Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 99

A general concept offered by Jensen ( 2007 ), is that in developing training statis-
tics, it is necessary to select a number of pixels in each class that is at least 10
times greater than the number of bands used during the classification process. This
is enough to allow good computations of variance-covariance matrices, which are
usually carried out with classification software. Related to size of sample sites, it is
noted that ''as sites grow larger than 10 pixels, there may be no new information
added. So, it would be better to have six sites of 10 pixels in each class rather than
one training site of 60 pixels'' (Schowengerdt 2007 ).
To classify the remotely sensed data, the classification algorithm needs to be
trained to distinguish one class from another. Representative identical class sites are
known as prototypes, exemplars or training samples. After the classifier is trained to
statistically analyze to ''distinguish'' the unlike classes represented by the training
sites, the ''rules'' that were developed during the phase of training are used to label all
pixels in the image to their ''in real world'' classes (Schowengerdt 2007 ).
A large enough number of training samples and their ability of representa-
tiveness are significant for image classifications (Mather 2004 ). When the bio-
physical structure of the study area is complex and heterogeneous, selecting
enough training samples will be difficult. This problem would be greater if med-
ium or coarse spatial resolution data were used for classification, because a large
number of mixed pixels may occur. So, the choice of training samples must
consider the three standards: (1) the spatial resolution of the available remote
sensing imagery; (2) availability of ground truth data; and (3) the complexity of
the biophysical structure (Lu and Weng 2007 ).
Training samples are usually collected from fieldwork/in situ, fine spatial res-
olution aerial photographs and satellite images/in-image, recently from Google
Earth, etc. Different gathering strategies, such as single pixel, seed and polygon,
can be used (Chen and Stow 2002 ).
Care must be taken to collect representative and non-auto-correlated training
samples. The problem in spatial autocorrelation occurring in remote sensing data is
that pixels in the image should not be considered as fully discrete features inde-
pendent of their juxtaposition, but rather a set of continuous features influenced by
their neighbors (Campbell 1981 ). This exists among pixels that are neighboring
(e.g., neighboring pixels have a high chance to have alike brightness values),
which can cause a decrease in variance between neighboring pixels (Campbell
1981 ). This decrease in variance can make large masses of neighboring training
pixels less representative of a particular LULC-class in the entire image; in con-
trast, the use of several single-pixel training samples that are situated spatially
separately from each other can result in better classifications than large masses
(polygons) of neighboring training pixels (Medhavy et al. 1993 ). Therefore, if such
care is taken, classification results for LULC-types (especially for crop recogni-
tion, since they have, generally speaking, a relatively small spatial distributions/
fields) can be more effective.
Google Earth ( http://earth.google.com/ ) contains ever more wide-ranging
coverage of the globe at very high spatial resolution 0.61-4 m, allowing the user to
zoom into particular areas to get great detail. Google Earth data were used in this