Image2006 ESAEUSC

1
GMES Fast Track Land Service 2006-2008 Production
of European orthoimage coverage
Rupert Müller, Thomas Krauß, Manfred Lehner,
Peter Reinartz
2
Outline

• Background of project Image2006
• Requirements The Statement of Work
• Processing chain and methodology
• Results and Experiences
• Processing infrastructure

3
Background of the Project IMAGE2006

• The GMES (Global Monitoring for Environment and
  Security) Fast Track Land monitoring Service
  (FTLS) is a service to provide on a regular basis
  
• European Mosaic (MOSAIC2006)
• land cover changes
• land cover map (CLC2006, CORINE)
• high resolution forest layer
• built-up areas including soil sealing.
• Since the mid 1980s a European Land Cover
  dataset has been regularly produced with now an
  increased frequency update to every five years
  from every ten years.
• On behalf of ESA (ESRIN) the DLR established an
  automatic processing chain to orthorectify about
  3000 satellite images (two European coverages).
• Within the GMES FTLS 2006-2008 a new dataset of
  orthorectified satellite images has to be
  produced covering the EU25 and neighbouring
  countries (total 38 countries), referred to as
  IMAGE2006.

4
From satellite raw image to orthorectified product
Definition Orthorectification Convert imagery
into map-accurate forms by removing sensor,
satellite motion and terrain related geometric
distortions from raw imagery
Satellite orbit
Satellite image (perspective projection per each
scan line)
Terrain Digital elevation model(DEM)
Orthoimage
5
high resolution satellite images shall be
orthorectified
up to 30 off-nadir view ? ?s 0.58 ?h
6
a DEM of accurate enough quality shall be used

• European DEM (lt60 latitude north)
• is derived from the data sets
• SRTM-C v2
• SRTM-X
• MONAPRO (parts in Alps)
• GLOBE (parts in east Turkey)
• within a fusion process
• using quality layers gt best-of-DEM
• (parts manually edited)
• Height accuracy (1?)
• 6m (flat areas)
• 30m (mountainous areas)
• Switzerland from maps 125000
• 1.5-10m accuracy 25m resolution

Projection Geographic (WGS84) Height Ellipsoid
(WGS84) Resolution 1 arcsec
Note countries above 60N latitude are
sub-contracted to the Swedish company METRIA
7
about 3000 scenes shall be processed within the
project life cycle

• Establish a Processing Storage environment
  suitable for parallel processing
• Establish administration tools to handle about
  3000 scenes
• Allocate sufficient staff for quality control and
  supporting tasks
• Implement automatic processing chain to handle
  bulk data

8
a geometric accuracy of RMSE lt 20m w.r.t. the
reference data shall be achieved

• Provided metadata for geometric correction are
  not accurate enough to achieve this requirement
  (deviation on ground about 30m for SPOT5 and up
  to 500m for IRS-P6 and SPOT4)
• Therefore Ground Control Points (GCPs) have to
• be used for improvement of geometric accuracy
• GCPs are points in image with known 3D object
  coordinates
• by manual in situ GPS measurement
• extraction from maps
• extraction from reference images and DEM
• Apply image matching techniques for automatic
  extraction of highly accurate and sufficient GCPs
  from the reference image data to improve the
  model
• (up to 0.1 pixel matching accuracy, about 450
  high quality points for 1000 km² found)

GCP 11.333456E 47.821445N 630m
Image line 4242.6 Column 101.2
Original Image
9
Overview of automatic processing chain
Data pool of original images
DEM database (w42)
Reference image database (Europe)
DEM
Original image
Reference image
Matching

• Metadata
• ephemeris
• attitude
• sensor
• time
• RPC
• .

Tie points GCP / ICP
Improvementsof model parameters
Generation of ortho image
Back to processing loop
Manually measured ground control points
Quality check
Ortho image
Data Delivery
10
absolute reference data sets shall be used

• Image2000 mosaic data
• covering 25 countries and
• derived from panchromatic
• ETM data
• Projection Geographic WGS84
• Resolution 0.000115 (12m)
• official absolute accuracy
• RMSEx 9-15 m
• RMSEy 7-18 m
• USGS ETM Land Cover Data
• covering all other countries
• Projection UTM WGS84
• Resolution 28.5 m
• official accuracy (global)
• RMSEx,y 50 m

Absolute geolocation accuracy of reference images
derived from 12 GCPs of superior quality in
southern Bavaria
11
Methodologies (Key Technologies)

