A fast classification method for hyperspectral imagery based on SOFM neural network

Rapid and accurate classification of hyperspectral images is a key technology in remote sensing image processing. The concept of region feature spectrum (RFS) was presented and RFS was extracted from hyperspectral image by the neighborhood clustering method. A fast classification method for hyperspectral image was proposed by taking RFS as input of self organizing feature mapping (SOFM), which can reduce training samples of neural network and suppress noise by RFS instead of single pixel spectrum. The simulated results of image data from the spectropholometer of type AVIRIS show that the classification accuracy of the RFS-SOFM is higher than those of the SOFM neural network and K-means, the corresponding computation cost increases by 63.6 percent compared with that of the K-means and decreases by 94.1 percent compared with that of the SOFM.