a recently developed classifier based on statistical learning
theory that seeks to find an optimal separating hyperplane
between classes by focusing on the training cases that lie at
the edge of the class distribution, i.e., the support vectors,
with the other training cases effectively discarded (Mercier
and Lennon, 2003). The
SVM
has been found to provide higher
classi cation accuracies than other widely used classifiers,
such as the maximum likelihood and the multilayer percep-
tron neural network classi ers (Melgani and Bruzzone, 2004;
Mountrakis
et al.,
2011).
Finally, the classification results of the non-shaded and
shaded areas were merged. All the classes in the merged
classification result were then aggregated into two classes;
impervious surface and pervious surface. Specifically, the im-
pervious surface class includes rooftop, road, other impervi-
ous surface, and
INS
(only in the winter image) in non-shaded
areas, and impervious surface in shaded areas. All other
classes were merged into pervious surface.
Result Assessment and Comparison
Impervious surface results created from the summer and
winter images were quantitatively evaluated and compared
in two aspects. First, the classification results from each step
were evaluated in terms of classification accuracy. Second,
the area estimation and spatial distribution of the extracted
impervious surface were produced and analyzed.
Accuracy Assessment
As in most existing studies that used very high resolution
(
VHR
) images for impervious surface extraction (e.g., Lu and
Weng, 2009; Lu
et al.,
2011; Zhang
et al.,
2013; Sugg
et al.,
2014), reference data for both training and validation were
mainly collected by visual interpretation of aerial photogra-
phy acquired by low altitude
UAV
(unmanned aerial vehicle),
pan-sharpened WorldView-2 images, and
VHR
images from
Google Earth
™
. Those samples of which the attributes are dif-
ficult to determine on remotely sensed data (e.g., the samples
in shaded areas) were checked from ground survey. A prob-
ability sampling protocol, namely stratified random sampling,
was implemented to specify the locations at which validation
points would be obtained. This sampling method has been
used widely and shown to improve sampling efficiency by
reducing variability within a stratum and increasing variabil-
ity between strata (Lohr, 1999; Jin
et al.,
2014).
There were three types of classification results created
from each image for quantitative evaluation: the general urban
land cover classification of the whole image, the shadow clas-
sification, and the final impervious surface extraction result.
Because of the difference in the number of classes identi-
fied from the summer and winter images (i.e.,
INS
class only
appeared in the winter images), the samples for validation
of general urban land cover classifications from summer and
winter images were independently generated using the strati-
fied random sampling method. Similarly, because of signifi-
cant difference in locations and areas of shadows between
summer and winter images, the samples for validation of
shadow classifications from summer and winter images were
also independently generated using the stratified random
sampling method. In both cases, the classification result of
each image was used as a prior stratification.
To fully assess and compare classification accuracy of the
final impervious surface extraction results from summer and
winter images, a reliable stratified random sampling method
was implemented. To highlight the difference between extrac-
tion results from summer and winter images and create reliable
stratification, four strata were created. Stratum I included pixels
identified as pervious surface in the summer image and as im-
pervious surface in the winter image. Stratum II consisted of pix-
els identified as impervious surface in the summer image and as
pervious surface in the winter image. Stratum III included pixels
identified as impervious surface from both season images, and
stratum IV included pixels identified as pervious surface from
both season images. Allocation of sample size to strata is propor-
tional to stratum area. Class labels were then assigned according
to the ground data for summer and winter seasons, respectively.
Thus, class labels of the validation samples in some areas could
be different for summer- and winter-derived classified maps. For
example, the samples located in the areas that were obscured by
deciduous tree canopies in summer and exposed as streets or
sidewalks in winter, were assigned pervious surface for the sum-
mer image and impervious surface for the winter image.
The error matrix approach was adopted to assess the clas-
sification results since it is the most common way to represent
the classification accuracy of remotely sensed data (Congal-
ton, 1991; Congalton and Green, 2008). A sample-based error
matrix was constructed inferring accuracy from a sample to
the population. Cell proportions of the error matrix and the
associated accuracy measures were estimated from the valida-
tion sample following standard probability sampling proto-
cols to construct statistically consistent accuracy estimates
(Stehman, 2000). For each matrix, cell entries are expressed
as percent area of the entire region.
In the accuracy assessment of the general urban land cover
classification of the whole image and the shadow classifica-
tion, the strata used in the stratified random sampling method
to select the validation sample correspond exactly to the
map classes. For these two cases, estimated population error
matrix was calculated using the method described in Pontius
and Millones (2011). However, in the accuracy assessment of
the impervious surface extraction, the four strata created for
selecting the validation sample differ from the map classes
of any classification performed, as described previously. The
accuracy estimators applicable to this situation described by
Stehman (2014) were used. In all cases, descriptive statistics
of overall accuracy (
OA
), class-specific producer’s accuracy
(
PA
) and user’s accuracy (
UA
) were computed.
Besides, quantity disagreement and allocation disagreement
(Pontius and Millones, 2011) were also calculated from each
estimated population error matrix and summarized to provide
a concise comparison of different impervious surface extrac-
tion results. Quantity disagreement is defined as the amount of
difference between the reference map and a comparison map
that is due to the less than perfect match in the proportions
of the categories, and allocation disagreement is the amount
of difference that is due to the less than optimal match in the
spatial allocation of the categories, given the proportions of the
categories in the reference and comparison maps (Pontius and
Millones, 2011). These two additional error terms (disagree-
ments) allow one to determine whether the classification was
properly allocating the cover classes to the right location but
detecting too few or too many of them (quantity disagreement);
or conversely, properly detecting the total number of cover
class pixels, but showing them in the wrong place on the map
(allocation disagreement) (Hladik
et al.,
2013). These two types
of disagreements are components of the overall disagreement.
Result Comparison
All the impervious surface extraction results obtained from
both summer and winter images were first visually compared.
The area proportions of impervious surface and pervious
surface extracted for each season were then estimated from
mapping results. Although the pixel counting method,
which counts the number of pixels allocated to a class and
multiplies this number by the area of a pixel to obtain the
mapped area of the class, provides a simple way to compute
the of area of the class, the method is often biased for the
true proportion of area (i.e., measurement bias) because of
classification error, even when the classification result has
a high overall accuracy (Olofsson
et al
. 2013). Thus, instead
340
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
