Store, manipulate and analyze raster data within

the PostgreSQL/PostGIS spatial database

Boreal Avian
Pierre Racine Modelling Project

Research Assistant
Centre for Forest Research

Steve Cumming
Centre for Forest Research
Wood and Forest Science Department
University Laval, Quebec, Canada

FOSS4G Denver
September 2011
 Introducing PostGIS Raster
• Support for rasters in the PostGIS spatial database
- RASTER is a new native base type like the PostGIS GEOMETRY type
- Implemented very much like and is as easy to use as the PostGIS
GEOMETRY type

One table row = one raster (or tile)

One table = one coverage
- Integrated as much as possible with the GEOMETRY type

SQL API easy to learn for users already familiar with PostGIS

Full raster/vector analysis capacity taking nodata value into account

Operators & functions works seamlessly when possible
- First release with PostGIS 2.0 (soon)
• Development Team
- Current: Bborie Park, Jorge Arevalo, Pierre Racine,
David Zwarg, Regina & Leo Obe
- Past: Sandro Santilli, Mateusz Loskot
Chapter 13 on
• Founding PostGIS Raster

- Steve Cumming through a Canada Foundation for Innovation grant

- Deimos Space, Davis University, Cadcorp, Azavea, OSGeo
 Georeferenced, Multiband,
Multiresolution and Tiled Coverages
upperLeftX,
• Each raster/tile is georeferenced upperLeftY skewX
scaleX
- Supports rotation (or skew)
skewY

• Supports multiple bands with different

pixeltypes in the same raster scaleY

1BB, 8BSI, 8BUI, 16BSI, 16BUI, 32BSI, 32BUI, 32BF, 64BF
- One nodata value per band

• Tiled & indexed e.g. SRTM Coverage for Canada

- No real limit on size

1 GB per tile, 32 TB per coverage (table)

Rasters are compressed (by PostgreSQL)
- Supports irregularly tiled & overlapping coverages

• Other resolutions (or overviews) are stored in sister tables

• List of raster columns available in a raster_columns table similar to
the geometry_columns table
 Supports Many Raster Arrangements

missing missing
tile tile

overlaps smaller same

tiles size tiles

a) warehouse of untiled b)irregularly tiled raster c) regularly tiled raster

and unrelated images coverage (36 tiles) coverage (36 tiles)
(4 images)

Table 1

empty space
Table 2

d)rectangular regularly f) rasterized geometries

tiled raster coverage e) tiled images (2 tables coverage (9 lines in the
(54 tiles) of 54 tiles) table)
 What You Can Do Now?
Store and manage rasters in the database…
• Batch import rasters
- raster2pgsql.py -r “c:/temp/mytiffolder/*.tif” -t mytable -s 4326
-k 50x50 -I | psql -d testdb
• Get and set the raster properties
- Upper left corner coordinates & transformation parameters
- SRID & number of bands
• Get and set band properties
- Pixel type & nodata value
• Reproject raster (ST_Transform)
• ST_Resample(raster), ST_Rescale(), ST_SnapToGrid()
• Convert a geometry to a raster (ST_AsRaster)
• Convert a raster to a set of geometries-values
(ST_DumpAsPolygons)
 What You Can Do Now?
Dump rasters from the database…

• With the 'PostGISRaster' GDAL driver

- Developed and maintained by Jorge Arévalo

• Read only, optimization in progress

• The write part is still to do (by you?)

• Two modes
1. ONE_RASTER_PER_ROW
2. ONE_RASTER_PER_TABLE
 What You Can Do Now?
Get raster statistics…
• ST_SummaryStats(raster)
- Return a set of (min, max, sum, mean, stddev, count (of
withdata pixels)) records
- 10 seconds for one SRTM tile of 3600 x 3600 pixels, 70MB

• ST_Histogram(raster, bin, width[ ])

- Return a set of (min, max, count, percent) records for an array
of bins

• ST_Quantile(raster, quantiles[ ])
- Return a set of values for an array of quantile

