, pp.143-149 http://dx.doi.org/10.14257/astl.2016.123.28 Development and application of hyperspectral image classification technology Jiang Wei 1, Su Wunhai 2, Fan Yongcun 2, and Li Jian 3 1 Harbin Finance University 2 Northeast Agricultural University 3 Harbin Finance University Harbin, China Abstract. The hyperspectral image is applied to many fields for the features as high spectral resolution, multi-band, much information included and so on. Several facets are studied on the development of hyperspectral applications, the classification and recognition technology of hyperspectral technique. The breadth and depth of the application of hyperspectral technique was also presented that would lay the foundations for the further study. Keywords: hyperspectral; image classification; application 1 Introduction In 20th Century, it was a great-step forward for man that earth observation can be actualized from the space, the earth observation means the earth, where human live, can be observed from high attitude and the space by remote sensing. In recent decades, a new era is now dawning upon the remote sensing which is marked by hyperspectral remote sensing technology. Hyperspectral remote sensing technology is currently of great use for many fields such as geologic survey, land resources survey, city monitoring and planning, vegetation analysis, marine monitoring, environmental protection, oil and gas exploration etc. [1, 2]. Technologies of hyperspectral remote sensing imaging and spectrum are combined together. When the spatial image of target is generated, each spatial pixel formed dozens or even hundreds of narrow bands. The acquired image contains rich spatial, radiative and spectral information. Therefore, the hyperspectral images can play an important role in classification and recognition [3]. The purpose of this paper is to combine with the results of latest hyperspectral technology research, summarize and analysis system for application of hyperspectral technology in the many fields and classification technology. Supported by Educational Commission of Heilongjiang Province of China ((Grant No.12521004)) ISSN: 2287-1233 ASTL Copyright 2016 SERSC
2 Application advances of hyperspectral remote sensing technology Hyperspectral images with high spectral resolution can provide more ground information, which has been paid widely attention at home and abroad. A. Geological field Since the mid-seventies of the twentieth century, researchers around the world have done a lot of research and applied basic test of rock and mineral spectroscopy system. Research shows that there are a series of spectral features of mineral between0.4~ 2.5 μm which can be used in identification. Zhang (2005) analyzed of various moss in the short wave infrared spectra and found that they were similar. So all kinds of moss can be attributed as a terminal element, when the 5 kinds of moss s spectrum between 2000 and 2400 wavelength are standardized to get more precise information of quartz deposit using spectral mixture analysis (SMA) method. SMA results show that the spectral correlation coefficient R2 of quartz deposit is higher than 0.9 after being standardized with grace and Qing Tai [4]. According to the "863" project, remote sensing archaeology and integrated geophysical exploration technology, hyperspectral remote sensing imaging was researched in the mausoleum of the first Qin emperor in 2003 with the OMIS2 imaging spectrometer. At the same time, spectra and temperature calibration and archaeological survey were processed. With the aid of spectral analysis and image processing technology, the archaeological information of study area could be extracted and analyzed well [5]. B. Areas of land resources survey From beginning of 80's in 20th Century, the data of resources satellite has been used in dynamic remote sensing monitoring of land and resource survey many times in China. From 1999, the Ministry of land and resources have been carried out monitoring the land using of the cities whose population is more than 500,000 in China[6]. There are some representative studies in the research of the land use/land cover dynamic monitoring as follow. Africa vegetation index with multi-temporal, AVHRR, was used in monitoring land cover by Tucker (1994). The NDVI feature of TM was used in land use monitoring in Egypt by Lenney whose classification accuracy was beyond 95%[7]. Land use / land cover classification is emerging with the development of remote sensing technology for land classification concept.using The TM data in Shangjing town of Fuqing city, the land use type spectrum of knowledge and the changes of land use type were used by Yang Cunjian and Zhou Chenghu (2000) to carry on land use / land cover classification, combined with prior knowledge of, which achieved higher accuracy than the conventional maximum likelihood method. Automatic update of the land use/ land cover remote sensing monitoring was tested [8]. Luo Jiancheng and Yang Yan presented RSIGIM classification system that was applicated 144 Copyright 2016 SERSC
