Investigation Of Hyperspectral Remote Sensing Data Classification Using Multiple Classifiers Combination

Since the advent of remote sensing in the second half of 20th century, nowadays there have been great changes in theory and technology. The advent of hyperspectral was one of the most significant breakthroughs in remote sensing. Hyperspectral remote sensing has higher spectral resolution as the same time retain higher spatial resolution, so its capability of distinguishing the different and describing the same ground objects in details enhanced greatly. Hyperspectral remote sensing was firstly used in geology and it was successfully used in ecology, atmospheric science, agriculture forest, etc. in 1988. Hyperspectral data has more bands and higher spectral resolution, so it makes future broad prospects of applications.There are two main kinds of extracting information from hyperspectral remote sensing image: based on feature space and based on spectral space. Using the method based on feature space the images should not need to atmospheric correct, but it will lead to curse of dimensionality and Hughes phenomenon for many bands. The solution methods usually are increasing sample numbers or dimension reducing processing. The former need huge sample number, so limted to a lot of human and material resources; the latter will lead to some information losed. Furthermore, it is hard to solve the mixed pixels unmixing. Therefore, the hypersepctral data should not be taken full advantage of using these methods. The methods based on spectral space are suceesfully used in hyperspectral data. Matched Filtering is one of the mature methods used in hyperspectral data classification. It classifies by measuring the similarity of the endmember and reference spectra, and it needs no sample data but the image should be atmospheric corrected. These methods based on the hypothesis that all spectra have been calibrated to apparent reflectance and the dark current of sensor and path radiation is removed. It is only the ideal state for these effects are hard to remove successfully, especially in low reflectance object, so this will lead to error using Matched Filtering. This thesis proposed the methods to improve the classification accuracy considering the advantage and disadvange of these two kinds of methods. One is to improve the classification accuracy of single classfier, including the atmospheric correction and extracting pure endmember; the other is combing multiple classifers. The main research results are as follow.1. The remote sensing image should be atmospherically corrected for quantitative application, while removing the adjacency effect is the important step in atmospheric correction. Removing adjacency effect depends on the knowledge of its main impact factors and mechanism. This thesis simulates the adjacency effect on different contrast targets in different conditions using MODTRAN model. The results show that visibility has great effect on adjacency effect; followed by satellite zenith angle; solar altitude angle is the least. When the visibility is higher, the curve of relatively variable percentage of radiance is more similar to the curve of background radiance. But in the lower visibility, this phenomenon is not appeared. When the visibility is higher, as the wavelength increases, the relatively variable percentage of radiance reduces gradually. But in the lower visibility, this relation has not appeared. It proves that as the visibility is reduced, its impact on adjacent effect is more complicated. This thesis investigates the main impact factors and mechanism of adjacency effect. It will provide evidence to remove the adjacency effect.2. This thesis review the method of hyperspectral remote sensing atmospheric correction. The hyperspectral remote sensing image was atmospheric corrected using ATCOR2, ATCOR3 and FLAASH and their results were compared. It concludes that pixel spectra of three result images are similar and the spectra of image corrected by FLAASH are smoother for using spectral polishing technology. Due to the higher different elevation, the corrected effect of ATC0R3 is not better.3. SAM, SFF and MTMF are the mature Matched Filtering methods. Firstly, the endmember spectra were acquired, and then the image was classified by comparing the pixel spectra and the endmember spectra. There are three kinds of methods to acquire the endmember: labortory spectra, field spectra and obtaining from the image directly. Acquiring endmember from the image directly is the easy way, and it is widely used in many researchs for it can prevent the atmospheric effect and scale effect. PPI is the commonly used algorithm to acquire endmember from the image directly.4. This thesis introduces geometrical, statistical and asymptotical properties of high dimensional spaces. As dimensionality increases, the volume of a hypercube concentrates in the corners and outside shell; the diagonals are nearly orthogonal to all coordinate axes. As dimensionality increases, the curse of dimensionality and Hughes phenomenon appear, so the required number of samples increases. However, as dimensionality increases, low linear projections have the tendency to be normal, or a combination of normal distributions. That means it is feasibility that the high dimensional data after dimension reducing processing can be classified using methods based on feature space.5. The vegetation classification, characteristic and growth vigour in large areas can be