2.2 Fusion Processing
Ten stations of the local triangulation network (Hong Kong 80 Grid System) were
selected as ground control points (GCPs) for geometric registration followed by the geo-locking
of the XS image to the PAN image. After the “geometric locking” process, fusion of
the images was performed by the three approaches, IHS, PCA and HPF.
The XS image was transformed into IHS and PC spaces, and the high spatial component
(PAN image) replaced the intensity and first principal component respectively. The reverse
function transformed the replaced and other components back into the RGB mode of the
newly generated, fused XS image. For the HPF fusion, the spatial detail from panchromatic
image (PAN) was extracted by differential high pass filtering and directly integrated into the
XS image to form HPF-fused image. The three child images generated by IHS, PCA and HPF
fusion process are shown in Figure 2.
 |
| (a) |
(b) |
(c) |
Figure 2. Fusion results of SPOT XS and SPOT PAN images: (a) IHS-fused image; (b) PCA-fused image; (c)
HPF-fused image.
2.3 Training Stage for Classification
The USGS classification systems for remote sensing (Anderson et al., 1972) was used to
define the classification themes. Using Anderson’s classification system, four categories -
water, vegetation, urban and barren land - were defined for level one classification. Another
eight categories - sea, inland water, grassland, woodland, mixed urban area, transportation,
sandy area and open space - were defined for level two classification. The descriptions of all
categories are summarized in Table 1.
Table 1 Description of designed categories
in level one and two.
| CLASS |
DESCRIPTION |
| LEVEL 1 |
| WATER |
Water coverage includes ocean, streams and reservoirs |
| VEG |
Vegetated coverage includes woodland, grassland and swamp. |
| URBAN |
Artificial coverage includes residential, commercial, industrial and transportation. |
| BARREN |
Uncovered coverage include beach, rock outcrop and construction sites |
| LEVEL 2 |
| SEA |
Predominant and large bodies of water separate the land mass |
| INLAND |
Water covered area within the land mass includes reservoirs, streams and estuaries. |
| GRASS |
Shrub and brush covered area |
| WOOD |
Forest covered area. |
| MXURBAN |
Intensive built-up area includes residential, industrial and commercial zone. |
| TRANS |
Linear routes as clearcuts and boundaries outline the other land use. |
| SANDY |
Un-vegetated sands dunes dominated along the coast. |
| OPEN |
Temporary barren land included construction sites and bare space. |
Due to the time difference bwteen imagery date and the date of this study, field
verification of the classified images was invalid. Total reliance was made on the ancillary
information sources for the ground reference data. Ancillary information used during training
included a number of aerial photos (SMO, 1993 & 1994), Hong Kong land utilization map
(SMO, 1988) and Hong Kong topographic maps (SMO 1990 & 1998).
In acquiring ground truth data, the size and delineated units were designed with reference
to Congalton and Green (1999). For the sample size of training data, a minimum of 50
samples for each category was used. For each class, a stratified random sampling technique
was used in capturing the location of each training set. The delineated locations were
distributed over the entire image. Based on the ancillary information, more than 130 sites
were delineated for each of the two classification cases. Two sets of data were identified for
each class, one set was used for the classification training and the other to assess the thematic
accuracy.
After the training process, all the child and parent multispectral images were classified
into the two levels through three selected approaches.
2.4 Classification Results
Table 2 lists the overall accuracy and the percentage of correctly classified pixels in
individual classes of the level one classification result using three child-images and parent
spectral image. For the level two classification, a second, independent data sets was obtained
for the training stage. The classification results are summarized in Table 3.
Table 2. The thematic accuracy of natural
land cover classification results.
| Thematic Accuracy of Classification Results (%) |
Accuracy
Assessment |
XS |
IHS-fused |
PCA-fused |
HPF-fused |
| MLC |
| WATER |
99.10 |
100.00 |
95.10 |
96.50 |
| VEG |
100.00 |
95.90 |
95.40 |
93.30 |
| URBAN |
92.90 |
88.80 |
86.30 |
76.10 |
| BARREN |
94.40 |
91.90 |
86.90 |
73.40 |
| Overall Accuracy |
93.45 |
91.66 |
88.78 |
80.51 |
| NNC |
| WATER |
99.30 |
100.00 |
84.50 |
97.20 |
| VEG |
100.00 |
100.00 |
99.50 |
99.00 |
| URBAN |
96.70 |
83.00 |
88.50 |
94.10 |
| BARREN |
100.00 |
91.00 |
99.50 |
80.60 |
| Overall Accuracy |
98.05 |
88.85 |
91.66 |
92.83 |
Table 3. The thematic accuracy of cultural
land use classification results.
| Thematic Accuracy of Classification Results (%) |
Accuracy
Assessment |
XS |
IHS-fused |
PCA-fused |
HPF-fused |
| MLC |
| SEA |
66.30 |
83.80 |
88.30 |
65.50 |
| INLAND |
92.00 |
73.60 |
74.70 |
85.10 |
| GRASS |
64.10 |
83.20 |
48.30 |
86.20 |
| WOOD |
64.40 |
72.60 |
56.40 |
79.50 |
| MXURBAN |
59.50 |
65.30 |
45.90 |
16.60 |
| TRANS |
37.90 |
29.20 |
45.10 |
35.10 |
| SANDY |
51.00 |
89.40 |
57.00 |
80.80 |
| OPEN |
79.20 |
59.50 |
30.90 |
83.50 |
| Overall Accuracy |
63.55 |
67.54 |
53.47 |
64.55 |
| NNC |
| SEA |
53.30 |
35.80 |
51.50 |
35.80 |
| INLAND |
96.60 |
74.70 |
79.30 |
82.80 |
| GRASS |
95.30 |
81.90 |
57.40 |
55.40 |
| WOOD |
55.00 |
76.90 |
56.40 |
54.10 |
| MXURBAN |
82.50 |
56.20 |
48.90 |
28.40 |
| TRANS |
21.90 |
16.90 |
30.70 |
46.10 |
| SANDY |
85.40 |
98.70 |
91.40 |
86.80 |
| OPEN |
92.40 |
51.70 |
44.10 |
86.90 |
| Overall Accuracy |
69.82 |
56.25 |
51.89 |
56.12 |