in the land cover change monitoring of the Hongkong area and provided a new idea for remote sensing classification[9]. C. Vegetation analysis In China, great progress has been made in recognition and classification of different vegetation and crop with using hyperspectral remote sensing technology. Hyperspectral remote sensing data was used by Liu Liang in the identification and classification of the crops whose classification accuracy was more than 95% in statistical analysis [10]. Vegetation index, one of the important parameter of vegetation monitoring application, is widely used in vegetation information. The extraction of vegetation information using the NDVI threshold method was used by Wu Jian to compress the hyperspectral remote sensing image using minimum noise fraction and hyperspectral vegetation classification method of spatial information, in ally completes the vegetation classification of study area was made in this way. The results show the average classification of each vegetation type precision can reach 90.3%. The average classification accuracy obtained by the maximum likelihood method is only 70.0%[11]. Hyperspectral monitoring in plant diseases and insect pests is based on measurement of biochemical composition and chlorophyll content of plants. The spectral prediction model for chlorophyll content of corn was constructed and verified by Zhang Dongyan. with acquired the hyperspectral image of corn canopy planted in pot and field were got by field scanning imaging spectrometer. The results show that there is certain application potentiality the imaging spectrometer at the micro scale, have certain of crop spectral information detection [12]. Crop growth monitoring is very important for the protection of food security and agricultural yield estimation. Through the comparison test of the cotton s density and moisture in Xinjiang, Jin Xiuliang analyzed the correlation of different density and water treatment in the whole growth period of cotton biomass and high spectral characteristic parameters. The result of study show that the mathematical relationship between the cotton growth parameters and hyperspectral characteristics can be reflected by growth monitoring model based on hyperspectral characteristic parameters[13]. D. Marine, environmental monitoring Imaging spectral technique has wide application prospect in the aspects of marine disaster and environmental monitoring. The spectral characteristics of the sea is one of the most important research contents of application of hyperspectral remote sensing in marine remote sensing. Through the study of ocean spectral properties, we can more accurately understand sea spectrum structure and identify the spectral characteristics of different components in seawater. Li Hongbo et al. PHI imaging spectrometer data image of our country obtained are analyzed, results show that both from the spatial resolution and spectrum resolution can clearly distinguish the oil. The coverage area and thickness of oil pollution can be preliminarily estimated because of the gray level difference in the image. Consequently, the contamination degree of sea Copyright 2016 SERSC 145
water can be known by the analysis [14]. The research on red tide has been developed by some domestic scholars. Hyperspectral remote sensing with unique hyperspectral high-resolution can be used in effectively identification of the sediment of water body and concentration of pollution. Hyperspectral remote sensing has the unique advantage in the environmental investigation and monitoring. Based on the spectrum characteristics, the visible and near infrared image of airborne imaging spectrometer were applied in the city environmental comprehensive evaluation by Ben-Dor (2001) [15]. The air pollution of Beijing area was monitored according to the aerosol optical depth of satellite remote sensing monitoring image by Li Chengcai et al., which provided satisfactory monitoring results and a new technology means for the study of air pollution g [16]. 3 Hyperspectral remote sensing classification and recognition A. Maximum likelihood classification Maximum likelihood classification is the most widely used classification method of the traditional remote sensing image classification that is based on statistical model [17]. The method of maximum likelihood classification shows the high speed characteristics in the classification of wide band remote sensing image, but it would have a significantly reduced classified speed in classification of hyperspectral remote sensing image that due to big size data. For the defects of the maximum likelihood classification in classification of hyperspectral data, some improved maximum likelihood methods have been researched. Jia and Richards (1994) found that the grouping high spectral data could present higher accuracy and efficiency and lower cost of classification in classification. Furthermore, the discriminant function and covariance matrix could be improved to be used in hyperspectral image classification which would