Roof-Cut Guided Localization for Building Change
Detection from Imagery and Footprint Map
Jinqi Gong, Xiangyun Hu, Shiyan Pang, and Yujun Wei
Abstract
Identification and monitoring of buildings are of considerable
practical value in three-dimensional (3D) reconstruction of
building models and urbanization monitoring. Especially for
the change detection of buildings with composite structures
and relief displacements, heterogeneous appearance and
positional inconsistencies are two
work, a novel roof-cut approach is
based model to locate rooftops an
through the use of imagery and pr
maps. The building region, bound
constraint terms were first formulated by multiple cues de-
rived from both data sources. Next, roof-cut segmentation was
performed by gathering all terms required for high-quality un-
supervised rooftop extraction. Finally, the positional displace-
ment statistics of similar adjacent buildings were collected to
accurately estimate the rooftop location and achieve building
demolition detection with the overlap ratio index. Experimen-
tal results indicated the effectiveness and generality of the
proposed roof-cut algorithm for aerial and satellite images.
Introduction
Retroreflectivity
Automatic building change detection from satellite and aerial
images is a relevant research area in the remote sensing field,
as the results are required for a range of applications such as
urbanization monitoring, identification of illegal or unauthor-
ized buildings, land use change detection, and digital map
updating (Akçay and Aksoy 2010). Similarly, information
about the change of buildings can be useful for aiding mu-
nicipalities with long-term residential area planning and the
analysis on the condition of damaged buildings after natural
disasters, supporting rescue activities and reconstruction
measures (Sofina and Ehlers 2017).
Conventional methods proceed with effective selection
of discriminative features according to the defined criteria
of buildings and comparison of features to achieve change
detection from remotely sensed images of the same scene
obtained at different times (Tiede
et al.
2011; Taskin Kaya
et al.
2011; Voigt
et al.
2011). Image-image comparison vary
between pixel-oriented methods and object-oriented meth-
ods, and between spectral characteristics-based methods and
artificial intelligence-based methods (Bouziani, Goïta, and He
2010). Pixel-oriented methods mainly use techniques based
on algebraic operations (Singh 1989), transformation (Coppin
and Bauer 1996) and classification (Lunetta and Elvidge 1999)
to recognize change information. However, when applied to
high-resolution remotely sensed images, there may appear
to be a large amount of small pseudo changes because of the
increased high-frequency components (Chen
et al.
2012). Fur-
thermore, pixel-based methods strongly depend on geometric
tric correction (Hussain
et al.
2013),
curacies and are commonly used
from low- or medium-resolution
feld, and Menz 2014). Object-
ng from the concept of object-based
image analysis (Thomas
et al.
2014) can not only employ the
spectral, texture, and transformed values, but also exploit
extra information about the shape features and spatial rela-
tions of objects by image segmentation techniques. As image
objects are used as the basic units in object-oriented methods,
they are more suitable for handling high-resolution remotely
sensed images and can achieve better performance (Myint
et al.
2011; Sellaouti
et al.
2013; Hussain
et al.
2013; Xiao
et
al.
2017). However, it is difficult to obtain a globally optimal
segmentation, thus geometric inconsistence is usually un-
avoidable. In addition, due to the complexity and diversity of
remotely sensed images, variable characteristics of buildings
with potentially infinite spatial layouts is hardly described
and the accuracy of the change detection suffers when only
image data are used.
Many modern approaches have focused on the integration
of high-resolution remotely sensed image and geographic
information system (
GIS
) technologies. The
GIS
data in these
studies were applied to select training areas for image classifi-
cation, finding differences for change detection, and providing
initial approximations for object detection and boundary cues
for three-dimensional (
3D
) reconstruction (Jolly and Gupta
2000; Vosselman 2008; Durieux, Lagabrielle, and Nelson
2008). Suveg and Vosselman (2004) integrate the aerial image
analysis with information from large-scale two-dimensional
GIS
databases and domain knowledge to get the possible loca-
tions of the building. Chesnel, Binet, and Wald (2008) used
ancillary data that agreed with the reference image, and auto-
matically searched for the damaged buildings in the postcrisis
image using the correlation between the homologous pixels
inside the building roof outline. For the updating of geoda-
tabases in urban environment, Mourad, Goïta, and He (2010)
proposed an object-oriented approach allowing the analysis
of the objects which are characterized by different attributes
to detect various types of building change. In this process, the
existing knowledge was used to improve image processing
and change detection. Automatic detection and classification
of damaged buildings by integrating high-resolution satellite
images and vector maps was proposed by Samadzadegan and
Rastiveisi (2008). This approach located buildings from vector
Jinqi Gong, Xiangyun Hu, and Yujun Wei are with the School
of Remote Sensing and Information Engineering, Wuhan
University, Wuhan 430079, China (
).
Xiangyun Hu and Shiyan Pang are with the Collaborative
Innovation Center of Geospatial Technology, Wuhan
University, Wuhan 430079, China.
Shiyan Pang is with the School of Resource and Environmen-
tal Sciences, Wuhan University, Wuhan 430079, China.
Photogrammetric Engineering & Remote Sensing
Vol. 85, No. 8, August 2019, pp. 543–558.
0099-1112/19/543–558
© 2019 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.85.8.543
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2019
543