• ST_ValueCount(raster, values[ ])
- Return the frequency for an array of value
 What You Can Do Now?
Display rasters…
• QGIS plugin by Maurício de Paulo
([email protected])
• gvSIG plugin by Nacho Brodin ([email protected])
• MapServer through GDAL
- Normally any software using GDAL to read raster and allowing
passing database connection parameters to GDAL

• Display a vectorization of the raster

- OpenJump

SELECT ST_AsBinary((ST_DumpAsPolygons(rast)).geom),
(ST_DumpAsPolygons(rast)).val
FROM srtm_tiled WHERE rid=1869;
- ArcGIS 10

Add Query Layer (same as OpenJump but without ST_AsBinary())
- Any software displaying vector PostGIS queries
 What You Can Do Now?
Edit and compute new rasters…
• ST_SetValue() of a pixel

• ST_Reclass() a raster

• ST_MapAlgebra(raster, band,
-4 2 0 6 null null
expression, null
-1 -4 2 9 6
nodatavalueexpr, null null
-2 0 1 8
pixeltype)
- Expressions are evaluated by the PostgreSQL parser
- You can use any complex SQL expression
- e.g. 'CASE WHEN rast < 0 THEN rast+10 ELSE NULL END'
- You can provide a nodatavalueexpr to handle source nodata
values
 What You Can Do Now?
Convert rasters to any GDAL format in SQL…

• Get the list of GDAL drivers available (ST_GDALDrivers)

• Convert to any of the available format (ST_AsGDALRaster)

- SELECT ST_AsGDALRaster(rast, 'USGSDEM')
FROM srtm_22_03

• ST_AsTIFF(), ST_AsJPEG(), ST_AsPNG()

 What You Can Do Now?
Intersects rasters with points and lines…
• Extract ground elevation values for lidar points…
- SELECT pointID, ST_Value(rast, geom) elevation
FROM lidar, srtm WHERE ST_Intersects(geom, rast)

• Intersect a road network to extract

elevation values for each road segment
- SELECT roadID,
(ST_Intersection(geom, rast)).geom road,
(ST_Intersection(geom, rast)).val elevation
FROM roadNetwork, srtm WHERE ST_Intersects(geom, rast)

1000m

0m
 What You Can Do Now?
Intersects rasters with polygons…
• Compute the mean temperature for each polygons of a table
SELECT bufID, (gv).geom buffer, (gv).val temp
FROM (SELECT bufID, ST_Intersection(geom, rast) gv
FROM buffers, temperature
WHERE ST_Intersects(geom, rast)
buffers temperature result
geom pointid raster geom pointID temp
id=24
polygon 24 raster polygon 24 11.2
polygon
polygon
46
31
∩ raster
raster
= polygon
polygon
53
24
13.4
15.7
temp=11.2

raster polygon 23 14.2

id=24
polygon 45
… … … …
temp=15.7
… …

• Results must be summarized per buffer afterward

• All analysis functions take nodata values into account
• Have a look at the tutorial in the PostGIS Raster wiki page!
 What You Can Do Now?
Create a high resolution analysis grid for a large area…

Compute values of many variables for each cell of a grid

• Road & river length, mean temperature, population, water surfaces, etc…
• Easy in vector mode (1 cell = 1 polygon) but

• What about all of USA at 10m?

500 000 x 300 000

= 3000 km
Way too many polygons!
5000 km
• Manageable in raster format!
• 15 000 000 tiles 100x100 pixels
1. Create a raster having a uid
per pixel
2. Intersect your vector layers
with your raster grid
3. Summarize per pixel uid
4. Create a new band for each variable and assign the values
 What You Can Do Now?
Create a specialised web or desktop GIS application…

• With the raster API, PostGIS is now a very complete SQL GIS
- All data are implicitly tiled and spatially indexed
- No need to write complex C,C++, Python or JAVA code to
manipulate complex geographical datasets.
Desktop or Web
- Use SQL: The most used, most easy and most
Applicaton
minimalist though complete language to work (query building
with data in general. Easily extensible (PL/pgSQL) & display)
- Keep the processes close to the data where the
data should be: in a database! attribute,
vector
• Lightweight multi-users specialized SQL or raster
desktop and web GIS applications table
- All the (geo)processing is done in the database
- Applications become simple SQL query builders Spatial Database
and data (results) viewers (geoprocessing)
 What You Can Do Now?
Implement a WPS server raster/vector geoprocessor…

