Remote sensing images are usually affected by light

angle

and other factors

thereby resulting in nonlinear mixed characteristics of a surface object spectrum. Thus

linear methods

such as Independent Component Analysis（ ICA）

have limitations. Kernel Independent Component Analysis（ KICA） achieves nonlinear transformation through the kernel function

and the data are mapped into a high-dimensional feature space

where the data are analyzed by ICA. As a result

detection errors from nonlinear mixing of the surface object spectrum are considerably reduced. Remote sensing images are usually large and complex. If they are analyzed directly by KICA

the computation will be large. Therefore

we propose a change detection method of remote sensing images of land cover based on multi-scale geometric analysis and KICA.First

multi-scale decomposition of remote sensing images is conducted by using multi-scale geometric analysis methods

such as contourlet transform

complex contourlet transform

and Nonsubsampled Contourlet Transform（ NSCT）. The decomposed components are transformed into partitioned vectors

which consist of low-frequency and high-frequency components. The vectors are then analyzed by KICA and mapped into a high-dimensional feature space by the kernel function

so that the mixing pattern of vectors is linear. In the space

mutually independent components are separated by ICA. The change component of land cover is selected and transformed into an image component. The change image is transformed into a binary image by using automatic thresholding method

and the final change detection result is obtained.Experimental results of the proposed method and recently proposed methods based on ICA

KICA

as well as wavelet and ICA are presented. Analysis and quantitative comparisons are conducted. Based on subjective visual effects

the isolated points and discrete regions in the results obtained by the proposed method decreased compared with those obtained by the methods based on ICA

KICA

as well as wavelet transform and ICA. The land edge is fully retained in the case of less isolated points

thereby reflecting more accurately actual change information of the surface. According to objective quantitative indicators

such as erroneous error

omission error

overall accuracy

and running time

the proposed method is more accurate than the methods based on ICA

KICA

as well as wavelet transform and ICA. The overall accuracy of the method based on NSCT and KICA is the highest

whereas the method based on contourlet transform and KICA shows a relatively high computational efficiency.Our method can separate change information of remote sensing images better. The method based on NSCT and KICA exhibits the smallest misjudgment and misdetection errors and preserves edge details better. The method based on contourlet and KICA shows relatively high detection efficiency.