An Entropy-Based Multispectral Image Classification Algorithm

Abstract

Employing the entropy theory, this paper presents a new and robust multispectral image classification algorithm. The digital number (DN) in remotely sensed multispectral images is considered as a random variable when judging the allocation of unknown pixels into predefined training classes. If an unknown pixel shows a similar DN vector as the pixels in a training class, it will increase the global entropy defined as the sum of DN probabilities multiplied by the logarithm of DN probabilities for all pixels within the training class. The unknown pixel is to be assigned to the class for which the entropy of the training class is increased most due to the inclusion of the pixel. The proposed entropy-based classification (EC) is compared with the maximum likelihood classification (MLC), parallelepiped classification, minimum distance classification, Mahalanobis distance classification (MDC), iterative self-organizing data analysis technique (ISODATA) classification, and K-means classification. These classifiers were applied to a Landsat Enhanced Thematic Mapper Plus image covering Houston, Texas, USA, acquired on October 16, 1999. A reference land cover map from the National Land Cover Data 2001 of the same area was taken as a ground reference to assess the accuracy of classification results, suggesting that the EC showed comparable overall accuracy as MDC, and they both outperformed other classifiers. The results of MLC can be improved by substituting the multivariate lognormal or gamma distribution for the multivariate normal distribution involved in its assumption. The EC algorithm has the potential to produce reliable land cover maps regardless of the distribution of DN vectors and relevant parameters of probability density functions involved in other classifiers.