With a mean of 0.87, the resultant Moran’s Is are apt to be
high, indicating the image surfaces of the generated texture
bands are inclined to be homogenous (Figure 8). The com-
plexity of the image surfaces of these texture bands does
not change considerably when different window sizes were
applied, because the range of the calculated Moran’s Is is only
0.10. Yet a close look of the graph reveals that, for windows
smaller than 33, the resultant Moran’s Is increase rapidly and
maximize at the 31 window. This demonstrates an obvi-
ous transition from complexity to regularity in the extracted
texture features due to the change of window sizes. Starting at
the 33 × 33 window and again the 49 × 49 window, the calcu-
lated Moran’s Is change little for their following windows and
the values from the latter (the 49 – 59 windows) are all higher
than those from the former (the 33 – 47 windows). This indi-
cates the change of window sizes at the 33 – 47 and 49 – 59
windows does not seem to impact much in generating texture
images with varying spatial distribution. For the last window
of 61, however, its result reduces again.
Maximum Likelihood Image Classification
Results and Results Comparison
The overall average
JM
distance for the spectral classification
by the raw G/R/NIR bands is found to be 1.806 (Table 2), in-
dicating an overall acceptable between-class separation. With
a
JM
value of 2.00, water indicates an excellent separability,
followed by forest, grass, and cropland and pasture (
JM
= 1.90
– 2.00), suggesting a good separability. As a result, these four
classes should be easily separated by spectral classification
alone. With a smaller average
JM
distance (1.890), commer-
cial and industrial is less separable from other classes with
an acceptable separability. For the three residential classes,
Table 2. The average Jeffries-Matusita distance of one class with others (values more than 1.900 indicating a good separability
are highlighted in bold).
GR
NIR
Window Sizes
15 19 21 23 27 31 33 35 37 39 41 47 49 55 59 61
CroPas
1.903 1.932 1.963 1.981 1.973 1.977
1.819 1.797 1.819 1.798 1.800
1.961 1.939 1.965 1.969 1.946
1.797
Water
2.000 1.985 2.000 2.000 2.000 2.000
1.838 1.838 1.838 1.838 1.838
2.000 1.985 2.000 2.000 1.985
1.838
ComInd 1.890
1.907 1.936 1.945 1.946 1.950
1.912 1.878 1.882 1.878 1.876
1.917 1.902 1.916 1.924 1.909
1.878
HDR 1.574 1.616 1.648 1.664 1.661 1.669 1.655 1.630 1.610 1.603 1.597 1.640 1.629 1.649 1.651 1.643 1.630
MDR 1.548 1.587 1.622 1.637 1.636 1.654 1.633 1.597 1.581 1.577 1.573 1.610 1.596 1.616 1.620 1.611 1.597
LDR 1.597 1.638 1.671 1.689 1.685 1.695 1.673 1.644 1.625 1.624 1.623 1.662 1.642 1.668 1.664 1.657 1.644
(RAve) 1.573 1.614 1.647 1.663 1.661 1.673 1.654 1.624 1.605 1.601 1.598 1.637 1.622 1.644 1.645 1.637 1.624
Forest
1.983 1.885 1.992 1.996 1.994 1.994 1.954 1.953 1.974 1.951 1.951 1.991
1.887
1.992 1.991
1.886
1.953
Grass
1.957 1.957 1.971 1.980 1.975 1.975 1.941 1.933 1.926 1.925 1.927 1.969 1.956 1.970 1.968 1.955 1.933
OAve 1.806 1.813 1.850 1.862 1.859 1.864 1.803 1.784 1.782 1.774 1.773 1.844 1.817 1.847 1.848 1.824 1.784
*CroPas: cropland and pasture; ComInd: commercial and industrial;
HDR
: high density residential;
MDR
: medium density residential;
LDR
:
low density residential;
RAve
: the mean of the average
JM
distance of three residential classes with others; OAve: the overall average
JM
dis-
tances for one classification;
GRNIR
: the average JM distance for the spectral classification.
Table 3. Error matrices for supervised spectral classification and hybrid classification.
a) Spectral classification only with the G/R/NIR bands (Overall accuracy: 52.5%; overall Khat: 0.67)
LULC classes
Reference Totals
Classified Totals
Number Correct
Producers Accuracy
Users Accuracy Kappa
CroPas
37
50
12
32.43%
24.00% 0.16
Water
49
50
48
97.96%
96.00% 0.95
ComInd
62
50
36
58.06%
72.00% 0.67
HDR
3
50
2
66.67%
4.00% 0.03
MDR
41
50
12
29.27%
24.00% 0.15
LDR
65
50
19
29.23%
38.00% 0.26
Forest
61
50
44
72.13%
88.00% 0.86
Grass
82
50
37
45.12%
74.00% 0.67
Totals
400
400
210
b) Hybrid classification of both spectral and fractal-based spatial information (Overall accuracy: 76.25%; overall Khat: 0.89)
LULC classes
Reference Totals
Classified Totals
Number Correct
Producers Accuracy
Users Accuracy Kappa
CroPas
17
50
18
47.06%
16.00% 0.12
Water
48
50
48
100.00%
96.00% 0.95
ComInd
60
50
42
70.00%
84.00% 0.81
HDR
1
50
0
0.00%
0.00% 0.00
MDR
94
50
69
52.27%
98.00% 0.97
LDR
90
50
50
41.11%
74.00% 0.66
Forest
53
50
41
77.36%
82.00% 0.79
Grass
37
50
21
56.76%
42.00% 0.36
Totals
400
400
305
*CroPas: cropland and pasture; ComInd: commercial and industrial;
HDR
: high density residential;
MDR
: medium density residential;
LDR
:
low density residential.
702
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
