Research On Multispectral (Hyperspectral) Images Analysis Based On Projection Pursuit

The earth observation technology is critical for national future sustainable development. With the tendencies of "3-High" (high spatial resolution, high frequency resolution and high spectral resolution), Spectral remote sensing has become one major technique to achieve earth surface information and has become more and more important for our lives and social. In processing and analyzing multi-spectral (Hyper-spectral) images the sufficient spectral information of images provide beneficial conditions to detect and recognize objects. But they also present more requests on processing and analyzing multi-spectral (Hyper-spectral) remote sensing images. Nowadays the researches of Hyper-spectral remote sensing images are active and have achieved considerable successes. The application of new theories and methods on processing and analyzing multi-spectral (Hyper-spectral) remote sensing images should have significant meanings.Projection pursuit is a new technology to process and analysis high dimensional data. By designing projection index it projects high dimensional data set to low dimensional space to reveal the internal structures and characters of high dimensional dataset. In processing multi-spectral (Hyper-spectral) remote sensing images, an important idea is to reduce data dimension and data amount and at the same time to retain as more information as possible which contained in the original data sources. This just coincides with the heart spirit of Projection Pursuit. This paper applies Projection Pursuit to the analysis of multi-spectral (Hyper-spectral) remote sensing images. The research work is represented as follows.1. Projection Pursuit is applied to explore the potential structures and characters of the multi-dimension data through projecting the high dimensional data set into a low dimensional data space while retaining the information of interest. While in processing and analyzing spectral images a common used method is to reduce the data dimension. Combined the two methods reasonably this paper developed the Projection Pursuit method based on global optimize algorithm used to the analysis of multi-spectral (Hyper-spectral) images.2. Analyzed the significance and possibility of band selection in processing multi-spectral (Hyper-spectral) remote sensing images. Because of the character of strong correlations among spectral bands, the bands selection of spectral images can be seemed as variable selection of general multivariate data analysis. This paperdeveloped the band selection method based on self-adaptive subspace decomposition. The advantages and effectiveness of this algorithm are verified by the experiments on classification of multi-spectral (Hyper-spectral) remote sensing images.3. Principal Component Analysis (PCA) is used widely in multi-spectral (Hyper-spectral) remote sensing image processing and analyzing. There are two methods to compute PCA. One is Projection Pursuit which based on genetic algorithm to compute PCA. The other uses the general analytical method to compute PCA. After analyzed the general meaning of Projection Pursuit and compared the two methods this paper developed the standard genetic algorithm and proposed optimize algorithm based on dynamical evolution to finding the optimal projection index. Let the information divergence as projection index, the advantages of this method on extraction high dimensional non-normal distribution data set are proved by the experiments on analog data and multi-spectral images data.4. Linear spectral mixture model is a most used method on mixture pixel separation. This paper assumed the multi-spectral (Hyper-spectral) remote sensing image as observed signals and was generated by mixture of various independent signals. Because the mixture signals are more close to normal distribution than each independent signal this paper proposed mixture pixel non-supervised classification based on Independent Component Analysis (ICA). And negative entropy is used as estimation standard of dependent. The method is successfully used on analog data separation and multi-spectral images non-supervised classification.5. Based on the band selection results, the projection index function of high order statistics is selected and designed through Projection Pursuit based on genetic algorithm (dynamical evolution algorithm). This paper developed the zero threshold detection method to detect the abnormity objects in hyper-spectral images. At last the computation advantages of Projection Pursuit based on dynamical evolution algorithm is validated by experiments.Projection Pursuit technology combined the Projection Pursuit method with the field of processing and analyzing multi-spectral (Hyper-spectral) remote sensing images. There are a lot of theme should be researched in the future, including spectral matching, mixture spectral separation, spectral classification, spectral data compression and spectral object recognition etc.