Abstract: Hyperspectral remote sensing image classification is an important part of hyperspectral remote sensing image processing and application. However, hyperspectral remote sensing image has the characteristics of large number of bands and high spatial resolution, which brings some challenges to image classification. To improve the classification accuracy of hyperspectral image and make full use of the spatial and local information, a classification method based on guided filtering combined with local discrimination embedding is proposed. Firstly, the hyperspectral remote sensing image is normalized, the feature extraction is realized by principal component analysis, and the extracted first principal component image is used as the guided image. Secondly, the spatial characteristics of each band are extracted by using the guided filter. Then, the extracted spatial image features are superimposed,and low-dimensional embedding is completed by local Fisher discriminant analysis. Finally, the embedded features are input into SVM to acquire classification results. The experimental results of two hyperspectral images, Indian Pines and Pavia University, show that compared to other relevant methods, the proposed method can obtain higher classification accuracy. When 10% and 100 samples from various ground objects were randomly selected as training samples, the overall classification accuracy increased to 98.28% and 99.45% respectively. At the same time of low dimensional embedding, the proposed method can effectively use the spatial information of images and improve the classification effect.

Key words: hyperspectral image classification, guided filtering, local discrimination embedding, feature extraction, principal component analysis