igery acquired in 
5 0T 5 image had 
>9 um, 0.61-0.68 
ution and one in 
olution (Figure 2 
rial photography 
rive training/test 
odel based and 
P mapping. The 
ced into a digital 
rhe SPOT 5 data 
lly corrected and 
rojection using a 
' Hong Kong. 
ita, several pre- 
a more direct 
.1 phenomena. In 
sing steps were 
above, mainly 
l (with 0.1 pixel 
ig applying to 
ures (coherence, 
r to geocode the 
t data and the 
d to generate a 
between DEM 
geometry. The 
xel accuracy by 
nma-rs.ch). The 
erage amplitude 
resampled at 10 
eloped by Yang 
percentage of 
)f SPOT multi- 
oach used high- 
for representing 
dium-resolution 
e spatial areas. 
;, Green- 
td3,) 
An advantage of this CART-based approach over other ISP 
estimation approaches based on spectral clustering (e.g., 
spectral mixture analysis) is their applicability to use multi 
source and multi-sensor remote sensing data which are 
characterized with different statistical distribution and different 
physical feature. In addition, it permits the exploration of large- 
area impervious surface information at sup-pixel level in a 
repeatable and objective way. In this study, the process of 
mapping impervious surface using the CART-based approach 
involves the following steps (as shown in Figure 3): 1) 
development of training/test data using 33cm-resolution digital 
aerial CIR photography, 2) design of predictive variables, 
establishment and assessment of final regression tree modelling, 
3) spatial extendibility of ISP prediction modelling with 
middle-resolution remote sensing dataset, and 4) accuracy 
assessment of ISP mapping. 
33-cm resolution CIR 
aerial images 
CART-based classification 
X 
ISP estimation (0.33m) 
I 
Middle-resolution 
sensing data:/ 
SPOT-4. 
InSAR-3,and/ 
SPOT-4 InSAT-T! 
Overlay on 10m image grid 
I 
ISP reference data (10m) 
I 
Development of 
training/tesf data 
_ J 
~ 1 
Sub-pixel ISP prediction 
modeling using Regression Tree 
algorithm 
(training data and features) 
X 
ISP prediction 
model 
buildirJ& ISP 
ISP mapping based on 
prediction model 
T 
mapping 
Accuracy assessment 
Figure 3 Block scheme of impervious surface mapping with 
remote sensing data and a regression tree model 
3.1 Development of ISP training/test data 
Successful ISP prediction modelling using CART algorithm 
relies on the quality of training-test data. In this study, training 
and test data were derived from four subset patches of the 
33cm-resolution CIR aerial image. These image subsets were 
approximately 2000 x 1500 m each and were visually selected 
to avoid areas where land cover changes occurred among the 
CIR aerial images, the SPOT imagery and the InSAR products. 
Initially, the selected CIR aerial images were classified using 
the CART machine learning algorithm. Each pixel was 
classified as one of five land cover classes: impervious surface, 
vegetated areas, bared soil, water, and shadow. The 
classifications were further modified by screen digitizing and 
recoding to reduce misclassification. Validated by the in situ 
data, the classification results archived overall accuracy 87.56% 
and Kappa coefficient 0.811. Once the final classification 
results were obtained, all 33-cm resolution pixels mapped as 
impervious surface were tallied using a 10 x 10 m grid 
geographically aligned with the middle-resolution data pixels 
(10m resolution corresponding to SPOT and InSAR data in this 
study) to compute percent impervious surfaces. This resulted in 
10-m resolution raster image of percent imperviousness. It is 
noted that the shadow class was excluded from the calculation 
of imperviousness in this process since the class could not be 
unambiguously merged with any single land cover class. 
Figure5 illustrates one of the four CIR aerial image subsets and 
its ISP image with 10-m spatial resolution. 
3.2 ISP prediction model building and ISP mapping 
An ISP prediction model based on the CART regression tree 
algorithm was developed and calibrated by using the 
training/test data above obtained from high-resolution CIR 
aerial image as dependent variable (target variable). In addition, 
the independent variables (predictor variable) were derived 
from the six layer including three SPOT multi-spectral images 
and three InSAR feature images. In this study, 2000 training 
samples and 2000 testing samples were randomly selected from 
the layers of target and predictor variables. It is worth nothing 
that the training samples were independent to the test samples. 
The training/test samples were utilized to build and improve the 
rule-based ISP prediction model by using the CART regression 
tree algorithm. The prediction model was composed of rule sets 
where each rule was defined by one or more conditions and 
corresponding confidence-level under which a multivariate 
linear regression model was established. Once the final ISP 
prediction model was built, it was applied to all pixels of the six 
middle-resolution remote sensing dataset to map percent 
impervious surfaces. 
3.3 Accuracy assessment of ISP estimation 
For accuracy assessment of ISP estimate, the reference data 
should ideally be collected from the field work based on a 
statistical sampling design. Due to time and resource constraints, 
the 10-m resolution ISP derived from the CIR aerial image was 
Figure 4 the CIR aerial imagery with 0.33m resolution and estimated ISP with 10m resolution from CIR imagery 
used for this purpose. To ensure the validity of the assessment,