Sliver Removal in Object-Based Change Detection
from VHR Satellite Images
Luigi Barazzetti
Abstract
This paper presents a novel strategy for object-based change
detection using very high spatial resolution (
VHR
) satellite images
captured under variable off-nadir view angles. The variable off-na-
dir angle, along with weak absolute orientation, generates spuri-
ous slivers during the multitemporal comparison of classification
results. The proposed solution for accurate object-to-object com-
parison is based on an intermediate registration of object-based
classification results with a piecewise affine transformation
followed by robust, geometry-based techniques for sliver removal.
Although different remote sensing applications require different
strategies and methods for object-based change detection, the ap-
proach developed in this paper can overcome the overall limita-
tion introduced by the slivers generated by weak geo-localization,
variable off-nadir angles, and image segmentation.
Introduction
Object-based Change Detection from VHR Images
Over the last decade, the availability of very high spatial
resolution satellite images has opened new opportunities in
the fields of photogrammetry and remote sensing (Rogan and
Chen, 2004). Nowadays, multispectral images with sub-meter
spatial resolution can be captured by numerous platforms
characterized by short revisit time (see Belward and Skøien,
2015). Metric accuracy has recently reached 30 cm with the
launch of WorldView-3 in August 2014. Although the metric
resolution is still not comparable with aerial images (
VVHR
,
i.e., very, very high resolution), the continuous acquisition
of
VHR
images is more straightforward than an aerial block
(Barazzetti
et al.
, 2014), which is expensive and requires a
flight plan. Some actual information on aerial cameras and
rules for image acquisition can be found in Kraus (2007).
Traditional pixel-based data processing (Lu
et al.
, 2004;
Radke
et al.
, 2005) is based on the use of the spectral informa-
tion encapsulated into the digital numbers of the image pix-
els. In the case of
VHR
images, pixels are significantly smaller
than real objects (buildings, tree crowns, roads, etc.) and data
processing may evolve from a basic pixel-to-pixel approach
toward a new object-to-object concept defined as Geographic
Object-Based Image Analysis (
GEOBIA
; Hay and Castilla, 2008).
GEOBIA
is based on segmentation of an image to generate
objects (i.e., groups of pixels that are consistent for geometry,
texture, or context). Objects have not only spectral informa-
tion but also spectral statistics (min-max values, mean, stan-
dard deviation, median, etc.), position information (distance
to, center coordinates, etc.) geometric characterization (area,
perimeter, width, height, elliptic fit, etc.), and relationships
with other objects (close to, far from, etc.).
GEOBIA
tools are
already available in several commercial software packages for
remote sensing image analysis, such as eCognition
®
(Definiens
Imaging GmbH, Munich, Germany), ENVI EX module (ITT
Visual Information Solutions, Colorado), ERDAS Imagine
®
Ob-
jective module (ERDAS, Inc., Norcross, Georgia), PCI Featu-
reObjeX (PCI Geomatics, Ontario, Canada), and IDRISI Selva
(Clark Labs, Massachusetts).
The advent of high spatial resolution remote-sensing im-
agery has also provided new opportunities for Object-Based
Change Detection (
OBCD
, see Löw
et al.
, 2015). In
OBCD
,
objects extracted and classified from satellite time series are
compared by using not only spectral information but also
object geometry (Löw
et al.
, 2015). OBIA has a strong connec-
tion to
GIS
data processing. However, one of the fundamental
challenges in
OBCD
concerns the lack of spatial correspon-
dence between objects detected in multi-temporal time series,
due to both geometric and spectral variability. In other words,
objects independently extracted from timed series can have
different boundaries (Blaschke, 2010) for the reasons illustrat-
ed in the next section.
Main Factors Affecting Object-to-Object Correspondence
Given two
VHR
satellite images, the main factors that affect ob-
ject-to-object correspondence are (a) the variable off-nadir angle,
(b) the overall geo-localization accuracy estimated with (or
without) a bias-compensated Rational Polynomial Coefficient
(
RPC
) camera model (Poli and Toutin, 2012), (c) the resolution of
the Digital Elevation Model (
DEM
) used in orthorectification, and
(d) the segmentation approach for object generation in OBIA.
Shown in Figure 1 is the effect of a variable off-nadir angle
in two QuickBird images. Large off-nadir angles reduce revisit
time, but images lack geometric consistency with terrain-cor-
rected satellite images and other geo-products (maps, spatial
databases, etc.; see Kapnias
et al.
, 2008). In Figure 1b, the
sub-vertical wall exhibits a non-constant spatial displacement
of about 9 m, whereas image spatial resolution is 0.6 m.
Overall geo-localization is another issue of primary impor-
tance in which map or geodetic coordinates are related to the
pixels of the image (Oh and Lee, 2015). Orientation parame-
ters of
VHR
satellite images are provided with a set of Ratio-
nal Polynomial Coefficients (
RPC
) derived by the image data
provider from the rigorous model using navigation data.
RPC
s
comprise 80 coefficients and allow sensor and camera model
data to remain unrevealed (Fraser
et al.
, 2006; Poli, 2007).
However, the direct use of
RPC
s for image orientation has
limitations in determining the true spatial orientation of every
scan line, errors within the direct measurement of sensor ori-
entation (especially attitude), and position and velocity (Fra-
ser and Hanley, 2003). Sensor orbit and data acquisition char-
acteristics lead to errors in direct geo-localization significantly
larger than ground resolution (ground sampling distance,
GSD
)
and require a correction as a bias in image space. The narrow
field-of-view of the satellite line scanner (approaching a paral-
lel projection for practical purposes) and the
Department of Architecture, Built Environment and Construc-
tion Engineering, Politecnico di Milano, via Ponzio 31, 20133
Milan, Italy (
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 2, February 2016, pp. 161–168.
0099-1112/16/161–168
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.2.161
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2016
161