• Direct Georeferencing (DG) for SPOT4 HRVIR and
  Spot 5 HRG
• Rational Polynomial Functions (RPF) for IRS-P6
  LISS III

12
Direct Georeferencing for SPOT4 HRVIR and SPOT5
HRG
m mapping frame
Basic Equation ( Collinearity Equation )
LoS correction by GCP
13
Direct Georeferencing for SPOT4 HRVIR and SPOT5
HRG

• satellite position and velocity every 30sec
  (DORIS ? lt 1m)
• attitude and attitude rate with 8Hz (STS and
  gyros)
• exterior orientation interpolated for each
  scanline (Lagrange, linear)
• sensor interior orientation look direction
  table (laboratory calibration)
• LoS model improved by GCP

• projection of DEM to Cartesian coordinate system
  LTS (local topocentric)
• DEM resampling to image resolution
• connection of DEM and LoS vector by iterative
  procedure

• European projection
• 38 different National map projections and
• geodetic datum transformations

• cubic convolution

14
Rational polynomial functions for IRS-P6 LISS III

• Basic Equations
• Mapping from object space to image space
  (replacement of rigorous collinearity equations
  by a so-called Universal Sensor Model)
• RPC coefficients are estimated via a 3D grid of
  object space points with known image coordinates
  generated by a rigorous modeling of the imaging
  geometry
• Standard form for satellite imagery (e.g.
  QuickBird / IKONOS-2 / IRSP6)

r, c image space coordinates row and column ?,
?, h longitude, latitude and ellipsoidal
height in WGS84 geographic coordinates

• RPC correction by GCP
• The accuracy of RPC is reflecting the knowledge
  of camera parameters (interior orientation) and
  orbit/attitude (exterior orientation)
• IRSP6 accuracy is in the order of a few hundred
  meters thus RPC have to be corrected
• This is done with an additional affine
  transformation estimated via ground control
  points

15
Overall Geometric Accuracy
Requirement RMSE lt 20m
Overall (3700 scenes) mean accuracy w.r.t.
reference data set RMSEx/y 10 m ( CE64
14m) 0.5 pixel size of resampled images
Mean number of ICPs per scene for accuracy
assessment IRS-P6 5496 points / scene SPOT
4/5 1360 points / scene Residual plots
available
16
Comparison SPOT 4/5 and IRS-P6
Coverage 2
Coverage 1
IRS
IRS
SPOT
SPOT
consider oblique view for SPOT gt DEM
error about same orthorectification accuracy
for both sensors
17
Example European Mosaick First Coverage
European Projection LAEA-ETRS89 Resolution
25m 1500 scenes SPOTIRS
18
Example Germany
Projection ETRS89-LAEA Scope pan-European Ellips
oid GRS80 Datum ETRS89 Latitude of origin 52
north Longitude of origin 10 east FN
3210000 FE 4321000 Unit meter Prime meridian
Greenwich
radiometric atmospheric correction is not part
of Image2006 project
19
Quality Assessment

• Correctness QC includes checking if the
• GCPs are distributed equally over the scene
• residual plots derived from the ICPs show no
  systematic behaviour
• RMS errors fulfil the requirement to be better
  than 20m in each direction
• orthorectified image fits with the overlaid
  (re-projected) reference image
• orthorectified image fits with overlaid neighbour
  scenes
• Completeness QC includes checking if the
• image mosaic shows no gaps
• data set for a country is complete (country
  frontier polygons)
• Others include the
• visual inspection of radiometric quality and
  cloud coverage of the images
• visual inspection of used DEM tiles (gaps,
  artefacts) and used reference tiles (cloud
  coverage, radiometric quality, artefacts,
  geometric errors)

20
Automatic processing chain some experiences

• Metadata quality (measurements) of SPOT 4/5 and
  IRS P6 LISS III used for orthorectification is
  excellent.
• Good candidates for image matching (in order to
  extract GCPs and ICPs) are arid areas (Spain,
  Turkey,) or mountainous areas with time constant
  sharp ridges.
• The time gap of 5-6 years between images and
  reference data can cause problems because
  matching strongly depends on image similarity.
• Geometric errors in reference images cannot be
  handled with RPC or DG methods rigorously (see
  next slides)

