Section 7 Orthoimage generation and object extraction (mainly roads and buildings)

ISPRS Technical Commission I Symposium “From Sensors To Imagery”, Paris – Marne la Vallée, France, 3-6 July 2006

Institute of

Photogrammetry and

GeoInformation

Section 7

Orthoimage generation and object extraction (mainly roads and buildings)

Emmanuel Baltsavias


Institute of

Photogrammetry and

GeoInformation

Orthoimage Generation

Older sensor models methods:

• Kratky’s Polynomial Mapping Functions (PMFs)

• Relief corrected affine transformation

- Project GCPs on reference plane (X, Y coordinate corrections) using sensor elevation and azimuth

- Reference plane -> reference plane of DTM

- Compute affine transformation between X, Y object and x, y pixel coordinates

- 3 GCP’s are needed but 4-6 are suggested

Current method:

• Use RPCs and subsequent shifts or affine transformation


Institute of

Photogrammetry and

GeoInformation


Test resultsLuzern

Method: Relief corrected affine transformation

Sens elev. (deg) 67.7DTM spacing/accur (m) 25 / 2.5 lowland, 10 AlpsGCP accuracy (m) 0.5 – 3GCP definition Very poor to goodElevation range (m) 400-2100

Version GCPs/CPs RMS/X RMS/Y Max. abs X Max. abs Y

1 0 /66 134.2 30.6 501.5 118.1

2 6 / 65 2.6 2.2 9.9 5.9

5 6/14 0.6 0.6 1 1.1

CPs = check points


Institute of

Photogrammetry and

GeoInformation


Test results

Method: Relief corrected affine transformation

Nisyros

Version GCP's/CPs RMS/X RMS/Y Max. abs X Max. abs Y1 0 / 38 106.1 75.5 153.1 122.82 4/34 1.7 1 4.4 2.33 4/15 0.9 0.6 1.5 1.4

Sens elev. (deg) 73.5DTM spacing/accur (m) 2 / 3.3 GCP accuracy (m) ca. 0.5GCP definition Poor to goodElevation range (m) 0-700


Institute of

Photogrammetry and

GeoInformation

Orthoimage generation (IKONOS, Quickbird) in Geneva

Input data

2 IKONOS Geo images (IKONOS-West / IKONOS-East)1 Basic QUICKBIRD Image

Orthoimages (for acquisition of GCPs):OP-DIAE: Digital Orthos of Canton Geneva (25 cm pixel size, 0.5 m planimetric RMS)Swissimage: Digital Orthos of Switzerland of Swisstopo (50 cm pixel size, 1m planimetric RMS)

DTMs:DTM-AV (from airborne laser scanning): 1 m grid spacing, 0.5 m height RMSDHM25 of Swisstopo (from digitised contours): 25 m grid spacing, 1.5-2 m height RMS

Measurement of GCPs with ellipse fit and line intersection.

Image orientation with various sensor models. RPCs with subsequent affine transformation used for orthoimage generation.


Institute of

Photogrammetry and

GeoInformation



Institute of

Photogrammetry and

GeoInformation


Pansharpened orthoimages. Left Ikonos, right Quickbird.


Institute of

Photogrammetry and

GeoInformation


Definition of lines and circles. Left Ikonos, right Quickbird.

Note the large visual difference although pixel size is 1m and 0.7m respectively.


Institute of

Photogrammetry and

GeoInformation


Image Number of GCPs/CPs

X RMS (m) Y RMS (m) X mean with sign (m)

Y mean

with sign (m)

Ikonos West 10/23 0.55 0.63 0.25 -0.49

Ikonos East 10/33 0.47 0.76 0.10 -0.59

Quickbird 10/53 0.56 0.60 -0.08 -0.38

Planimetric accuracy of panchromatic orthos with GCPs from OP-DIAE

(CPs = check points)

Quickbird is not more accurate than Ikonos although GSD was 0.7m and 1m respectively.

Planimetric accuracy could be even higher with well defined GCPs measured with GPS.

In Y mean (bias) large due to coordinate system differences (Geneva and national systems differ).


Institute of

Photogrammetry and

GeoInformation


Image Number of GCPs/CPs

X RMS (m) Y RMS (m) X mean with sign (m)

Y mean

with sign (m)

Ikonos West 10/58 0.91 0.72 -0.07 -0.30

Ikonos

East

10/57 0.67 0.75 0.00 -0.33

Quickbird 10/93 0.66 0.77 -0.06 -0.11

Planimetric accuracy of panchromatic orthos with GCPs from OP-DIAE and Swissimage

Submeter accuracy even with GCPs from not so accurate Swissimage orthos.


