Land Use Information Extraction Technology Based On Hyperspectral Remote Sensing

The strict management of land is related with the safety of food supplies, the sustainable development of economy, and the social stabilization. The State Department has set the land as the most important macro-economic management method. Management lies on the decision-making, and the correct policy depends on the exactly and timely information. Meanwhile, the information is also the feedback of the policy performing. The land use investigation and monitoring based on remote sensing data are the major way to acquire land information, inspect the land management, and acquire the feedback of the land-policy performing.Hyperspectral remote sensing, especially, airborne hyperspectral remote sensing plays an important role in the international remote sensing research. Hyperspectral remote sensing is mostly used in the geology, ecology and vegetation study at present. The classification methods of hyperspectral data are mainly based on the spectral characteristics analysis at pixel level, and there are seldom methods at object level. There are also less study of information extraction based on the fusion of hyperspectral data, any other remote sensing and non-remote sensing data, and the existing study mainly focused on the classification rather than change information extraction. In this dissertation, we used the hyperspectral data acquired by airborne imaging spectrum instrument OMIS-1, which was developed by the Shanghai Institute of Technical Physics, Chinese Academy of Sciences, and took Yixing city, Jiangsu province as the experimentation area to analyze the method to extract land use information. The preprocessing and bands selection of hyperspectral data, classification methods at pixel and object level, and the change information extraction methods based on OMIS-1 hyperspectral data, SPOT-5 panchromatic data and land use database were studied respectively.The major results of this paper are:(1) The theoretical models of bands selection include the models based on the statistical characteristic of remote sensing data and the models based on the land surface spectrum separability. Based on the statistical characteristic of remote sensing data, we put forward a method that uses dark current to estimate the ratio of signal to noise. This method is a simple and practicable method for the image with dark current. For the band selection, it is optimal to integrate the band selection theoretical model and visual evaluation.(2) The reflectance conversion is to calibrate the image through atmospheric correction. So far, there have been many calibration algorithms for hyperspectral data. We adopted the linear empirical calibration model to retrieve the reflectance from hyperspectral data. The visual and quantitative evaluation results show that the reflectance retrieved by this method can satisfy thedemands of ground spectrum rebuilding.(3) There are obvious distortions such as radiometric distortion, random points noise and stripe noise in the hyperspectral image in the study area. Based on the analysis of distortion characteristics, different methods were used to eliminate the distortions and to optimize the radiometric quality of the image. Considering the fact that the radiometric distortion does not always appear as low-brightness values at edge area and high brightness values at middle area for the same ground features, we put forward a new concept, geometric radiometric distortion. We also developed a simple, practicable elimination method based on low pass filter. This method resolves the problem that the model is too complicated, and the parameters are difficult to acquire in other methods.(4) We first analyzed the information extraction methods based on spectral characteristic at pixel level, and then we introduced the object-oriented method based on the fractal theory into analysis of hyperspectral image. The overall accuracy of decision tree classification at pixel level is 73.8%, the Kappa coefficient is 0.6480, and there is large confusion among various classes. Comparing with the decision tree method, the object-oriented method can obviously reduce the confusion, with the overall accuracy 94.5% and Kappa coefficient 0.9295. It demonstrates that the object-oriented method can satisfy the demands of land use classification, and can be used as a new technique in land resource survey.(5) In order to acquire the change information, especially the new built-up information distinctly, the multi-source data fusion, color composition, and abnormal spectral detection technology were adopted to process the multi-temporal OMIS-1 hyperspectral data and SPOT-5 panchromatic data. Meanwhile, we also studied the change information extraction method based on the OMIS-1 data and land use database. These methods expand the abnormal spectrum detection technology, enrich the technically system of hyperspectral remote sensing application, and provide the technical support for the new land resource survey. IHS transformation fusion and simple color composition method can only be applied to the 3 or 2 bands multi-spectral image. When these two methods are applied to hyperspetral data, the result image couldn't keep the spectral wave shape, so, they couldn't be used to automaticly extract the change information and the change information can only be extracted by visual interpretation. While the principal component transformation, multiply fusion and Brovey transformation are not limited by band number, and the fusion data can keep the spectral wave shape of the hyperspectral image, these three methods can be used to automatically extract the change information. The commonly used change information auto extraction methods include spectral angle mapping, binary encoding matching, spectral matched filtering and so on. We adopted these methods to extract the change information from the fusion data produced by the principal component transformation, multiply fusion and Brovey transformation respectively. Theresults show that the multiply fusion is better. Although the old land use vector data and new remote sensing data are widely used to extract change information, the precision is limited by classification accuracy of remote sensing data and the precision of land use database, and there are always false changes included in the change information. So, the extracted change results have to be testified through investigation on the spot.