Feature-level Change Detection Research Based On Spectral-textual Information Integration In SAR Image

Synthetic Aperture Radar (SAR) is of important use for military and civil activity.Change detection in SAR images is a technology, which can determine the changecharacteristics and processes of surface features by analyzing quantitatively SARimages of the same region obtained at different times. With the rapid development ofSAR imaging technology, change detection in SAR images has become an importantresearch direction.Through the research and analysis of problems existing in the traditional methodsof change detection, focusing on the characteristics of multi-resolution and full ofinformation in SAR images, a change detection algorithm for SAR image based on thecombination of NonSubsampled Contourlet Transform(NSCT) and kernel method isproposed in this paper. In the time-frequency domain, NSCT has unique multi-scaleanalysis technique, which uses the nonsampled pyramid decomposition and thenonsampled direction filter for multi-scale detailed analysis. The algorithm firstestablishes spectral kernel in NSCT domain from low-pass sub-band as well as textualkernel from high-pass sub-bands at each scale to construct spectral-textual compositekernel by feature-level fusion, and then uses the support vector classifier (SVC) basedon the difference kernel to obtain the change detection map at each scale, finallyinter-scale fusion is performed to obtain the final decision-level change detection resultwith the speckle noise reduced significantly.Many real SAR image datas for different scenes from different regions are used inthis paper, Using these experimental datas, contrast to the kernel method based onGabor transform, according to the experimental effect, the proposed method based onNSCT has more accurate and faster change detection effect than based on Gabortransform. The algorithm is fast, robust and effective for change detection.