21
Errors in Reference Dataset (1)
22
Errors in Reference Dataset (2)
23
Errors in Reference Dataset (3)
IMAGE2000 and USGS land cover datasets are
considered as absolute reference, which contain
systematic and local geometric distortions. ?
Models (DG and RPC) are not designed to handle
unrealistic errors rigorously
Probably wrong mosaicking of reference scenes or
DEM error
Good distribution of ICPs no systematic
behaviour
Residual plots Deviations in pixel (enlarged) of
ICPs from the reference image versus image
coordinates of orthorectified scene after
correction with GCPs
It can be assumed that an increased absolute
geometric accuracy is achieved for IMAGE2006
products (has to be verified using GCPs of
superior quality)
24
European GCP Chip Database
GCP chip naming convention ltEastinggt_ltNorthinggt_lt
EllipsoidHeightgt.png Example 4179645_2766680_385.
png Projection LAEA-ETRS89 Resolution 25 m Chip
size 101 x 101 pixel Channels Green, Red, NIR
Origin Image2006 data set of first coverage
and interpolated best of DEM Accuracy
10m w.r.t. reference data (as Image2006) Density
5 GCP chips _at_ 1000 km² Total amount of GCP
image chips EU38 61053 Extraction manually from
ICPs/GCPs
4179645_2766680_385.png
25
Overview IT-Infrastructure
10 workstations for automatic job- and process
assignment
GBit-Ethernet
Internet
2 servers (active and passive cluster node)
synchronized via heartbeat for automatic backup
and fallback
Firewall
Fibre-Channel-Net
2 Storage-Units (RAID 5) with 20 TByte (netto)
Administration system with Web-Interface (mySQL)
HW-/Process- Monitoring, Watchdog
SMS EMail
26

• The DLR in-house developed processing chain
  consists of
• a web based interface (mySQL) for administration
  of data and processes
• as well as GUI for QA tasks (check-in /
  check-out capability)
• an autonomous process queue to distribute jobs
  on the workstations
• which realizes parallel processing
• processing modules which are assembled to S/W
  processors
• using higher level script languages
• ? after import of scenes the processing chain is
  fully SW controlled till delivery

27
System Monitoring
Connectivity CPU Discs H/W Memory System
Messages Network Processes
28
Data throughput for one workstation in hours
The maximum throughput of the whole processing
system (10 desktop workstations, 10 operators)
including manual Quality Control is therefore
for one IRS-P6 scene about 0.5 h for one SPOT
scene about 0.2 h. About 3-5 of the images have
to be processed manually (additional GCP
measurements) in cases the product does not pass
the internal quality control.
29
Conclusion

• An operational and automatic processing chain to
  orthorectify satellite images from SPOT 4 HRVIR,
  SPOT 5 HRG and IRS-P6 LISS III covering EU25 and
  neighbouring countries has been developed within
  two months and successfully applied for the
  IMAGE2006 GMES Fast Track Land Service project.
• An overall geometric accuracy of about 10m RMSE
  (0.5 pixel size) in each direction w.r.t. the
  IMAGE2000 and USGS Land Cover reference datasets
  is achieved.
• About 3-5 of the images have to be re-processed
  (manual GCP measurements) in cases the product
  does not pass the internal quality control.

30
Similarity of images for image matching

• Number of automatically derived ICPs (and GCPs)
  strongly depends on the similarity of the images
  having a time gap of 5-6 years
• For example
• Agriculture areas show changes over the years
• Heavy floodings (Romania) changed the landscape

31
Tie point generation (GCP/ICP) by automated image
matching
on each level of image pyramid
32
Requirements - Statement of Work (Technical)

• From high resolution images of SPOT4 HRVIR, SPOT5
  HRG and IRS-P6 LISS III orthorectified images
  shall be produced using a DEM of accurate enough
  quality.
• Orthorectified products in European Projection
  (LAEA-ETRS89) with 25m resolution and National
  Projection for the countries with 20m resolution
  shall be generated in one resampling step using
  Cubic Convolution for the resampling.
• The IMAGE2000 panchromatic layers shall be
  considered as absolute reference. For countries
  not covered by IMAGE2000 the USGS ETM Land Cover
  dataset shall be used.
• Orthoimages with an overall accuracy of 20m RMSE
  in each direction w.r.t. the reference dataset
  shall be produced and internal quality control in
  order to generate quality records shall be
  performed.
• A consistent European wide GCP chip database with
  more than 15 GCPs for an area of 3600 km² shall
  be established.