Creating Spectral Words for Large-Scale Hyperspectral Remote Sensing Image Retrieval

Content-Based Image Retrieval (CBIR) for common images has been thoroughly explored in recent years, but little attention has been paid to hyperspectral remote sensing images. How to extract appropriate hyperspectral remote sensing image feature is a fundamental task for retrieving large-scale similar images. At present, endmember as hyperspectral image feature has presented more spectral descriptive ability. Visual words feature is a feasible method to describe image content, which can achieve scalability for large-scale image retrieval. In this article, spectral words are created for hyperspectral remote sensing image retrieval by combining both spatial and spectral information. Firstly, spatial and spectral features are extracted respectively using spectral saliency model and endmember extraction. Then a spectral vocabulary tree is constructed by feature clustering, in which the cluster centers are considered as the spectral words. Finally, the spectral words are compared for finding the similar hyperspectral remote sensing images. Experimental results on NASA datasets show that the spectral words can improve the accuracy of hyperspectral image retrieval, which further prove our method has more descriptive ability.