Variante Do Algoritmo Isodata Para Aplicacao Em Alvos Agricolas (Variant of the Isodata Algorithm for Application to Agricultural Targets)

摘要

A variant is presented of the ISODATA algorithm for automatic clustering of agricultural targets in multispectral images. In the same way as ISODATA, the algorithm dispenses with a priori knowledge of the number of classes. Compared to ISODATA, the algorithm has the following advantages: it depends on less parameters, it converges quicker and is guaranteed to terminate. The proposed algorithm start with a single class and, in the following iterations, splits classes for which the variance is greater than a given constant C. The algorithm terminates when all classes have variances within the desired limit. To minimize the mean square error, at each iteration, a k-means algorithm is performed where k is the iteration index. Examples are shown of the application of the algorithm for LANDSAT/TM images.