work well with higher efficiency and accuracy [18]. B. Spectral angle mapper Being a rather peculiar method of high spectral data classification recognition methods, the spectral angle mapper is a kind of spectral matching classification technology, which involves analyzing the similarity of spectral angle between the image spectrums and the reference image spectrums vectors. The spectral angle mapping is used in many research fields by lots of Chinese scholars. The spectral angle mapping technology was used in recognition of hyperspectral image data before and after radiating calibration by Liang Xinlian et al. The study show the results of recognition of data before and after calibration basically coincide [19]. The accuracy of spectral angle mapping is higher than the accuracy of traditional classification methods. Its accuracy will get improvement when the spectral angle mapping is improved. The experiments of the spectral angle mapping method combined with decision tree were testified that presented higher accuracy than maximum likelihood classification. Another classification method of spectral irradiance direction 146 Copyright 2016 SERSC
combined with the spectral angle mapper was used in classification research of alteration information of the Zhongpo and Bijia mountain in Xinjiang province whose result show that this method has high reliability. C. Neural network classification Artificial neural network (ANN) is a kind of imitation of human neural network, the mathematical model of distributed parallel processing algorithm. It has strong adaptive, self-organizing, self-learning and adult-tolerant ability. The classification accuracy can be improved by means of improvement of the neural network technique. The BP neural network used in analysis of remote sensing image will be better in accuracy and speed improved by involving momentum factor and bias vector[20]. The applicability of three layers MLP neural network and linear spectral mixture model in the ASTER and Landsat TM images was analyzed and compared by Weng and Hu (2008). The impervious layer of the city was synthetized with decomposed endmember abundance image. The results show that the use of neural network to extract impervious layer accuracy is better than that of the linear spectral mixture model [21]. a. Fuzzy classification Fuzzy classification, based on fuzzy set theory, is an object-oriented image classification method that is used in calculation of the subjection degree in all sets by using mathematical model. Then the attribution is determined according to the subjection degree Compared with the traditional methods of image classification, the object oriented classification method show better result in the classification of hyperspectral image and broadly application prospect. Li Xican established a general spectral inversion model of soil properties according to the theory of fuzzy pattern recognition whose application results show that the model is of high accuracy [22]. Fuzzy recognition can be applied to the investigation of land, crop identification and soil monitoring. For example: the fuzzy classification is used in investigation of use or cover change land. Through the actual verification, this method show better classification effect in the complex region and worth studying. b. Support vector machine classification In addition to the above classification and recognition technology, the SVM method for remote sensing image classification is proposed in recent years. SVM is a machine learning method based on the statistical learning method whose biggest property is strong generalization ability. Tan Kun designed a multi class SVM classifier based on separability measure two binary tree that showed higher accuracy in test than the conventional classification and the other known SVM method [23]. The SVM dynamic integration method is improved by Niu Peng et al.. The result show the highest accuracy in the classification of hyperspectral image compared with other SVM integration methods respectively and single optimized SVM. Palza and Sanchez (2011) decomposed mixed pixel of hyperspectral image with parallel computing respectively that brought the higher speed and better results [24]. Copyright 2016 SERSC 147
c. Classification accuracy evaluation The evaluation of hyperspectral classification accuracy refers to assessing the accuracy of classified hyperspectral image according to the referenced data of real ground. The CM or EM is often used in the accuracy of classification method. The assessment indexes of CM include overall accuracy, Kappa coefficient, mapping accuracy, user accuracy, misclassification error and omission error. Scholars have conducted for evaluation of the accuracy of the maximum likelihood method, spectral angle mapping method and hierarchical classification. The analysis show that the overall accuracy of maximum likelihood classification and spectral angle mapper classification is same basically, the accuracy of hierarchical classification method is higher than others. The stability of spectral angle mapper and hierarchical classification are better than the stability of maximum likelihood classification that is not good for fluctuates greatly. 