On the detection of land cover change using fraction images

The aim of this study consists in investigating a new methodology to detect changes in land cover that have occurred over a given period of time, based on the concept of mixture pixels. This approach allows the interpretation of remote sensing digital image data at sub-pixel level. It is expected that this feature will create the right conditions for a more accurate identification of changes in land-cover as compared with the results obtained directly from multispectral image data. The Mixture Models allow the creation of fraction images, which convey the information regarding the proportion of each component class within any pixel. By subtracting fraction images associated with the same component at two different dates, fraction-change images can be obtained. As the magnitude of the changes may be either positive or negative, each component will originate two fraction-images, one displaying positive changes in the magnitude of component´s proportion and the other the negative changes. Therefore 2m fraction-change images will be generated, m being the number of component classes. The proposed methodology to detect land-cover changes is then performed by analyzing the fraction-change image data in this 2m-dimensional space.