Feature Extraction of Hyperspectral Image Structure Based on Spatial-Spectral Fusion

摘要

Hyperspectral images (HSIs) contain abundant spectral information and spatial information. Fully mining the hidden information of the data itself can effectively improve the classification accuracy. For this reason, a spatial-spectral fusion classification method combining the spatial structure characteristics of data is proposed in order to improve the conventional spatial-spectral fusion method. The spatial structure information of the data in the spatial domain is extracted through Fourier transform, so that it has position invariance. With this operation, the comparability and analyzability are increased. Then the cross convolution is used to improve the response of different ground objects to features Then the spatial-spectral fusion method is used to fuse the original spectrum information to classify. Tree model is picked as the method of classification, in consideration of its strong interpretability and good selectivity for structure feature. Experimental results show that this method can refine the extraction of spatial structure information in the spatial domain. When this method is applied to HSI dataset, the classification accuracy can be significantly improved even with a small number of samples. Three types of HSI datasets (Indian Pines, Salinas, Tea Farm) have 12.69%, 9.09%, and 4.23% accuracy improvements combining with spectral information respectively.