object size or segmentation scale. As demonstrated in Figure
8, when over-segmentation happens, the image objects are
generally of small sizes and accompanied by high intraseg-
ment homogeneity and low intersegment heterogeneity; when
the scale is suitable, the image objects can optimally represent
the geo-objects and are accompanied by moderate intraseg-
ment homogeneity and high intersegment heterogeneity; and
when under-segmentation happens, the image objects are gen-
erally of larger sizes and comprise mixed geo-objects, also the
spectral difference between the under-segmentation image ob-
ject and the average gray value of the whole image is usually
slight than that with optimal scale. So under-segmentation are
usually accompanied by low intrasegment homogeneity and
low intersegment heterogeneity..
A General Discussion of These Coupling Relationships
The analysis given above are based on
MRS
,
MSS
and
ECW
.
Additionally, there is an available tool related to scale pa-
rameters selection for
MRS
in eCognition. Dr
ǎ
gut et al. (2010)
proposed the idea that the local variance (
LV
) of object het-
erogeneity within a scene can indicate the appropriate scale
level, and they introduced an
ESP
(Estimation of Scale Param-
eter) tool to find optimal parameters for the multi-resolution
segmentation. However, only intrasegment homogeneity is
taken into consideration in
ESP
tool. This paper concludes
that intrasegment homogeneity plays more important role in
GeOBIA
, which further theoretically proves the rationality of
ESP
tool.
Ming
et al
. (2012) proposed a semivariance based spatial
bandwidth selection method for multi-scale
MSS
. Spatial
bandwidth is a scale parameter that control the segmented im-
age object size, the resulted changing trends of intrasegment
homogeneity and intersegment heterogeneity were similar
with those changing with
SP
in
MRS
. Considering both spatial
and spectral bandwidth selection and taking
MS
segmentation
as an example, Ming
et al
. (2015) proposed a spatial statistics
based scale selection method for multi-scale segmentation.
Both two resulted changing trends of intrasegment homo-
geneity and intersegment heterogeneity along with spatial
and spectral bandwidth are also similar with those changing
with
SP
in multi-resolution segmentation. More experimental
details, please refer to (Ming
et al
. 2012; 2015; and 2016).
Although Ming
et al
. (2012) and Ming
et al
. (2015) did not
further carry out series of classification, the similar changing
trends of intrasegment homogeneity and intersegment hetero-
geneity are effective and powerful supplements for general
discussion of these coupling relationships.
Espindola
et al
. (2006) proposed an objective function for
selecting suitable parameters for region-growing algorithms.
In region-growing segmentation, area threshold play the role
of controlling the image object size, and spectral similarity
play the role of controlling the spectral heterogeneity of image
object. The changing trends of weighted variance (intraseg-
ment homogeneity) and Moran Index (intersegment heteroge-
neity) were similar with this paper. However in the objective
function for selecting suitable parameters, the weights of
intrasegment homogeneity and intersegment heterogeneity are
equal to each other (both are set 0.5), just because the aware-
ness of the relative importance of the two measurements was
not built. The research findings in this paper just provide a
reasonable support for setting the suitable weights of intraseg-
ment homogeneity and intersegment heterogeneity for scale
selection in
GeOBIA
.
Conclusions
This paper employs series of segmentation and classification
with different segmentation scales to discuss the coupling
relationship between image segmentation and classification
accuracy. The research idea and research findings are in a
general sense. The main conclusions coming from the experi-
mental results and analysis are as follows:
1. With increase of spectral heterogeneity parameter, image
object amount decrease; the intrasegment homogeneity
also decreases, however the intersegment heterogeneity
increases or increases first then decrease; there is highly
positive correlation among intrasegment homogeneity,
intrasegment homogeneity and classification accuracy; an
appropriate scale parameter means compromise between
intrasegment homogeneity and intersegment heterogeneity.
2. Intrasegment homogeneity plays more important role than
intersegment heterogeneity in
GeOBIA
. In other words,
over-segmentation is not absolutely unacceptable because
over-segmentation corresponds to high intrasegment homo-
geneity. If the classification samples can cover all kinds of
sub-category to the utmost, the classification accuracy can
be ensured. However, when under-segmentation happens,
low intrasegment homogeneity directly leads to poor classi-
fication even though the intersegment heterogeneity is high.
3. Shape heterogeneity is a special scale parameter in
MRS
.
With increase of shape value, the amount of image object
is on the decline, however the decline trends is not dra-
matic. Under such condition, the intrasegment homogene-
ity decreases and the overall trend of intersegment hetero-
geneity is downward but with irregular fluctuation. A high
shape value leads to deviation between the segmented
image objects and the actual border of the geo-object. It is
not suggested to set shape value too large. When the shape
of geo-objects is regular, parameter shape of about 0.5 is
applicable.
4. Similarly as many previous researches, this paper uses
Moran Index to express the intersegment heterogeneity.
The result of correlation analysis between intersegment
heterogeneity and classification accuracy when changing
shape parameter is inconsistent with that of theoretical
analysis. The intersegment heterogeneity based on Moran
Index is computed by using the global average gray value,
however the theoretical intersegment heterogeneity is
toward local adjacent image objects. So, the two computa-
tion is naturally different. The validity of using Moran
Index to express the intersegment heterogeneity should
be further discussed and a more appropriate intersegment
heterogeneity computation should be further explored.
The conclusions provide theoretical supporting for select-
ing suitable scale parameter or setting the suitable weights of
objective function like used by Espindola
et al
. (2006), so the
research findings are practically meaningful for scale process-
ing in
GeOBIA
.
Acknowledgments
This work is supported by the National Natural Science Foun-
dation of China (41371347) and (41671369) and ‘‘the Funda-
mental Research Funds for the Central Universities’’. Many
thanks to Erdas Imagine for their providing with sample data.
Many thanks to the reviewers for their helpful comments and
suggestions.
References
Anders, N. S., A.C.,and W. Bouten, 2011. Segmentation optimization
and stratified object-based analysis for semi-automated
geomorphological mapping,
Remote Sensing of Environment
,
115(12):2976-2985. doi: 10.1016/j.rse.2011.05.007.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
691
