methods. After introducing the depth-searching process, these
errors were magnified in the
LBLW
&DS method, as shown in
Plate 2f. In the two experimental areas, the LW method retrieved
the minimum road regions and formed a broken road network.
The
LBLW
and
LBLW
&DS methods obtained significantly denser
road regions and formed continuous road networks while in-
ducing additional errors. Visual interpretation was completely
coherent for the aforementioned quantitative analyses.
An enlarged section of experimental area 1 is shown in Plate
1g to 1i. In Plate 1g, a road intersection (region 2), parts of a
horizontal road (regions 3 to 6), and an arc-formed road (region
1) were not extracted as roads because their
LW
s were less than
3.0. In Plate 1h, which illustrates the
LBLW
method, straight
lines overlap onto regions. The
LBLW
of region 1 was larger than
3.0 based on the straight lines fitted on the edge lines. Similar-
ly, the
LBLW
of region 6 was larger than 3.0, which was calcu-
lated based on the below tangent straight line. Thus, regions 1
and 6 were classified as roads by the
LBLW
method. However,
road regions 3, 4, and 5 were not extracted as roads by the
LBLW
method because neither their
LW
nor their
LBLW
reached 3.0.
Such defects of a broken road were solved by the depth-
searching process. For example, region 7, which was extracted
as road by the
LBLW
method, acted as the starting point of depth
searching for neighboring road regions in the
LBLW
&DS method.
First, a tangent straight line (A) of region 7 was fetched (region-
to-line conversion). This line served as the direction guide for
the next possible road regions, and thus, regions 3 and 4 were
found. Region 3 did not share a common boundary with region
7 because of some segmentation errors, which would fail in
common neighborhood searching. However, the depth-search-
ing process successfully overrode the gap and fetched regions
3 and 4 (line-to-region conversion) because these regions were
located on the same negative side of straight line A and were
IPSL
-neighbors. These regions were accepted as members of
the road network because they were classified as impervious
surface in the first-level classification and had suitable
LBLW
s.
By contrast, neighboring parallel lines B and C were
fetched from line A (line-to-line conversion). Based on the
IPSL
-neighborhood relationships with respect to lines B and C,
the impervious surface region 5 was retrieved and extracted as
road (line-to-region conversion). The depth-searching process,
which involved the aforementioned conversions between
regions and lines, were iteratively performed from the left and
right sides of region 7 along the straight line direction until no
IPSL
-neighbor was found, which resulted in a complete road.
Region 2 was also extracted as a road by another depth-search-
ing process that was conducted from the vertical direction. In
the experiments, straight lines played at least two important
roles. (a) Region shape analysis became more accurate than
that in common region only-based
OBIA
, and (b) The line-
based, depth-searching process offered precise directions and
the capability to override gaps. These advantages exhibited the
feasibility and superiority of the proposed technical schemes.
However, the road network obtained by the proposed meth-
od remained imperfect. For example, defects such as broken
roads and burrs, as shown in Plates 1f and 2f, were commonly
found in urban areas with complex backgrounds. These defects
were comprehensively caused by possible segmentation and
classification errors, as well as the simplicity of current road-
extraction rules. Sophisticated
OBIA
rules or processing are
necessary to improve the accuracy of road extraction, particu-
larly in urban areas. Nevertheless, the proposed scheme can
serve as an initial step for road extraction because of its sim-
plicity and effectiveness. A more precise road extraction tech-
nique may be achieved based on this scheme and by including
additional optimization steps. In particular, a set of continu-
ous road regions may form a homogeneous
IPSL-H
chain. Short
chains can be trimmed by calculating the chain length because
roads are generally long. Furthermore, by checking whether
the nodes of a chain touch other roads, dangle chains can also
be trimmed because roads are generally connected. In addition,
the
IPSL-H
chain may also be used to smoothen road regions
and remove blurs because roads generally have a fixed width,
and thus, a set of regions provides more optical clues than a
single region when estimating road width. In addition to these
extensions, factors that include occlusions by other ground
objects (e.g., high and large buildings, trees, and shadows)
and highly complex road shapes (e.g., multiple overpassed
road intersections), may also be considered to improve road
extraction. Considering that this study mainly focused on
validating
RLPAF
, particularly several of its derived concepts
and operators, complex spatial constraints and rules were not
investigated comprehensively. The extracted road network
still exhibited many defects and might only serve as an initial
extracted result for subsequent refinement. In the future, the
aforementioned spatial relationships should be considered for
a mature
OBIA
-based road extraction system.
Conclusions
In this study, we propose
RLPAF
to extract information from
remote sensing images. This framework comprehensively
utilizes line and region primitives in
OBIA
by image segmen-
tation, straight line detection, and region-line relationship
modeling. The proposed framework is then applied and
validated in extracting road networks from
HSR
images. In this
framework, regions and lines are closely integrated through-
out the entire
OBIA
process. During image segmentation, lines
(edges) are first embedded into sub-region merging as spatial
constraints, which outputs regions with precise boundaries.
During feature extraction, the direction and topology features
of regions and lines are used to build the association model.
Then, several newly formed region-line associated features
are derived. During the information extraction stage, region
and line primitives are collaboratively used for rule-based
object discrimination. Such a highly systematic technical
framework of line and region integration has not yet been
reported in state-of-the-art
OBIA
studies and software systems,
and thus, is the main contribution of our study.
In the road extraction task,
RLPAF
improves region shape
analysis and spatial relationship reasoning. It can also be
utilized in many other forms. For example, regions may be
classified using the “line-based texture” by calculating the
orientation, length, and density of their intersecting straight
lines. In addition,
IPSL
-neighborhood relationships may screen
out highly homogeneous neighbors in context-based image
classifications. Extending the applications of the proposed
framework and methods, as well as optimizing their perfor-
mance, will be the focus of our future investigation.
Acknowledgments
This work is jointly supported by the National Natural Sci-
ence Foundation of China (41171321), the Natural Science
Foundation of Jiangsu Province, China (BK20140042), the
Natural Science Foundation of the Jiangsu Higher Education
Institutions of China (11KJA420001), the Jiangsu Surveying,
Mapping and Geoinformation Project (JSCHKY201503), the
Qing Lan Project, and the Priority Academic Program Devel-
opment of Jiangsu Higher Education Institutions.
References
Anders, N.S., A.C. Seijmonsbergen, and W. Bouten, 2011.
Segmentation optimization and stratified object-based analysis
for semi-automated geomorphological mapping,
Remote Sensing
of Environment
, 115(12):2976–2985.
158
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
