Environmental Engineering Reference
In-Depth Information
Fig. 1
A location map of the study site
third steps were combined to get the final conclusions. In
this way, combination of the thematic information and the
NDVI was possible in order to infer the nature of change,
between-class or within-class change (Lunetta and Elvidge
1998
), or simply an error.
samples becomes difficult to perform when the landscape is
complex and heterogeneous as they would influence clas-
sification results, especially for classifications with fine
spatial resolution image data (Chen and Stow
2002
). This
problem would be complicated if medium or coarse spatial
resolution data are used for classification, because a large
volume of mixed pixels may occur (Lu and Weng
2007
).
Despite precautions made in the training samples' col-
lection, it is sometimes difficult to identify the most sensi-
tive reference land-cover class for some observations, even
resorting to ancillary data (Carrão et al.
2008
). Thus, the
original training sample, i.e., the sample of pixels that was
directly collected by the analyst, contained unusual training
units. According to Johnson and Wichern (
1998
), unusual
observations are those that are either too large or too small
compared to the others. Thus, in order to identify these
anomalies, it is necessary to apply a statistical procedure
based on the distance of each training unit to its mean class.
Detecting Outliers and Cleaning Training Sample
A suitable classification system and a sufficient number of
training samples are prerequisites for a successful classifi-
cation (Lu and Weng
2007
). Training samples are usually
collected from fieldwork or from fine spatial resolution
aerial photographs and satellite images, and sampling of
sufficient number and their representativeness is critical for
image classifications (Landgrebe
2003
; Mather
2004
). Dif-
ferent collection strategies, such as single pixel, seed, and
polygon, may
be
used,
but
selecting
sufficient
training
Search WWH ::
Custom Search