Remote Sensing and Geospatial Analysis for Landscape Pattern Characterization - Landscape Ecology for Sustainable Environment and Culture - page 202

by USGS EROS Data Center over nearly the past four decades can be downloaded
via internet at no charge ( http://landsat.usgs.gov ). This moderate-resolution image
archive is unmatched in quality, details, coverage, and length, and has been an
invaluable resource for examining natural and anthropogenic changes on Earth's
surface (Yang 2011a ).
11.2.1.2 Information Extraction from Remote Sensor Data
Both visual interpretation and computer-based image classification techniques can
be used to extract land cover information from remote sensor data. Through the
combined use of various image elements with human intelligence, visual inter-
pretation can achieve excellent mapping accuracy. And this technique can be
implemented through on-screen digitizing in a GIS environment. However, it is
manual by nature, and can be very much labor intensive for land over mapping
over a large area. On the other hand, computer-based image classification can
automate the entire land cover mapping process although some further research is
still needed towards the operational use of this promising technique. In general,
image classification is preferred over visual interpretation for land cover mapping
over large areas (Jensen 2007 ).
Remote sensor image classification is largely based on the manipulation of
statistical characteristics of one or more multispectral scenes. A variety of clas-
sification methods have been developed for remote sensing applications. By using
specific criteria, these methods can be grouped in different ways: a parametric or a
nonparametric classifier; a supervised or an unsupervised classifier; a ''hard'' or a
''soft'' classifier; a spectral or a spatial classifier; a sub-pixel, a pixel or an object-
based classifier (Jensen 2005 ). Each of these methods has its own advantages and
disadvantages, and there is no any single classifier that can be superior to another
in all aspects (Duda et al. 2001 ).
Among all existing image classifiers, some are considered as advanced ones.
However, most of the mapping applications have relied upon the use of a con-
ventional classifier that largely manipulates a single image element (e.g. color or
tone) in the multispectral pattern recognition (Jensen 2005 ). Conventional pattern
recognition methods are largely based on the use of parametric statistics, which
generally work well for medium-resolution scenes covering spectrally homoge-
neous areas, but not in heterogeneous regions or when scenes contain severe noises
due to the increase of image spatial resolution. For years, substantial research
efforts have been made to improve the performance of pattern classification for
working with different types of remote sensor data and with spectrally complex
landscapes. Some strategies have been developed as a result of such efforts: (1) the
identification of various hybrid approaches that combine two or more classifiers, or
incorporate pre- and post-classification image transformation and feature extrac-
tion techniques (e.g. Yang and Liu 2005b ); (2) the development of 'soft' classifiers
by introducing partial memberships for each pixel to accommodate the hetero-
geneous and imprecise nature of the real world (e.g. Shalan et al. 2003 ); (3) the