Analysis of remotely sensed image data by means of category decomposition
M. Inamura · Electronics and Communications in Japan (Part II Electronics) · 1988
Abstract The pattern recognition of a spectral characteristic which is the same method as the latter recognition, is one of the most widely used in remotely sensed multispectral image processing. There are many pixels in remotely sensed image data in which multiple categories are mixed within an instantaneous field of view. This paper points out that a conventional category classification in which the generalized distance of a feature space is used as the evaluation standard of the similarity is not necessarily suitable for the contents of pixels in remotely sensed image data. Also, it is more logical to use a method based on the calculation of the areaoccupation rate (the mixture ratio) of each category. From the viewpoint of the composition and decomposition of the categories, the generalized inverse matrices, least‐square, and quadratic programming were reexamined, and they are compared with conventional methods by using numerical examples. The results confirm the usefulness of the categorydecomposition method and problems of the conventional category‐classification methods including the maximum‐likelihood method (i.e., misclassification of pixels containing multiple categories). As a conclusion, if a more accurate calculation of the mixture ratio becomes available by a further development of the method, a higher order of information than that by a conventional categoryclassification processing applied to the image data on the earth surface will be obtained; it will even be possible to exceed the limit of the space‐resolution of a remote sensor.