all the reference data were selected randomly and
independently from the training/test data used in building ISP
prediction model above.
In addition, the ISP map produced from the remote sensing
dataset in this study belongs to continuous data. Therefore, the
accuracy assessment method based on error matrix and kappa
index used often in classification cannot be appropriate here
because it is designed for categorical data only. In this study,
correlation analysis was used for the quantitative comparison
between the estimated ISP and the reference data. Three
indicators, Average Error (AE), Relative Error (RE) and
Pearson coefficient (R2), were calculated to assess accuracy of
the ISP estimation.
4. EXPERIMENT RESULTS
To test the effectiveness and feasibility of the synergistic use of
optical and InSAR data in urban ISP mapping, three group of
data sources used in this study were (1)4 SPOT HRG multi-
spectral bands, here referred to SPOT 4, (2) 3 InSAR products
(referred to InSAR_3), and (3) 4 SPOT bands plus 3 InSAR
products (referred to SPOT_4+InSAR_3).
estimate derived from the three groups of middle-resolution
remote sensing dataset were illustrated in Figure6. They were
post-processed by masking sea water boundary. Figure6 shows
that The ISP of the urban area located in the commercial
business district of Kowloon and Hong Kong Island was greater
than 60%, and that of the suburbs and hill areas was less than
40%.
Visually, three ISP results shown in Figure 6 were quite
reasonable in spatial distribution pattern and the density urban
areas with high ISP value were similarly mapped. But the ISP
estimate results using just InSAR products were fragmentized
due to coarse spatial resolution and signal distortion in roll
areas (Figure 5(a)). In addition, checking by the use of high-
resolution CIR aerial photographs suggested that the SPOT-
derived ISP was slightly overestimated in the low to middle ISP
areas (e.g., bared soils or vegetated fields) as has been reported
in the previous works (Figure5(b)). This ISP overestimation
problem using SPOT data may be due to similar spectral
reflectance between bared soils or spared vegetation and urban
imperious surface such as parking plots. However, this
overestimation was greatly reduced by synergistic use of SPOT
ISP value
Water
0-30%
30-46%
46-63%
63-78%
78-91 %
91-100%
(a)
(b)
(c)
Figure 5 ISP classification results derived from the three groups of remote sensing datasets in the study area (a) InSAR_3;
(b) SPOT 4; (c) InSAR 3+SPOT 4
The CART based ISP estimate approach was respectively
applied to the three data sources and corresponding ISP
prediction models were built. The classification results of ISP
and InSAR dataset (Figure 5(c)). It should be attributed to the
fact that InSAR features, in particular long-term interferometric
coherence, possessed the capacity to distinguish the two land-
cover classes mentioned above. It further and in detail was
