Evaluation of the use of spectral and textural information by an evolutionary algorithm for multi-spectral imagery classification

Considerably research has been conducted on automated and semi-automated techniques that incorporate image textural information into the decision process as an alternative to improve the information extraction from images while reducing time and cost. The challenge is the selection of the appropriate texture operators and the parameters to address a specific problem given the large set of available texture operators. In this study we evaluate the optimization characteristic of an evolutionary framework to evolve solutions combining spectral and textural information in non-linear mathematical equations to improve multi-spectral image classification. Twelve convolution-type texture operators were selected and divided into three groups. The application of these texture operators to a multi-spectral satellite image resulted into three new images (one for each of the texture operator groups considered). These images were used to evaluate the classification of features with similar spectral characteristics but with distinct textural pattern. Classification of these images using a standard image classification algorithm with and without the aid of the evolutionary framework have shown that the process aided by the evolutionary framework yield higher accuracy values in two out of three cases. The optimization characteristic of the evolutionary framework indicates its potential use as a data mining engine to reduce image dimensionality as the system improved accuracy values with reduced number of channels. In addition, the evolutionary framework reduces the time needed to develop custom solutions incorporating textural information, especially when the relation between the features being investigated and the image textural information is not fully understood.