Multi-Channel Remote Sensing Imagery Change Detection Based on Orthogonal Transformations

An approach based on canonical correlation analysis in multivariate statistics is introduced to change detection of multi-temporal/multi-channel remote sensing imagery. The basic idea is to take multichannel remote sensing imageries acquired at different times as groups of random multivariates, then construct linear combinations to explore correlations between them, thus finding out biggest differences over the time span. In our approach, MAD transformation is firstly conducted on original imageries to produce a difference image, then MNF transformation is utilized as a postprocessing step to separate noise from signal in difference images, and change information can be effectively concentrated into a few components of the final result. Another change detection method based on PCA is also described briefly for comparison. Experimental results of a case study using Landsat5 TM imageries are presented to demonstrate the effectiveness of our method. And the characteristics of correlation between results and original imageries are discussed in detail.