Remote Sensing Image Fusion Based On Wavelet Kernel Filter And Sparse Representation

With the development of remote sensing technology, its applications areincreasingly being used in varying degrees in geology science, agriculture, meteorology,forestry, urban planning, environmental monitoring fields, et al. However, due to thelimitations imposed by remote sensors, remote sensing data obtained often do not reflectall of the information from the geographical area. To better understand the content of thearea, the image information is merged by the different remote sensor data. It has becomea very economical and effective solution. In recent years, information fusion technologyhas been introduced into the process of multi-sensor and satellite image fusion in orderto improve the interpretation ability of remote sensing images.In this dissertation, remote sensing image fusion is as the research background. Thefusion methods are proposed based on the multi-scale geometric analysis, machinelearning, evolutionary algorithm and other tools. These methods can achieve the taskand requirement of projects such as the National Natural Science Foundation, theNational ‘863’,‘973’ program and the The Fund for Foreign Scholars in UniversityResearch and Teaching Programs (the111Project), et al. The main work of thisdissertation is summarized as follows:1. Images are approximated by using the principle of support vector machine anddescribed as kernel functions, moreover, a new multi-scale transform tool namedwavelet kernel filter (WKF) is proposed. WKF is applied in bridge classification,MSTAR recognition and SAR image de-speckling. Results of bridge classificationand MSTAR recognition validate that the WKF coefficients of images contain themore image information, and the application of speckle images also show thetranslational invariance due to the transformation, for speckle ringing effectsappear almost clear.2. WKF is applied in remote sensing image fusion. It has multi-scale, translationinvariance and perfect reconstruction property. All the properties ensure that theWKF has advantages in image fusion. Considering the characteristics ofmulti-sensor image, the region energy maximum value is applied as fusion rule,where WKF is used to extract the fusion features. Fusion effect is carried out insome multi-scale geometric analysis tool wavelet, such as non-sampled wavelet,contourlet, non-sampled contourlet. Four group fusion results on the multi-sourceimage fusion library from University of Manchester validate that application of WKF in image fusion is effective to obtain better fused results. The fused resultscan avoid ringing effect and retain the detail information. For the fusion problemof multi-spectral and panchromatic image, two fusion strategies are proposed, thefirst one is combined with traditional Intensity-Hue-Saturation transform, whichprocesses intensity component; the other one adopts ARSIS (Amélioration de laRésolution Spatiale par Injection de Structures) concept for enriching the detailinformation in missing multi-spectral image. Fusion results demonstrate showsthat both fusion strategies can obtain the multiple spectral images with highresolution. It is helpful for the subsequent operation.3. A new method is proposed by combining WKF and Particle Swarm Optimization(PSO) algorithm for remote sensing image fusion. The detail subbands andapproximation subband are extracted using WKF. The detail subbands from thedifferent images are fused according to the region energy maximum fusion rulesand the approximation ones are fused using the PSO. The fused image can beobtained by the inverse transformation of WKF. Experimental results validate thismethod is effective and achieves the better optimal fusion result. For the MS andPAN image fused problem, two strategies are given. The first one constructsmulti-WKFs through the change of parameters in WKF, ARSIS concept withclone selection algorithm (CSA) can find the optimal weighted values and thenobtain optimal fusion result. The second one attempts to find the optimal intensitycomponent by CSA. The fused results indicate both combing algorithms areeffective and can obtain better results.4. In recent years, with the development of sparse representation theory, the imageprocess methods appear more and more applications by sparse representation.Considering the coefficients of images will be obtained by sparse representation,and finish fusion by these coefficients, and then we apply sparse representation inimage fusion. According to features of multi-sensor images and the coefficientsobtained by sparse representation, obtain fusion results by five fusion rules andgive the comparison results. According to characteristic of MS image, superresolution method by sparse representation method is applied in fusing MS andPAN image, and obtain the better results.5. Then, a novel pan-sharpening method, GIHS-WKF-BEMD-TP, based on GIHStransformation and combined WKF with bi-dimensional empirical modedecomposition (BEMD), is applied in fusing MS and PAN images. The BEMD is a highly adaptive method; each intrinsic mode functions (IMF) from thedecomposition describes local salient information and produces better spatialresponse. Finally, a tradeoff parameter is used to control the fused image withboth high spectral and spatial performance in terms of quality indices and visualeffect effectively.