Unsupervised change detection in remotely sensed images.

摘要

This study proposes a method to utilize remotely sensed pre- and post-disaster imagery data in order to detect the change specifically associated with structural and major regional damage caused by natural disasters such as a strong earthquake. The input is a pair of coregistered remotely sensed images of the same scene acquired at different times and the output is a binary image in which ‘changed’ pixels are separated from ‘not-changed’ ones. Correlation analysis generally fails to detect structural change, especially if images are acquired under different illumination conditions. In fact, automated detection in such a case becomes problematic since making distinction of change due to structural damage from that associated with the difference in the illumination condition is difficult. To overcome this problem, a method of principal component analysis (PCA) is employed.; Using PCA enables one to quantify changes proportionally to the actual change in a sequence of remotely sensed images; however, it requires the analyst to set a threshold on change measure that in fact is a linear product of the second principal component. In order to make the procedure entirely unsupervised, a probabilistic method by assuming Markovianity property in pixel class assignment is employed based on which a code in MATLAB image processing toolbox is implemented. It is further concluded that for disaster management purposes, regional change detection is more practical and precise compared with change in individual structures. The success in computing regional change in imagery data is a very strong step to take and should be extendable to compute change in higher resolution imagery data. The proposed approach produced promising results on the model images and also on the real sample images from Turkey earthquake.; This study further implements a protocol by which detecting change, using data fusion techniques in individual structural becomes feasible. It is important to note that for the purpose of this study structural change refers to any major change in geometrical shape of a structure that can be translated to change in pixel intensity values.