detected quickly using remote sensing. Hyperspectral remote sensing has potential application in ecology field for its higher spectrum resolution. Vegetation information as an important parameter of entironment evaluation can benefit for construction and supervision in region entironment. In this thesis, the vegetation information in Beiya region can be extracted by SAM algorithm using hyperspectral remote sensing image. The spectrum acquired by ASD handhold spectroradiometer in situ was used as reference spectrum. This thesis researches the main influence factors and its response measure of acquiring vegetation reflectance spectral and choosing the endmember spectrum. The classification result was made by SAM match and overall accuracy and Kappa coefficient was calculated. Finally the vegetation distribution map was made by projection transformation.6. The absorption vale in vegetation visible spectrum is caused by intensive absorption of chlorophyll, and its depth and shape are different in different vegetation coverage, so the vegetation coverage can be extracted by comparing the depth and shape of the absorption vale. This thesis extracts the vegetation coverage from the hyperspectral remote sensing data using Spectral Feature Fitting (SFF) method which is based on the spectral absorption feature. The reference spectral is acquired by the ASD Field Spec Pro handhold spectroradiometer in situ. The vegetation coverage was extracted using matched continuum removal curves of the image spectral and the field spectral using SFF method. The vegetation coverage image was outputted at last. Comparing with the result extracted by VI (Vegetation Index), SAM (Spectral Angle Mapper) and the filed data, they are consistent with each other.7. Matched filtering method is successfully used in information extraction from hyperspectral remote sensing image. However, it is hard to extract the low reflectance ground object due to atmospheric influence. In this paper, an improved method based on spectral clusting was used to extract low reflectance ground object. Firstly, using Pixel Purity Index (PPI) to find the endmember from hyperspectral image and computing the spectral angle between the pixel spectrum and each endmember spectrum, the pixel was classified into the endmember category with the smallest spectral angle. Then, the endmember spectral was clustered using K-mean algorithm. Finally, the endmember categories were clustered according the K-mean algorithm and the final classification result was projected and outputted. Comparing the result with the original image and field data, they are consistent with each other. This method can lead to the objective result without artificial interference. It can be used as an unsupervised classification method in hyperspectral remote sensing image.8. In order to obtain the best classification performance in pattern recognition, firstly, the data set should be classified using different classiers, then, choose the best classification result as the final conclusion. With the complexity of pattern recognition increased and novel algorithm developed, researchers find that although different classifiers have different classification performance, their misclassification set are not consistent with each other. That is, some sample misclassified by one classifier may be recognized by another classifier. Different classifiers are complementary to eacher other. If only the best performance classifier is choosed, some valued information of other classifiers may be ignored. In order to solve this problem, multiple classifiers combination was put forward. Aiming at the advantage and disadvantage of different hyperspectral classification methods, this thesis proposed the combination method based on decision tree, and extracted the rock body from the hyperspectral remote sensing image using this combination method. The study area located at the Pulan porphyry copper and gold deposits in southwest of China, and the composition of the deposits is mainly dioritic porphyrite. Firstly, dioritic porphyrite was extracted from hyperspectral remote sensing image using SAM, SFF and MTMF respectively. The image was atmospheric corrected before processing, and endmember was extracted by PPI algorithm from the intersection area of multisegmentation and geology map. Then, three classification results were outputted using combining classifiers by decision tree. At last, comparing all classification results and geology map, it is concluded that combining multiple classifiers has the best classification performance and SFF has the better capable of pixel unmixed than SAM and MTMF.In order to improve the classification performance of hyperspectral remote sensing, this thesis investigation mainly includes two parts, one is to improve the performance of the single classifier, and the other is to combine multiple classifiers. In the fomer, the thesis investigates the methods of atmospheric correction and extracting pure endmembers. Hyperspectral remote sensing atmospheric corrections are reviewed and different atmospheric correction model are compared. A novel endmember extracting method combining multisegmentation and geology map is proposed. In the latter, the extracting the low reflectance ground object based on spectral clusting from hyperspectral remote sensing image and combining multiple classifiers based on decision tree was proposed. Comparing with the field data and geology map, it concludes that these two methods are effective.