2 where georid = 3;
commit;
end;
Figure 1 shows a small area of the original ETM+ 543 image
and the resulting NDVI image after running the above script.
Fig. 1, ETM+ 543 color image (left) and NDVI image (right). Image
Courtesy of PCI Geomatics.
The concept of tasseled cap transformation is a useful tool for
compressing spectral data into a few bands associated with
physical scene characteristics (Crist and Cicone 1984). TCT
helps analyze the physical ground features. With Landsat
imagery, it uses 5 bands of either original digital number (DN)
or reflectance data to generate 6 new bands, each of which
represents different ground features. The 6 resulting bands are