Desktop or Web
WPS Client

WPS WPS
query response

WPS Server

attribute,
vector
SQL
or raster
table

PostGIS
(geoprocessing)
 Performance?
• Import of 1 GB SRTM DEM files
- tiled to 48373 100x100 pixels tiles: 3 minutes
- tiled to 525213 30x30 pixels tiles: 6 minutes
• ST_Intersection() of 814 buffers with the 30x30 SRTM
- 4 minutes
• ST_Intersection() of 100 000 lines with a 300 MB landsat
coverage
- 8 minutes
• Recently selected by the main Canadian governmental
provider of geospatial data (GeoBase)
- online on-the-fly and internal elevation product generation
• PostGIS raster is still a baby, many optimizations are
still possible
 Summary
• PostGIS Raster is multiband, tiled, multiresolution
- Each band supports one nodata value, one pixel type.
- One row = one raster, one table = one coverage.
- Supports many tile arrangement.
- Very much like a vector coverage.
- Import is done the same way as usual with PostGIS:
raster2pgsql
• There are plenty of functions to…
- manipulate,
- edit,
- do raster and raster/vector analysis,
- get raster statistics,
- create new rasters,
- write web and desktop applications.
 Summary

• Roadmap…
- Two raster version of ST_Intersection()
- Neighbor version of ST_MapAlgebra()
- Two rasters version of ST_MapAlgebra()
- Aggregate rasters with ST_Union()
- Statistic functions as aggregates
- ST_Interpolate() from irregular grid of point (lidar)
- ST_AsDensity() to produce density maps

• Third party developments…

- GDAL write driver
- Support in GeoServer
- Read/write in FME
 What You Can Do Soon?
More complex analyses…
• Two rasters ST_Intersection()
- ST_Intersection(raster, raster) -> raster
- Equivalent to ST_Clip(raster, ST_AsRaster(geometry))

• One raster neighbor version of ST_Mapalgebra()

- or “ focal function ” or “moving window” computation
- User function taking a 3x3, 5x5, 7x7, or more
raster and optional parameters and returning a value
-10 0 0
• Two rasters version of ST_MapAlgebra
-4 -12 2
20-2 -6
0
- Useful to implement most overlay functions and more
-4-5 20-2
-1 -4.5 1

ST_Union(raster, raster) -> raster

ST_Intersection(raster, raster) - > raster -2 0 1

ST_BurnToRaster(raster, geometry, value)…
- Resulting extent can be FIRST, SECOND, UNION or INTERSECTION.
 What You Can Do Soon?
Aggregate many tiles into one raster… (or merge)

• Use ST_Union() as an aggregate function

- Taking a state, a temporary and a final function specifying how to
aggregate pixel values in a state, a temporary and a final raster
- User can defines their own expressions or use predefined
functions like FIRST, LAST, MIN, MAX, SUM, MEAN, COUNT

• Ex. SELECT ST_Union(raster, 'MEAN')

- Compute the mean pixel value of many overlapping pixels
- The state function 'SUM' accumulate pixel values
- The temporary function 'COUNT' count the number of pixels
- The final function 'state raster/temporary raster' divide the sum
by the count
- See pl/pgsql code in raster/script/plpgsql/st_union.sql
 Thanks!
http://trac.osgeo.org/postgis/wiki/WKTRaster

Chapter 13 on PostGIS
Raster April 2011
Some extra slides…
 Comparison with Oracle GeoRaster

Oracle GeoRaster PostGIS Raster

• Stored as a one to many relation • Stored as a unique type, in one
between two types, in two table
different tables - RASTER (or tile)
- SDO_GEORASTER (raster) - Each raster is geoferenced
- SDO_RASTER (tile)
- Only SDO_RASTER is georeferenced
• Supports (too) many raster • Supports the minimal set of
features for any kind of raster characteristics for the geospatial
application industry
- bitmap mask, two compression - georeference, multiband, tiling,
schemes, three interleaving types, pyramids, nodata values
multiple dimensions, embedded
metadata (colour table, statistics,
etc…), lots of unimplemented features • Easy to load data
• Hard to load data • Designed for raster/vector
• Designed for raster storage analysis
 What You Can Do Now?
Store and manage rasters stored outside the database…

