PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2014
161
Multi-Agent Recognition System based on
Object Based Image Analysis Using WorldView-2
Fatemeh Tabib Mahmoudi, Farhad Samadzadegan, and Peter Reinartz
Abstract
In this paper, using spatial and spectral characteristics of
the WorldView-2 satellite imagery, capabilities of multi-agent
systems are used for solving multiple object recognition
difficulties in complex urban areas. The methodology has two
main steps: object based image analysis (
OBIA
) and multi-
agent object recognition. In the first step, segmentation and
multi-process object classification based on spectral, textural,
and structural features are performed. Classified regions are
used as an input dataset in the multi-agent system in order
to modify object recognition results. According to the results
from the object based image analysis process, using contex-
tual relations and structural features, the overall accuracy
and Kappa improved by 17.79 percent and 0.253, respec-
tively. Using knowledge-based reasoning and cooperative
capabilities of agents in the multi-agent system in this paper,
most of the remaining difficulties are decreased and values
90.95 percent and 0.876 are obtained for the overall accuracy
and Kappa, respectively, of the object recognition results.
Introduction
Recent advances in airborne and spaceborne sensor tech-
nology and digital imaging techniques has lead to very high
resolution (
VHR
) remotely sensed data that has reached the
level of detail of classical aerial photography. New sensors
and
VHR
images may provide a wealth of information for auto-
matic recognition of objects in urban areas in order to reduce
time and labor intensive tasks associated with field surveying
or manual digitizing. Hence, one of the objectives of deploy-
ing the advanced commercial remote sensing sensors was to
increase the visibility of urban objects by reducing per-pixel
spectral heterogeneity, and thereby improving the land-cover
identification (Myint
et al
., 2011). But, the need for highly ac-
curate and regularly updated geo-spatial information cannot
be met only by the advancements of sensor technology.
Because of the complex nature and diverse composition
of land-cover types found within the urban environment,
the production of urban land-cover maps from very high
resolution satellite imagery is a difficult task (Shackelford
and Davis, 2003). This is because many small objects are
concentrated in a small area, and they become more visible
as the spatial resolution gets finer. This situation potentially
leads to lower accuracy in pixel-based image classification
approaches which only use spectral characteristics of input
remotely sensed data. In other words, spectral heterogene-
ity severely limits the application of traditional pixel-based
approaches for accurate urban object recognition and classifi-
cation (Blaschke, 2010; Meng
et al
., 2012; Myint
et al
., 2011;
Shackelford and Davis, 2003; Zhou and Troy, 2008). However,
many researchers have investigated the potential of the object
based image analysis approaches for dealing with very high
resolution images and complexities in urban areas (Blaschke,
2010; Lewinski and Zaremski, 2004; Jacquin
et al
., 2008;
Maclean and Congalton, 2011; Myint
et al
., 2011; Peets and
Etzion, 2010; Zhou and Troy, 2008; Zhou and Wang, 2008).
The object based image analysis approaches first seg-
ment imagery into small objects, which provides the object
primitives for image analysis. Image segmentation decreases
the level of detail and reduces image content complexities.
Therefore, segmentation is an efficient mean of aggregation
the high level of detail and producing usable objects. Each
segmented region has additional information compared to
single pixels (e.g., mean values per band, etc); but of even
greater advantage is the additional spatial information for
objects (Blaschke, 2010).
Once these image objects are derived, topological relation-
ships between them, statistical summaries of spectral and
textural values, and shape characteristics (e.g., area, perime-
ter, rectangularity, etc) can all be employed in the rule-based
classification procedures (Blaschke, 2010; Platt and Rapoza,
2008; Zhou and Troy, 2008).
Literature Review
The idea of object based image analysis has been around since
the early 1970s. In many research studies, various segmenta-
tion algorithms such as the multi-resolution and the fractal
net evolution followed by a rule-based classification scheme
are performed on high-resolution remotely sensed data such
as Spot4, Spot5, Ikonos, QuickBird and aerial photographs
(Ivits
et al
., 2005; Myint
et al
., 2011; Platt and Rapoza, 2008;
Zhou and Troy, 2008).
The accuracy of object based image analysis results in
urban areas directly depends on the segmentation and rule-
based classification processes. Segmentation quality is one
of the most crucial aspects in object based image analysis
(Clinton
et al
., 2010; Marpu
et al
., 2010; Meinel and Neubert,
2004). It is commonly believed that there is no universal
segmentation method that can be successfully used for all
types of land-cover objects. Therefore, over-segmentation
(when the real world object is segmented into smaller sub-ob-
jects) and under-segmentation (when the real world object
or parts of it becomes a part of another object) may occur in
the segmentation results (Marpu
et al
., 2010). The selection
Fatemeh Tabib Mahmoudi and Farhad Samadzadegan are
with the Department of Geomatics, University of Tehran, P.O
Box: 11365-4563, Tehran, Iran (
.
Peter Reinartz is with the Department of Photogrammetry
and Image Analysis, Remote Sensing Technology Institute,
German Aerospace Centre (DLR), Oberpfaffenhofen 82234
Weßling, Germany.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 2, February 2014, pp. 161–170.
0099-1112/14/8002–161
© 2013 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.2.161
