The Clustering Research On Oil Spill Hyperspectral Image Based On ICA

With the development of society, the demand of energy continues to grow. As a non-renewable resource, the distribution of oil is extremely unbalanced, its demand grows constantly in social environment and supply grows slowly, oil prices remain high. The uneven oil distribution,the vast ocean area and the rich submarine energy make the offshore oil drilling and ocean transportation actively, and lead to marine oil spill occur continually.At home and abroad, remote sensing method is used to detect accidents of marine oil spill. Remote sensing image of aviation hyper-spectral has advantages with more band, high resolution,rich data and strong timeliness etc, which make it have a higher research and application value in marine oil spill detection. This paper takes the accidents of Qingdao oil spill and Dalian oil spill as background and study the ground features classification and data dimension reduction of aviation hyper-spectral oil spill image.In this paper, the main work is as follows.Firstly, the commonly used clustering methods is theoretically studied and the results of them are compared. Then, considering spectrum curve characteristics, two new clustering methods are proposed to enhance the classification or clustering effect of hyper-spectral image. One is fuzzy clustering method based on the difference of standardization bands. The other is clustering method based on terrain curvilinear characteristics.Secondly, to solve the problems of large amount of data caused by large number of bands and a lot of stripe noises in aerial hyper-spectral images, independent component analysis method is used to extract features, reduce space dimension of hyper-spectral images and separate the characteristics of noises. Then, in order to improve the efficiency of the image analysis and processing, clustering method is used on effective features wavelengths to classify them.Finally, aerial oil spill hyper-spectral images captured after the oil accidents in Dalian and Qingdao are used to test the methods. Preprocessing is conducted by using reflectivity inversion, then principal component analysis method is used to reduce space dimension of hyper-spectral images and separate the characteristics of noises. In the end, proposed clustering methods are used to classify the images and classification system for oil spill hyper-spectral images is established.Experiment results demonstrate the effectiveness of the proposed clustering methods.In a word, the classification system and classification results illustrate the feasibility of the proposed methods on hyper-spectral image.