131
such as CASI and MERIS (Babey and Anger, 1989, Rast, Bäzy et al. 1991)
provides tools that may help to address the problem, though, at present, the
use of these for change detection remains at best experimental. Seasonal
changes in spectral signature, in principle, may be exploited to enhance class
differentiation (Schriever and Congalton 1995), but this is obviously at the
cost of detecting the seasonal changes themselves.
7.
IMAGE SEGMENTATION METHODS
The most common approach to the segmentation and interpretation of
multi-spectral remotely sensed data for land cover mapping utilizes a suite of
probabilistic classification and clustering algorithms. Supervised
classifications exploit the radiometric properties of known ‘training’ regions
to identify areas elsewhere on the image with similar spectral properties. The
hypothesis is that land cover of the training regions is identical to regions
elsewhere in the scene with similar spectral characteristics. Kershaw and
Fuller (1992) demonstrated the importance (and difficulty) of selecting
training data that are truly representative of spectrally unique classes.
Unsupervised clustering methods have been successfully employed,
especially for the segmentation of natural areas, where the choice of training
regions is notoriously difficult (Belward, Taylor et al., 1990; Lark, 1995).
However, the use of unsupervised methods for change detection presents
unsolved problems, relating to the difficulty of generating comparable
clusters from images with even slightly different radiometric characteristics.
Spectral mixture modeling has been widely advocated (e.g., Adams, Sabol et
al., 1995; Foody and Cox, 1994) as a means of dealing with continua, rather
than distinct classes, but examples of operational applications are rare.
More advanced techniques, based to a greater or lesser extent on the use
of artificial intelligence, include rule-based methods which recognize the
context of pixels, rather than treating them in isolation (e.g., Alonso and
Soria, 1991; Groom, Fuller and Jones, 1996; Binaghi, Madella et al., 1997)
and methods based on neural networks and other fuzzy logic approaches
(e.g., Gopal and Woodcock, 1996; Chen, Tzeng et al., 1995; Schaale and
Furrer, 1995; Dreyer, 1993). Smith et al. (1997) have recently developed
map- and image-based segmentation approaches, in which the classification
unit is the land parcel, rather than the pixel; these appear to deliver
substantially better performance than conventional per-pixel classifiers.
Nevertheless, despite the richness of this recent literature and the claimed
improvements in performance over conventional multi-spectral
14. REMOTE SENSING OF LAND COVER AND LAND COVER
CHANGE
such as CASI and MERIS (Babey and Anger, 1989, Rast, Bäzy et al. 1991)
provides tools that may help to address the problem, though, at present, the
use of these for change detection remains at best experimental. Seasonal
changes in spectral signature, in principle, may be exploited to enhance class
differentiation (Schriever and Congalton 1995), but this is obviously at the
cost of detecting the seasonal changes themselves.
7.
IMAGE SEGMENTATION METHODS
The most common approach to the segmentation and interpretation of
multi-spectral remotely sensed data for land cover mapping utilizes a suite of
probabilistic classification and clustering algorithms. Supervised
classifications exploit the radiometric properties of known ‘training’ regions
to identify areas elsewhere on the image with similar spectral properties. The
hypothesis is that land cover of the training regions is identical to regions
elsewhere in the scene with similar spectral characteristics. Kershaw and
Fuller (1992) demonstrated the importance (and difficulty) of selecting
training data that are truly representative of spectrally unique classes.
Unsupervised clustering methods have been successfully employed,
especially for the segmentation of natural areas, where the choice of training
regions is notoriously difficult (Belward, Taylor et al., 1990; Lark, 1995).
However, the use of unsupervised methods for change detection presents
unsolved problems, relating to the difficulty of generating comparable
clusters from images with even slightly different radiometric characteristics.
Spectral mixture modeling has been widely advocated (e.g., Adams, Sabol et
al., 1995; Foody and Cox, 1994) as a means of dealing with continua, rather
than distinct classes, but examples of operational applications are rare.
More advanced techniques, based to a greater or lesser extent on the use
of artificial intelligence, include rule-based methods which recognize the
context of pixels, rather than treating them in isolation (e.g., Alonso and
Soria, 1991; Groom, Fuller and Jones, 1996; Binaghi, Madella et al., 1997)
and methods based on neural networks and other fuzzy logic approaches
(e.g., Gopal and Woodcock, 1996; Chen, Tzeng et al., 1995; Schaale and
Furrer, 1995; Dreyer, 1993). Smith et al. (1997) have recently developed
map- and image-based segmentation approaches, in which the classification
unit is the land parcel, rather than the pixel; these appear to deliver
substantially better performance than conventional per-pixel classifiers.
Nevertheless, despite the richness of this recent literature and the claimed
improvements in performance over conventional multi-spectral
14. REMOTE SENSING OF LAND COVER AND LAND COVER
CHANGE