and 8 bands (4 standard colors: red, blue, green, near-IR; 4
new colors: red edge, coastal, yellow, near-IR2). Image E is
a subset Spot-5 pansharpened image with 2.5-meter resolu-
tion multispectral band. The sizes of image A, B, D and E are
all 1000*1000 pixels and the size of image C is 1,201*1,137
pixels. To clearly reflect the experimental results, a local part
of image A (Sub-A) is specially shown in Figure 2 (f).
Selection of Classification Features and Samples
In object based classification, simpler features or simpler clas-
sifiers can reflect the influence of scale or segmentation more
directly than complex features or classifiers, so not taking
samples into account, the classification results are mainly de-
termined by segmentation quality under simple experimental
conditions. Based on this consideration, this paper used aver-
age brightness of all bands, Normalized difference vegetation
index (
NDVI
), Normalized difference water index (
NDWI
) and
Gray-level co-occurrence matrix of band1 (texture feature at
directions of 90°) as the features for classification based on the
membership function classifier, though there are some other
effective multiple features (such as Differential morphological
pro les (
DMPs
) (Aytekin and Ulusoy, 2011; Benediktsson
et al
.,
2003), Morphological building index (
MBI
) (Huang and Zhang,
2011), Morphological shadow index (
MSI
) (Huang and Zhang,
2012a), Urban complexity index (
UCI
) (Huang and Zhang,
2012b), multiple features (Li
et al
., 2017a; Huang
et al
., 2017)
and etc.) and classifiers (such as SVM ensembled classifier
(Huang and Zhang, 2008; Huang and Zhang, 2013) and multi-
index learning based classifier (Huang and Zhang, 2014), and
RF
based classifier (Huang
et al
., 2015), deep learning tech-
nique (Li
et al
., 2017b) and etc.) that can be used in
GeOBIA
.
Of all experimental images, image A, B and C were
segmented and classified by using Definiens Professional 8
software, and image A, D and E were segmented by
MS
and
ECW
segmentations and all classifications were performed in
Definiens Professional 8. In
MS
segmentation, the parameter
(
h
s
,
h
r
) for image A, D, and E are respectively (15, 8), (7, 8),
and (16, 7) by spatial statistics based scale parameter estima-
tion (Ming
et al
., 2015). In the experiments, all the classifi-
cation systems and the samples were visually selected. For
every image, all the classification samples were selected from
some certain scale segmentation results on which the com-
pleteness of parcels were visually better and the types of the
parcels could be utmost contained. Additionally, all samples
used in the different groups of experiment are basically the
same controlled by the sample locations to maintain the reli-
ability of the experimental results.
About the selection of testing samples, as Castilla
et al
.
(2012) points out, there is no objective ‘ground truth” to use
as reference for accuracy. To evaluate the overall classification
accuracy (
OA
), this research selected the testing samples at seg-
mentation level from the original image with step sizes of 50
pixels row by row and column by column. The testing sample
selection criterions are similar to those of classification sample
selection. Table 1. gives the details of the classification system
and the samples related to the five experimental images.
Experimental Results of Segmentation and Classification
Thirty-three series of segmentation and classification were
totally performed. Limited by the paper length, only part of
segmentation and classification results with different param-
eter combination of Sub-A are demonstrated in Figure 3.
For series of segmentation, the segmentation evaluation
indices
AV
,
MI
were normalized to 0 to 1 respectively by
Equations 3 and 6), thus intrasegment homogeneity
N
(
AV
)
and intersegment heterogeneity
N
(
MI
) were calculated. The
larger the two evaluation index (
N
(
AV
) or
N
(
MI
)) is, the better
the segmentation is. The change tendencies of the average
intrasegment homogeneity (
N
(
AV
)) and the intersegment
heterogeneity(
N
(
MI
)) with different parameters by using differ-
ent segmentation methods are shown as in Figure 4.
Figure 5 illustrates the changing trends of overall classi-
fication accuracy (
OA
) with different segmentation methods
and different parameters. It is needed to be stressed that the
results of different segmentation methods for the same image
cannot be absolutely compared because the scale parameters
used in different segmentation is naturally different, such as
Figure 5 (a), (d), and (e).
Analysis and Discussion
Scale Effect on the Image Object Amount with
Change of Segmentation Scale Parameter
For series of segmentation, different parameters directly
impact image object amounts or object sizes. Image object
size is of spatial scale meanings, and image object amount is
not, though large image object amount means small average
image object size and small image object amount means large
average image object size. However, because the segmented
image object size varies from each other in a certain scale
segmentation, the average object size is somewhat unpracti-
cally meaningful. Figure 6 demonstrates the changing trends
of image object amount with different segmentation methods
and different parameters.
Table 1. Classification system and the sample amounts of the experiments.
Image
Class name
Ncs Nts
Image
Class name
Ncs Nts
Image
Class name
Ncs Nts
A
construction land 25 46
B
construction land 57 61
C
construction land 52 72
grassland
14 65
grassland
19 58
vegetation
60 45
woodland
13 32
woodland
36 70
water area
18 40
road
14 97
road
80 44
road
42 46
shadow
11 23
-
bare land
10 11
D
construction land 45 94
E
construction land 28 24
vegetation
20 34
vegetation
29 45
water area
5 31
dry farmland
50 126
road
13 71
paddy field
14 15
shadow
19 110
road
24 18
bare land
5 10
-
Note: Ncs-number of classification samples; Nts-number of testing samples
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
685
