The Hyperspectral Image Noise Assessment And Data Dimensionality Reduction Method

Because of the large quantity of hyper spectral image data, it is necessary todimension data before target detection and classification. In this process, we should tryour best to retain signal and compress the noise. Noise assessment to hyper spectralimage can provide reference for the following image processing. At the same time, italso supports the credibility of the evaluation to the result of target detection andclassification.Because of the large quality of hyper spectral image bands、strong correlationbetween adjacent ones, redundancy is inevitably in the observed data，therefore，weneed to reduce the data’s dimension. Through using low dimension data to express theinformation of high-dimensional data effectively, it is more conducive to data’svisualization、understanding、interpretation, and improve operational efficiency toreduce error. Base on the development and research status of hyper spectral image, thispaper carried on a systematic study to noise assessment and dimension reductionmethods of hyper spectral data. The completed work and achievements are as follows:1、Discusses and implies the noise assessment methods that commonly used inprocessing a variety of remote sensing image, such as: uniform area、geologicalstatistics、 spectral and spatial de-correlation method (SSDC). At the same time，I tryto improve the SSDC algorithm, and compare the applicable range、advantages anddisadvantages of various assessment methods. SSDC noise evaluation method is moresuitable for hyper spectral data.2. Discusses and implies dimension reduction methods commonly used in avariety of hyper spectral image data, include independent component analysis,principal component analysis, the maximum noise fraction (MNF). Implies the MNFmethod base on SSDC noise assessment Finally, it carries on a systematic analysis andevaluation to variety of dimension reduction methods. and obtain the advantages、disadvantages and applicable range of different dimension reduction methods.3. Carry on a procession of denoising for hyper spectral data by using inverse MNF transform, and ideal results are obtained.