Application Of Hyperspectral Image Feature Extraction In Oil Slick Identification

Hyperspectral remote sensing technique is developing rapidly and has been applied into many areas such as agriculture and mineral. It is the result of this technology's own merits:multi-band, combination of image and spectrum, high resolution, rich data including multiple and narrow continuous spectrum of visible light, near infrared, SWIR and thermal infrared bands. The emergence of hyperspectral remote sensing predicts or distinguishes materials easily, while the materials could not be predicted or distinguished with wide-band remote sensing. Existing researches on oil spill identification are mostly based on the image data of infrared or near infrared, etc., while the ones based on hyperspectral image data are very few, so this research is of great value.There are many bands in hyperspectral image (some hyperspectral image are up to hundreds) so that the amount of data will be very large and also the data processing will also be very complex. Moreover, the information of some bands is extremely little or none and these bands may even interrupt follow-up processing. As a result, dimension reduction is one of the key steps of hyperspectral image processing. There are two ways for dimension reduction. One is feature extraction and the other is band selection. This paper conducts research on the hyperspectral feature extraction as well as classification and recognition of spilled oil thoroughly. The method of feature extraction based on GA-PCA and the oil classification method based on SAM-SFF are studied. Besides, experiments on spilled oil are carried out under existing equipment conditions in our laboratory and then spectral library of spilled oil was built. The feature extraction algorithm based on GA-PCA which imports genetic algorithm into the process of feature extraction and combines genetic algorithm and principal component analysis to implement dimension reduction. The oil classification method based on SAM-SFF which combines spectral angle matching method and spectral feature fitting method together to realize relatively good recognition effect.To evaluate the performance of GA-PCA and SAM-SFF, hyperspectral image obtained from Cuprite District, Nevada, USA and the oil images is used in the experiment. After feature extraction with the method based on GA-PCA, the classification accuracy of hyperspectral image in Cuprite District could reach 96.4706%, which increased 5.8824% higher than the accuracy just using PCA. On the basis of images from experiments of spilled oil, this paper validated the superiority of oil classification method based on SAM-SFF, whose recognition accuracy can reach up to 86.7%.