ESAEUSC2006 Image Information Mining

1
ESA-EUSC2006 Image Information Mining

• Automatic Heritage Observation and Retrieval
  Under Sand Using a Shape Detection Algorithm with
  Satellite Images
• Ian Bridgwood (Rovsing A/S)
• Djenan Ganic
• Michael Schultz Rasmussen (GRAS A/S)
• Mikael Kamp Sørensen (GRAS A/S)
• Renzo Carlucci (Sogesi Aps)
• Fabrizio Bernardini(Sogesi Aps)
• Ahmed Osman(Centre for Documentation of Cultural
  and Natural Heritage of Egypt)

2
Contents

• Introduction
• Implementation
• Results
• Conclusion

3
Introduction

• An application is presented for the automatic
  identification of lost or undiscovered
  archaeological sites in Egypt using a shape
  detection technique on satellite Earth
  observation (EO) imagery superimposed in a
  geographic information system (GIS) environment.
• A shape detection algorithm that employs an
  operator matched to the shape of buried
  archaeological structures is described.
• The implementation is based upon the optimal
  edge-based shape detection approach described by
  1. The algorithm is applied to EO images and
  the results are presented.
• Moon H., Chellapa R. and Rosenfield A., Optimal
  Edge-Based Shape Detection, IEEE Transactions on
  Image Processing, Vol. 11, No. 11, November 2002.

4
Heritage Observation and Retrieval Under Sand

• The project has been performed under the ESA Data
  User Element (DUE) Innovator User Partnership
  project Heritage Observation and Retrieval under
  Sand, (HORUS).
• The main objective of the Innovator User
  partnership is to demonstrate an innovative Earth
  Observation (EO) based information service to
  End-users.

5
Heritage Observation and Retrieval Under Sand

• Horus was the Egyptian god of the sun and moon.
• Horus uncle, Set gouged out one of Horus eyes,
  thus explaining why the moon isn't as bright as
  the sun.

6
Innovative End-user Service

• The End-user is CULTNAT, the Centre for
  Documentation of Cultural and Natural Heritage of
  Egypt.
• Current resources and practices do not allow the
  analysis and control of all archaeological areas
  in Egypt.
• An increasing amount of time and expense must be
  spent in detecting new archaeological sites, in
  protecting actual sites and in discovering
  possible illegal excavation sites.
• Innovative service
• Identification of lost or undiscovered
  archaeological structures by
• combining the historical information and
  knowledge of archaeological structures from the
  End-user with automatic procedures for searching
  EO data and integrating the results in the
  End-users geographic information system (GIS)
  environment

7
End-user Scenarios

• The following scenarios envisaged by the End-user
  include
• An imprecise historical map indicating the shape
  of the archaeological structures exists. The
  location of the structures is known to a good
  approximation i.e. the same order of magnitude as
  the location spread of the objects.
• A precise historical map indicating the shape of
  the archaeological structures exists. The
  location of the structures is known to a poor
  approximation i.e. several orders of magnitude
  greater than the location spread of the objects.
• No map is available of possible structures but
  knowledge exists that in a given location (with
  area dimensions depending on the size of the
  objects targeted by the search) objects may be
  present underground.

8
Process Description

• The process to provide the End-user service
  consists of
• EO image acquisition and pre-processing
• Archive search for historical maps and drawings
  of archaeological structures of interest
• Construction of shape matched operator and 2D
  shape detection
• Integration into GIS framework

9
EO Images

• SIR-C/X-SAR (STS59/64) project previously
  demonstrated that C and L band radar has
  effective ground penetrating ability in dry sand.
• Consequently, the C band synthetic aperture radar
  (SAR) images of the areas of interest generated
  by ERS1, ERS2 and Envisat have been used along
  with the L band SAR images generated by JERS1.
• Optical data is also considered as sub surface
  structures can cause changes in topography, sand
  accumulation, humidity and other features. In
  some cases, very high resolution (VHR) optical
  data can also be used to verify the results
  obtained and QuickBird images have therefore also
  been considered.
• Before performing shape detection, the EO images
  are calibrated, ortho-rectified and
  geo-referenced.

10
End-user Input

• Serapeum complex at Saqqara

Mariette Plan 1882
Picard-Lauer Excavations
11
2D Shape Detection

• Define a 2D shape from archaeological records
• Construct an optimal edge operator for the
  specific shape.
• Apply the operator to detect the shape in the EO
  images.

12
2D Shape Detection
13
Operator Preparation

• The 2D shape to be detected is supplied by the
  End-user and the shape boundary cross sections
  can readily be modelled as ideal step functions
  and supplied to the application as a monochrome
  TIF file.
• To define the optimal step edge operator,
  consider that the areas of interest are desert
  regions and the EO images typically have low
  contrast, also in boundary regions.
• In their paper, Optimal Edge-Based Shape
  Detection, (IEEE Transactions on Image
  Processing, Vol. 11, No. 11, November 2002), Moon
  H., Chellapa R. and Rosenfield A, define an
  optimal step edge operator. This application is
  based upon their approach.

14
Optimal Edge Operator

• Edge detection is essentially finding a high
  intensity gradient.
• Ideally, the differential operator should be the
  optimal edge edge operator.
• The derivative of the optimal smoothing operator
  should be the optimal edge operator, according to
  the differential rule
• (h I) h I

15
Optimal Edge Operator

• The 1D optimal step edge operator, which
  minimizes both the noise power and the mean
  squared error between the input and the filter
  output, was derived to be the derivative of the
  double exponential, (DODE).
• An operator for shape detection was defined by
  extending the DODE filter along the shape
  boundary.

16
Optimal Edge Operator

• The optimal edge operator is the piecewise
  derivative h of the smoothing operator and using
  the differentiation rule
• (h I) h I

Optimal edge operator
17
Shape Matched Operator

• The ideal step edges in the 2D shape file are
  extended according to the optimal edge operator

• Edge operator is the second derivative of the
  smoothed edge profile.

18
Shape Detection

• The optimal edge operator for the specific shape
  is convolved with each of the RGB layers from the
  EO TIF images.
• The results are combined to produce a detection
  result image file indicating the most probable
  location of the shape.

19
Test Results

• Initial simple shape recognition performance
  tests

20
Test Results Detection in Optical EO image
21
Test Results Detection in Optical EO image
22
Test Results JERS1 L Band Radar Image

• Serapeum Complex at Saqqara

23
Test Results JERS1 L Band Radar Image
24
Integration into GIS

• The results of the shape detection algorithm will
  then be combined with existing GIS layers of the
  area of interest.

25

• Just add water

26
Alternative Test

• Green fields in Denmark

27
Performance

• The application has been implemented in C and
  can be optimized for performance.
• A SAR image measuring 250 by 250 pixels and a
  shape image of 25 by 25 pixels is processed in
  seconds on a 1.83 GHz CPU with 2 GB RAM.
• Correspondingly, for a 5000 by 5000 pixel image
  and a shape of 100 by 100 pixels the processing
  time is in excess of 3 hours.

28
Conclusion

• An innovative EO information based End-user
  service has been implemented for the detection of
  archaeological sites, combing historical data and
  EO data and within a GIS environment
• The reliable detection of 2D shapes has been
  implemented using the aproach derived by Hankyu
  et al. using the DODE optimal edge operator.

29

• Thank you