The Research On Hyperspctral Remote Sensing Image Segmentation

The hyperspectral remote sensing image has many features, such as numerous channels, spectral resolution with high bands, large amount of data etc., although it brings us rich information, it also brings great challenge for data processing. How to reduce the noise of hyperspectral image effectively, reduce hyperspectral dimension, decrease the data volume is of an important issue in the hyperspectral image segmentation.Noise reduction aspects: This paper chooses two traditional frequency domain noise reduction methods: high-pass filter and low-pass filter to deal with the hyperspectral image. On the aspect of airspace noise reduction, the paper chooses the method based on mathematical morphology and designs the serial morphology filter method based on open and closed operation. The high-pass filter sharpens the image, highlights the image edges information, while retaining the most part of noise image is retained, but the result of this filter will be used in reduced-dimension operation. Low-pass filtering makes the image smooth and reduces the image's noise. The effect of morphological filtering is same with the effect of low- pass filtering; the image's noise is also reduced by image smoothing. But its smoothing effect is better than the low-pass filter.Dimension reduction aspects: Based on feature extraction aspects,the paper chooses two algorithm processing methods which are the Principal Component Analysis and Minimum Noise Fraction method. Principal Component Analysis is a powerful tool of multivariate data analysis, and it is widely used in hyperspectral image dimension reduction processing based on feature extraction. But this method is quite sensitive to noise, in the process of realizing dimension reduction, and also retains the part of the noise image information. Based on this, the paper proposes a minimum noise fraction transform, which is a kind of measure standards of linear transformation to maximize signal-to-noise ratio. It separates the noise in the process of realizing dimension reduction.Segmentation aspects: This paper takes the hyperspectral image segmentation algorithm as spindle which bases on mark watershed and briefly introduces the basic operation about the pretreatment of mathematical morphology. The paper put forward the segmentation technology of mark watershed algorithm by using threshold segmentation and k-means clustering image segmentation. This paper uses Dexing copper hyperspectral as research data and compares three segmentation algorithms and gives the superiority of segmentation algorithm which bases on mark watershed.The main innovation in this paper is putting forward the method of segmentation through selecting the main component feature of the image, which is first processed by noise reduction, and then by dimension reduction, presenting the technology of noise reduction based on mathematical morphology, and applying the maker watershed algorithm to the segmentation of hyperspectral image successfully.