4 Conclusion The development of hyperspectral data classification technology is developing well, but its development is slower compared with the development of hardware of hyperspectral remote sensing. Therefore, several aspects should be further researched and explored, including the data mining technology in the mass hyperspectral data, improvement and fusion traditional classification technology, new pattern of recognition classification for the characteristics of hyperspectral data. References 1. Tong, M., Xi, Q., Zhang, B., Zheng, L. F.: Hyperspectral Remote Sensing: Principle, Technology and Application. Beijing: Higher Education Press. 2006. p. 206. 2. Pu, M., Liang, R., Gomg, P.: Hyperspectral Remote Sensing and It s Applications.Bei jing: Higher Education Press. 2000. p. 1 3. Tong, J., Xi, Q., Zheng, L. F., Wang, J. N., Wang, X. J.: Study on Imaging Spectrometer Remote Sensing Information forwetland Vegetation. Journal of Remote Sensing. 1997.1(1): 50-56 4. Zhang, J., Benoit, R., Sa nchez-azofeifa, A.: Spectral unmixing of normalized reflectance data for the deconvolution of lichen and rock mixtures. Remote Sensing of Environment, 95: 57-66, 2005 5. Wan, J., Qingm, Y., Tan, K. L., Zhou, R. P.: Application of Hyperspectral Remote Sensing [M]. Science Press. 2005. 218-241 6. Liu, X. J. J., Tang, W. Z.: The application of remote sensing in a new round of the survey of land and resources. Chinese Remote Sensing ahead 20 years of innovation research papers. Beijing: Meteorological Press. 2001.126-132 7. Tucker, J., C J, Dregne H E, Newcomb W W. AVHRR datasets for determination of desert spatial extent. Int J Remote Sensing, 1994.15:3547-3565 8. Yang, J., Cun, J.,Zhou, C. H.: Extracting Residential Areas on the TM imagery. Journal of Remote Sensing.2000. (02): 146-150 9. Luo, J., Cheng, J., Yang, Y.: Remote Sensing Intelligentized Geo-Analysis. Remote Sensing Technology and Application. 2000. 15 (3). 199-204 148 Copyright 2016 SERSC
10. Liang, J. L., Guang, J. X., Bin, L.X.: Study on Classification of Agricultural Crop by Hyperspectral Remote Sensing Data [J]. Journal of Graduate University of Chinese Academy of Sciences. 2006, 23 (4): 484-488 11. Jian, J. W., Peng, D. L.: Vegetation classification technology of hyperspectral remote sensing based on spatial information. Transactions of the Chinese Society of Agricultural Engineering. 2012.28 (5): 150-153 12. Zhang, J., Yan, D., Liu, R. Y., Song, X. Y.: A Field-Based Pushbroom Imaging Spectrometer for Estimating Chlorophyll Content of Maize. Spectroscopy and spectral analysis. 2011. 31 (3): 771-775 13. Liang, J. J. X., Li, S. K., Wang, K. R.: Monitoring of Cotton Growing Based on Hyperspectral Parameters. Journal of Northwest Agriculture. 2011, 20 (9): 73-77 14. Li, J., Bo, H., Rong, S., Qi, S. Y., Pushbroom Hyperspectral Imager and It s Potential Application to Oceanographic Remote Sensing. Infrared Millim.Waves. 2002.(6) 15. Ben-Dor, J. E.: A Spectral based recognition of the urban environment using the visible and near-infrared spectral region(0.4~1.1mm).a case study over. Int. J. Remote Sensing, 2001. 22 (11): 2193-2218 16. Tang, J., Zhao, H., Lei, Y., Li, C. C., Gao, P. Q.: Atmospheric Correction for ETM+ Imagery Based on High Resolution Aerosol Optical Depth Retrieved from MODIS Data. Geography and Geo-Information Science. 2010. (4) 17. J. Mclver, D., Friedl, M.: Using prior probabilities in decision-tree classification of remotely sensed data.remote Sensing of Environment, 81 (2-3): 2002. 253-261 18. Jia, X., Richards, J.: Efficient maximum likelihood classification for imaging spectrometer data sets. IEEE Transactions on Geoscience and Remote Sensing, 41 (5): 1994. 1129-1131 19. J. Liang, X. Lian, Li, H. T., Zhang, J. X.: The possibility of applying pre-calibrated hyperspectral data before radiometric correction for classification An experiment of spectral angle mapping. Science of Surveying and Mapping. 2004. 29 (4): 37-40 20. J. Gong, P., Pu, R. L., Miller, J. R.: Coniferous forest leaf area index esti2mation along the Oregon Transect Airborne Spectrographic Image data. Photogrametry Engineering and remote sensing, 1995. 61 (9): 1107~1117 21. J. Weng, Q., Hu, X.: Medium spatial resolution satellite imagery for estimating and mapping urban impervious surfaces using LSMA and ANN. IEEE Transaction on Geosciences and Remote Sensing, 2008. 46 (8): 2397-2406 22. J. Yu, T., Li, X. C., Wang, X., Shang, X. D.: Application advances and outlooks of hyperspectral technique. Science of Surveying and Mapping. 2012. 37 (2): 115-118 23. Tan Kun, D., Research on Hyperspectral Remote Sensing Image Classification Based on Support Vector Machine. China University of Mining Technology. 2010 24. D. Gao, Heng, Z.: Research on Classification Techinique for Hyperspectral Remote Sensing Imagery. National University of Defense Technology. 2011. pp. 5-18 Copyright 2016 SERSC 149