Adaptive Threshold Hyperspectral Image Classification Based On Gauss Distribution

Hyper-spectral image classification provides a strong support for geological exploration, environmental monitoring and military reconnaissance. The existing K-MEANS and ISODATA have been used in a variety of image processing software. But these methods only consider the spectral information and ignore the spatial information, and they need to set the number of categories in advance. Therefore, in this paper, we propose a method of hyper-spectral image classification with an adaptive threshold based on Gaussian distribution.Considering the traditional threshold methods that rely on the one-dimensional histogram, and the threshold needs to be set manually, in this paper we propose a method of an adaptive threshold based on Gaussian distribution. This method is modeled as Gaussian distribution, and uses the superposition of multiple Gaussian distribution to fit the spectral angle histogram of the hyper-spectral images, and the censored data of peak district is analyzed according to the peak and valley character of the fitting curve, finally the threshold is determined adaptively. For adjacent pixels with a higher similarity possibility, we establish a dimensionality reduction model based on the minimum correlation window, which combines the spectral and spatial information. Then we regard the spectral angle which is less than the adaptive threshold as the similarity condition between pixels, finally the unsupervised adaptive classification of hyper-spectral images is finished.The simulation experiment and target recognition application demonstrate that our method is adaptive, and has advantages of high clustering accuracy, strong edge recognition and good robustness.