International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
both in principal purpose and data characteristics. Such video
surveys tend to be (i) unplanned in terms of data acquisition
scheme, (ii) focus on highly damaged areas at the expense of
complete coverage, and (iii) produce uncalibrated data with
comparatively low spatial resolution. Commonly used devises
used are analogue or digital BetaCams (360 and 720 vertical
lines, respectively) or HDTV (High Definition Television) with
1920 columns. All PAL-compatible systems have 576 lines. In
addition to the news media, law enforcement agencies are also
increasingly using rapid airborne surveys following urban
disasters, where also more sophisticated cameras that acquire
infrared or thermal imagery are used.
While the acquisition of video data is straightforward, the use of
such data poses challenges, in that established image analysis
methods can only be used within limits. Reasons for that
include the oblique nature of the data, an unstable imaging
platform, frequent changes in focal length and, consequently,
image scale during data acquisition. In addition they frequently
lack camera orientation and location information.
The 1995 Kobe earthquake led to efforts to use non-calibrated
data for damage assessment, partially fuelled by delayed
overpasses of satellites following the event (Landsat TM: 7
days; JERS: 20 days; Yamazaki, 2001). Improving on previous
work by Hasegawa et al. (2000b) that only used visual image
interpretation, as well as on analysis of multi-temporal satellite
data on grounds of impracticality and data unavailability,
Hasegawa et al. (2000a) explored quantitative analysis methods
applied to noncalibrated imagery. Training data of areas
showing 3 levels of damage were extracted from individual
HDTV frames and used for a classification. Similar methods
were applied by Mitomi et a/. (2000) on earthquakes in Taiwan
and Turkey, and by Yamazaki (2001) on the 2001 Gujarat
disaster. All studies applied threshold values based on training
data to identify damaged areas with varying success, and all
were applied to individual video frames only, limiting the
practical value of the approach.
1.3 Research aims
A number of issues are addressed in this paper: (1) given the
limited resolution and, therefore, detail of standard video data,
compounded by further quality loss resulting from data format
conversions, we investigated the potential to improve image
quality (in terms of higher signal/noise ratio), using a stacking
method that incorporates adjacent frames, as well as with a
synthetic aperture approach similar to the one described by
Gornyi and Latypov (2002); (i) we applied a damage
assessment method to a video stream instead of individual
frames based on damage training area characteristics; (iii) we
used imagery with encoded GPS data, and relative camera
azimuth and inclination information. We extracted this auxiliary
information automatically to aid in the mapping of the
flightpath as well as spatial registration of the damaged areas.
The overall aim of the project reported on here is to provide a
video processing environment that combines the above
elements, and that allows near-real time processing of video
data to detect damage. Results can be displayed related to
geocoded pre-event and auxiliary data, in 2 or 3D as applicable.
An important difference to previous work is that our study deals
with an industrial accident marked by a radial damage pattern
quite different from earthquake damage. Further improvements
to our processing environment are planned to make it flexible
581
enough to deal with any type of urban disaster damage. The
event addressed in this study is briefly described below.
1.4 The 2000 Enschede fireworks disaster
On 13 May 2000, a series of explosions occurred at a fireworks
factory located within a residential area in Enschede, the
Netherlands. An area in excess of 40 ha was affected, 22 people
killed, and close to 500 buildings severely damaged or
destroyed. The force of the explosions led to complete
obliteration of buildings close to the site of the factory, while
damage severity rapidly declined towards the furthest affected
structures, approximately 1 km from ground zero (van Westen
and Hofstee, 2001; Figure 1).
Video image acquisition by the police began approximately two
hours after the disaster, and was repeated on the following days
and augmented by high-resolution vertical aerial photographs
12 days later. These data, however, were not used for the actual
damage assessment, which was instead based on ground-based
surveys. Disaster response was hindered by outdated map
material. Incidentally, a planned aerial survey of Enschede at a
scale of 1:18,000 was carried out just 4 hours before the disaster
occurred, providing reference data that could have been, but
were not, used in the disaster response phase.
RTE
"Enschede!
AR
ft % 4
~ Netherlands +
: i
| Belgium, i Germany:
Figure 1. Illustration showing (a) location of Enschede, (b) an
overview map of the disaster site, and (c) a photo taken during
the explosion of the fireworks factory.
2. DATA AND METHODOLOGY
2.1 Data used
A comprehensive database related to the disaster was compiled,
comprising of pre- and post-event, space- and airborne, still and
video imagery, in addition to pre- and post-disaster vector data.
The principal data sources used for this paper are optical video
data acquired by the Dutch National Police Aviation Branch on
May 13 and 15, as well as a pre-event Ikonos MS reference
imagery acquired on 3 April 2000. Video data were captured
with a Sony-HAD camera with 720x576 pixels, as well as an
Agema Thermovision 1000 that acquired data between 8-12
um.
2.2 Methodology
The following principal steps were carried out in this study:
(1) Investigation of frame quality enhancement
using AstroStack (www.innostack.com) and a