• By default raster are stored INSIDE the database in the

PostGIS raster format
Web Client
• It is also possible to register in the database
rasters stored OUTSIDE in the file system HTTP

- Stored in any GDAL format Web server

- Faster direct access for web apps (JPEGs) Web service
- Avoid useless database backup of large
datasets not requiring edition SQL
JPEGs
- Avoid importation (copy) of large BD Image01.jpg
landcover
datasets into the database raster Image02.jpg
raster
- Provides an easy SQL API to raster Image03.jpg
raster
Image04.jpg
manipulate/analyse raster files raster
… …
- Use the –R raster2pgsql.py option
- All functions should eventually works
seamlessly with out-db raster. Now only a few.
 What You Can Do Now?
Develop new raster processing functions…

• ST_MakeEmptyRaster()
• ST_AddBand()
- Empty band or copy a band from another raster
• All georeference setters
- ST_SetScale (), ST_SetSkew(), ST_SetUpperLeft(), ST_SetGeoReference()
• ST_SetBandNodataValue
• ST_SetValue()
• Coordinates transformation helpers
- ST_World2RasterCoordX(), ST_World2RasterCoordY(),
ST_Raster2WorldCoordX(), ST_Raster2WorldCoordY()
• ST_Intersection() & ST_intersects()
- To interact with vector data
• Many more…
 What You Can Do Now?
Develop new raster processing functions…

• PL/pgSQL example for ST_DeleteBand

CREATE OR REPLACE FUNCTION ST_DeleteBand(rast raster, band int)
RETURNS raster AS $$
DECLARE
numband int := ST_NumBands(rast);
newrast raster := ST_MakeEmptyRaster(rast);
BEGIN
FOR b IN 1..numband LOOP
IF b != band THEN
newrast := ST_AddBand(newrast, rast, b, NULL);
END IF;
END LOOP;
RETURN newrast;
END;
$$ LANGUAGE 'plpgsql';
 What You Can Do Soon?
Write to PostGIS raster with GDAL…

• A write GDAL driver does not exist yet.

• It should allows
- loading raster in the database using gdal_translate
- loading many raster at the same time
- any application writing to GDAL to write to PostGIS raster
- tiling a raster to any tile size
- to create overviews
 What You Can Do Soon?
Complex MapAlgebra analyses…
• Already available: One raster version of ST_MapAlgebra()
• Soon: Faster user-defined function version
- Function taking a pixel value and some
parameters and returning a computed value

CREATE FUNCTION polynomial(x float,
VARIADIC args TEXT[])
RETURNS FLOAT AS $$
DECLARE
m FLOAT;
b FLOAT;
BEGIN
m := args[1]::FLOAT;
b := args[2]::FLOAT;
return m * x + b;
END; $$ LANGUAGE 'plpgsql';

SELECT ST_MapAlgebra(raster, 'polynomial', ARRAY['1.34', '5.2'])

 What You Can Do (maybe not too) Soon?
Interpolate a raster coverage from a point coverage…

• ST_Interpolate(pts geometry)
- Should be an aggregate returning one raster
(or a set of tiles)
- Implementing many different interpolation
algorythms

Nearest neighbor, linear, polynomial
- Very useful to convert lidar data to raster

• ST_AsDensity(geometry)
- Count the number of features touching each
pixel and then smooth the surface using a
moving window (neighbor map algebra)
 What You Can Do (maybe not too) Soon?
Create a clean raster coverage… from a messy one…

1. Load a bunch of unaligned overlapping rasters

(e.g. landsat)

2. ST_SetBrightness() & ST_SetContrast()

- or ST_NormalizeColor('table', 'rasterColumn')

3. ST_MakeEmptyRasterCoverage()
- Create a vector grid or an empty raster coverage based on a set
of parameters

4. ST_MapAlgebra(emptyRaster, messyRaster, 'MEAN',

'FIRST') -> raster