Hyperspectral imagery principle component extraction method based on SOFM neural network with region feature spectrum

Unsupervised classification algorithms used for principle components extraction in hyperspectral imagery are sensitive to noise and data distribution, which result in low accuracy. This paper presented an unsupervised classification algorithm based on SOFM (Self Organizing Feature Mapping) neural network with region feature spectrum. In place of pixels spectrum, the algorithm firstly extracted region feature spectrum obtained through spatial neighborhood clustering as training samples, and then acquired the classificatory character adaptively by training SOFM neural network. Finally, principle components were successfully extracted based on spectral vectors clustering. Considering insensitivity of region feature spectrum to noise and self-organizing abilities of the neural network to cluster accurately, this method can enhance the extraction precision in evidence. The experiments on the hyperspectral imagery of Shenzhen Red-forest show the numbers of the training samples are reduced by about 95%. Moreover, this method can accurately extract principle components obviously better than K-means algorithm does.