ISODATA 15-class unsupervised classification of LandSat imagery
Classification of a portion of LandSat imagery from Path 90 Row 79 (around Lake Wivenhoe and the Somerset reservoir, between Brisbane and Toowoomba in southern Queensland), using bands 1,2,3,4,5 & 7. It was performed in ERMapper.
Unsupervised classification is an exercise in point-cluster delineation, each point corresponding to a pixel in the image. In this case, because six wavelength bands were used, clustering was analysed in 6 dimensions. ISODATA is an iterative algorithm, repeatedly merging and re-dividing classes it has previously created, to arrive at an optimally distinct and robust collection of point clusters.
Correspondence between the classes it comes up with and human-conceived land cover type classes is not guaranteed, as in this case. To get round this, the algorithm can be asked to erect many more classes than are envisaged for human use, and then hopefully it is possible to selectively combine ISODATA classes that are more-or-less subsets of a particular human-conceived cover type class.
The class that I have coloured blue was the one that principally corresponded to open water, but it also included areas of particularly dark, lush vegetation, both natural forest and certain arable fields. This probably happened because of green algae (and hence chlorophyll) in the freshwater bodies.
ISODATA 15-class unsupervised classification of LandSat imagery
Classification of a portion of LandSat imagery from Path 90 Row 79 (around Lake Wivenhoe and the Somerset reservoir, between Brisbane and Toowoomba in southern Queensland), using bands 1,2,3,4,5 & 7. It was performed in ERMapper.
Unsupervised classification is an exercise in point-cluster delineation, each point corresponding to a pixel in the image. In this case, because six wavelength bands were used, clustering was analysed in 6 dimensions. ISODATA is an iterative algorithm, repeatedly merging and re-dividing classes it has previously created, to arrive at an optimally distinct and robust collection of point clusters.
Correspondence between the classes it comes up with and human-conceived land cover type classes is not guaranteed, as in this case. To get round this, the algorithm can be asked to erect many more classes than are envisaged for human use, and then hopefully it is possible to selectively combine ISODATA classes that are more-or-less subsets of a particular human-conceived cover type class.
The class that I have coloured blue was the one that principally corresponded to open water, but it also included areas of particularly dark, lush vegetation, both natural forest and certain arable fields. This probably happened because of green algae (and hence chlorophyll) in the freshwater bodies.