AN EFFICIENT MIXED PIXEL RESOLUTION MEHTODOLOGY USING BIO-INSPIRED HEURISTICS

Abstract

Remote sensing has been used for extracting information about regions or objects for a long time. The process of gathering information without coming in direct contact with the object, is advantageous in identifying even those areas which seem to be nearly inaccessible to humans for example in monitoring earthquake hit areas. These remote sensing systems record brightness values emitted from the object at different wavelengths that commonly include not only portions of the visible light spectrum, but also infrared and, in some cases, middle infrared bands. The brightness values for each of these bands are typically stored in a separate grayscale image (raster). However, certain regions have brightness value which is composed of brightness from more than one region. Pixels depicting brightness for such regions in an image are known as the Mixed Pixels. In the process of image classification, the mixed pixels pose a major problem. It becomes difficult to assign a accurate single class to these pixels. If these pixels are left as such then complete and accurate information may not be obtained from the remotely sensed data. Our work intends to solve the problem of mixed pixels in a remotely sensed image using recently (December, 2008) introduced bio inspired algorithm known as Biogeography based optimization. In last few years, bio-inspired methods have rapidly gained importance in computing due to the need for flexible and adaptable ways of solving engineering problems. They are a class of algorithms that imitate specific phenomena from nature. Bio-inspired algorithms are based on the structure and functioning of complex natural systems and tend to solve problems in an adaptable and distributed fashion. Several systems such as the antcolony system, bee foraging, bird flocking etc. have been used as the basis for developing models and algorithms to solve various issues. We have utilised the migration behaviour of a species residing in a habitat. Species immigrate to a new habitat based on several factors such as vegetation, rainfall, temperature etc. These Page | iv factors decide the fitness of a habitat for supporting the species residing in it. Based on this approach we find the fittest habitat for the mixed pixel and assign a class to it accordingly. The species are migrated taking into account the effect of external factors, in the form of sinusoidal migration pattern. A good candidate solution has relatively high emigration rate and low immigration rate, while the converse is true for a poor candidate solution. The results produced have been compared with the previous work to show the efficiency of the proposed algorithm. In previous work, linear migration has been applied for the mixed pixel problem. We have designed our algorithm based on sinusoidal migration. The sinusoidal migration curve has been modelled in two approaches. The first approach incorporates all the bands in a remotely sensed image. The second approach finds the non redundant bands and then algorithm is applied on the selected bands.