A study of image fusion applied in land cover classification

• Main Authors: Meng-Che Wu, 吳孟哲
• Other Authors: K.S. Chen
• Format: Others
• Language: en_US
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/91064278686944920619

碩士 === 國立中央大學 === 太空科學研究所 === 93 === There are two main topics will discuss in this paper, pixel-level image fusion based on Principle Component Analysis (PCA), and feature-level image fusion based on Dempster-Shafer evidence theory. In pixel-level case, the SAR image at HV polarization is relatively sensitive to the vegetation canopy. We combined the HV polarization information from SAR and spectral characteristic from SPOT images in an effort to enhance land cover classification. Before the fusion process, wavelet transform was first applied to denoise the SAR image which suffers from speckle contamination due to coherent process. The principle component analysis (PCA) is used to fuse the SPOT and SAR images. In so doing, the PC-1 component is replaced by SAR image (approximation image, after wavelet transform) and then the inverse transform is followed. At last, the maximum likelihood classifier was used for both SPOT-XS images and fusion images. In feature-level case, fully polarization information from SAR is used to combine with spectral characteristic from SPOT images, mainly to enhance land cover classification as well. We first denoise the SAR image by Lee filter. Next, the maximum likelihood classifier based on different distribution was used for SAR and SPOT images ( Based on Wishart distribution and multivariate Gaussian distribution respectively), to extract the conditional probability of each pixel for each class. Dempster-Shafer evidence theory is then applied, to combine the classified results of SAR and SPOT data. Experimental results show that the classification accuracy is dramatically improved by making use of the proposed methods above. Data fusion can take advantage of the use of complementary information to obtain a better overall accuracy than using single data source only.