Remote sensing imagery segmentation based on multi-objective optimization algorithms

Remote Sensing (RS) has been used to obtain relevant information about objects without the explicit necessity to stay in contact with them. RS collects measured data from the emanated energy of the surface of the earth. This process aims the construction of knowledge-based systems to identify interesting geographic features automatically. In RS, multispectral image segmentation is one of the most widespread methodologies for information extraction, using schemes comprising a wide variety of hard and soft grouping mechanisms based on different non-standard similarity measures making the classification problem to be application dependent. This procedure uses the spectral information contained in an image to recognize regions of interest. The segmentation of multispectral images is usually conducted by performing segmentation over a specific band according to the application. However, the segmentation of a specific channel might not perform well on the other bands of the image. This paper proposes a general scheme for multispectral imagery segmentation using multi-objective evolutionary algorithms (MOEAs) to identify thresholds encoding the best trade-offs between the segmentation criteria of various channels of the multispectral image. An evaluation of the performance of the proposed methodology is presented over a multispectral benchmark set composed of different images complexities and compared with several multi-objective algorithms.