Institute of

Photogrammetry and

GeoInformation

Road Extraction – Project ATOMI

• Automated Reconstruction of Topographic Objects from Aerial Images using Map Information.

• It’s a co-operation between swisstopo (Swiss Federal Office of Topography) and ETH Zurich, financed by swisstopo.

• ATOMI uses edge detection and existing knowledge and cues about road existence to detect road centrelines from orthophotos.

• ATOMI is used to remove cartographic generalisation and fit the geometry of roads to the real world to an accuracy of better than 1m in x, y and z

• ATOMI keeps the topology and attributes of the input vector map data set (VECTOR25)

• The result is a new accurate 3D road centreline data set without gaps, containing the topology and attributes of the input data as well as new weighted mean road width attributes


Institute of

Photogrammetry and

GeoInformation

ATOMI input and output data

3D accurate structured road vector data

2D inaccurate structured road vector data

DTM/DSM

Ortho images

ATOMIAutomatic

road centreline extraction


Institute of

Photogrammetry and

GeoInformation

Classification of roads according to landcover

foresturbanrural area

Current work only in rural areas


Institute of

Photogrammetry and

GeoInformation

General Strategy• Use of existing knowledge, rules and models

• Use and fusion of multiple cues about road existence

• Creation of redundancy through multiple cues to account for errors

• Early transition to object space, use of 2D and 3D interactions to bridge gaps and missing road parts

• Object-oriented approach in multiple objects (e.g. road classes)

+


Institute of

Photogrammetry and

GeoInformation

15

Features, Cues and Algorithms

2-D straight edges2-D road marks

zebra crossingsStereo color

aerial images / orthosFeature

Extraction

Geodatabase InformationDerivation

road geometryroad attributesroad topologylandcover

3-D straight edges

3-D road marksImage

Matching

road regionsshadowsvegetationbuildings

ImageClassification

above-ground and

ground objectsDSMnDSM

Generation

Input Features, cuesProcessing


Institute of

Photogrammetry and

GeoInformation

a. Straight edge

extraction

b. Removal of irrelevant

edges

c. Detection of Parallel

Road Sides

d. Evaluation of Missing

Road Sides

e. Bridging Gaps

f. Linking Road Sides to

Extract Roads

How ATOMI road extraction works


Institute of

Photogrammetry and

GeoInformation

Test site in Switzerland

Geneva 7km²Aerial Film 50cm (summer ‘98) IKONOS PSM 100cm (May ‘01)Quickbird PSM 70cm (July ‘03)Manually measured reference data from 50cm orthophotos


Institute of

Photogrammetry and

GeoInformation

50cm Aerial Film 100cm IKONOS 70cm QUICKBIRD

Results from Geneva (yellow VECTOR25, black result)


Institute of

Photogrammetry and

GeoInformation

50cm Film 100cm IKONOS 70cm QUICKBIRD


Institute of

Photogrammetry and

GeoInformation

Quality evaluation of the results of Geneva


Institute of

Photogrammetry and

GeoInformation

Results from Geneva

• The system achieved good results with the 50cm aerial film imagery with 90% of rural roads extracted.

• The performance (mainly the completeness) of the satellite data was inferior to aerial imagery, especially the 1m IKONOS imagery

• In the satellite data, higher class (wider) roads were usually extracted, while most lower class (narrower) roads were not. This is because of ATOMI’s algorithm requirement of min. 3 pixels road widths was not fulfilled.

• The smaller GSD of Quickbird made more roads visible and the road surface and road edges were clearer. But compared to aerial film the completeness was lower.

• With smaller GSD, correctness and accuracy do not deteriorate much, but completeness yes.

• Other road extraction methods and results with aerial and Ikonos images as part of EuroSDR working group on road extraction

• See paper Mayer et al. in coming ISPRS Com. III Symposium, Bonn, September 2006


Institute of

Photogrammetry and

GeoInformation

Building Extraction, Ikonos, Melbourne

Roof corners

• 19 roof corners measured by GPS

• Measured in mono and stereo in all three images of Melbourne

Results from stereo images and 6 GCPs (RMSE):

RPCs: XY = 0.7m Z = 0.9m


Institute of

Photogrammetry and

GeoInformation

Building Extraction

Aerial Photography (1:15,000) Ikonos 1m Pan Stereo

Omission of 15% of buildings (small & large)


Institute of

Photogrammetry and

GeoInformation

Building Extraction

3D Model of University of Melbourne Campus from Ikonos 1m PAN Stereo

Produced with CyberCity Modeler


Institute of

Photogrammetry and

GeoInformation

Building Extraction

Aerial Photography (1:15,000) Ikonos 1m Stereo Imagery


Institute of

Photogrammetry and

GeoInformation

Building Extraction

Aerial Photography (1:15,000) Ikonos Stereo Ikonos Nadir Pan-Sharp.

