Abstract:

The fusion of multispectral and synthetic aperture radar (SAR) images can retain the advantages of each data and improve the accuracy of land cover classification. However, some current image fusion methods cannot fully utilize the spectral information and texture details of the original data. In order to overcome these problems, a fusion method based on space-spectrum information collaboration and Gram-Schmidt transform is proposed. In the proposed method, Sentinel-2A images and GaoFen-3 (GF-3) images are preprocessed by different methods. Since the gray co-occurrence matrix can effectively extract the texture information of the image, it is applied to the Sentinel-2A image to extract the structural features, and the multispectral image coordinated by the space-spectrum information is fused with GF-3 image by the Gram-Schmidt transform. Principal component analysis (PCA) and the traditional Gram-Schmidt transform are used as the comparison methods in this experiment. In order to determine the effectiveness of the fusion algorithm, this paper uses five evaluation indicators including average gradient, spatial frequency, mean, standard deviation and correlation coefficient to measure the quality of the fusion image. In addition, due to its excellent training speed and excellent classification performance, random forest is used for land cover classification. The classification accuracy of random forest, Kappa coefficient and classification result graph are used as the evaluation criteria of the fusion method. Experimental results show that, compared with the original Sentinel-2A alone, the proposed fusion method can improve the overall accuracy by up to 5%, and has the potential to improve the accuracy of land cover classification in remote sensing satellite images.

Key words: image fusion, classification, multispectral, remote sensing