Conducive to building feature measurement


Institute of

Photogrammetry and

GeoInformation

Building Extraction

Aerial Photography (1:15,000) Ikonos Stereo Ikonos Nadir Pan-Sharp.

Ikonos stereo of questionable value to building feature measurement in this case


Institute of

Photogrammetry and

GeoInformation

Building Extraction

Aerial Photography (1:15,000)

Ikonos Stereo Ikonos Nadir Pan-Sharp.

Ikonos stereo of questionable value to building feature measurement in this case


Institute of

Photogrammetry and

GeoInformation

DEM

(X,Y,Z)

Image plane, (line,sample)

Sensor model

DEM

Initial Z from DEM

Line, Sample

XY Solver

Interpolating Z

Convergence

XYZ

Z

No

Yes

Monoplotting: 3D Information Extraction from Single Satellite Images(implemented In Barista software, C. Fraser, Univ. of Melbourne)


Institute of

Photogrammetry and

GeoInformation

3D mapping of points and linear features by monoplotting

Measure single points in monoplotting mode using one image in combination with a DEM

Measure polylines/closed polylinein monoplotting mode

3D mapping of pointsand linear features


Institute of

Photogrammetry and

GeoInformation

Measure groundpoint using one image in combination with a DEM

XYZ

Measure roofpoint Z from line,sample and

XY

Height ΔH

Height measurement of buildings by monoplotting


Institute of

Photogrammetry and

GeoInformation

3D mapping of buildings from single IKONOS image

Measure first ground point of thebuilding (by monoplotting)

Measure first roof point, assuming thesame XY-position as ground point

Force same height for other roof points and find XYGet other ground points bykeeping XY-position and interpolateheight from underlying DEM

>>> 3D building reconstruction


Institute of

Photogrammetry and

GeoInformation

Building reconstruction from single IKONOS image

Export of buildings in VRML dataData collection with Barista


Institute of

Photogrammetry and

GeoInformation

Quality assessment of results derived from IKONOS and Quickbird imagery

• GCPs and multi-image data used as reference

• Single image point positions and building heights were measured

• Bias-corrected RPCs and affine sensor model was used

• Measurements were performed in three IKONOS and two Quickbird images

• Regular monoplotting mainly depending on DEM quality

• Single-image height measurement affected significantly by off-nadir angle


Institute of

Photogrammetry and

GeoInformation

Quality assessment of results derived from Ikonos and Quickbird

Sensor orientation

model

Height RMS discrepancy at CPs (m)

e = 83° e = 61° e = 61°

RPCs 6.59 1.08 1.31

No. of CPs 20 23 22

3D-Affine 3.05 0.69 0.76

No. of CPs 20 24 22

Accuracy of Building Height Determination

Sensor orientation

model

Height RMS discrepancy at CPs (m)

e = 60° e = 58°

RPCs 1.01 1.03

No. of CPs 30 30

3D-Affine 1.34 1.33

No. of CPs 29 29

Accuracy gets worse with higher elevation

Quickbird Ikonos


Institute of

Photogrammetry and

GeoInformation

3D site modeling (Silvereye from Geotango, CA)

Semi-automatic system, using one oriented image (monoplotting) and additional information (similar to Barista). Software for 3D site modeling and orthoimage generation. Not available any more (company bought by Microsoft)


Institute of

Photogrammetry and

GeoInformation

3D city modeling (Cybercity AG)

Quickbird stereo images, Phoenix, USA


Institute of

Photogrammetry and

GeoInformation

WEB-based 3D earth visualisation and Location Based Services

• Google Keyhole (Quickbird)

• Similar devepments by Microsoft (Orbimage)

• Geotango‘s (CA) Globeview (company bought by Microsoft)

• Use of satellite images, espec. HR

• DSM from usually unknown sources

• 3D building models

• Driving directions

• Location of various services

Documents

Section 7 Orthoimage generation and object extraction (mainly roads and buildings)