Research On Dimension Reduction Algorithm Of Hyper Spectral Remote Sensing Data Based On Rough Set

A great deal of spectral data provides human with the rich information forunderstanding surface features, which are very useful for subsequent applications. Butbig data and redundant information of the hyper-spectral images leads to theconventional data processing for facing the challenge. Therefore, a method of theeffectively utilizing the maximum information of hyper spectral data and quicklyprocessing hyper spectral data is proposed to study.The dimension reduction of hyper spectral image data is to find a new methodfor processing and analyzing data. Without loss of effective information, the data isreduced in order to obtain the core knowledge of data and realize the dimensionreduction of data. Rough set is a kind of data processing method. Its maincharacteristic is it can do data reducing to get core data, without losing effectiveinformation. Based on rough set, hyper spectral data are proposed for surface featuresspectrometer and imaging spectrometer.The reduction dimension method is proposed to the surface features spectral data.According to ground spectral data, eighteen hyper spectral indices were selected asvariables to estimate chlorophyll content of rice canopy in this paper. However, thecorrelation between these indices causes data redundancy, which affecting theretrieval rate of the leaf chlorophyll content. This paper presented a chlorophyllestimation method based on rough set attribute reduction and support vectorregression (SVR) to solve this problem. Firstly, data space was reduced by usingrough set algorithm and then the SVR algorithm was introduced into estimatingchlorophyll content. There are six indices reserved in the reduced kernel after attributereduction. The R2of retrieval results based on all indices and reduced kernel are0.8586,0.8506respectively.The feature selection method is proposed for imaging spectrometer data. Asupervised band selection reduction dimension method of the hyper spectral imagesbased on rough set is proposed in this paper. In this method, the interest region isselected according to the image-based species type. So all bands of the selected samples are regarded as the condition attributes of the rough set, the species type isregarded as the decision attribute. The reduced kernel (the selected band combination)is obtained by the reduction operation. The airborne hyper spectral HYDICE data inWashington D.C. area are used to verify the effectiveness of the proposed method.The six major species types are selected in order to obtain the reduction results of theeight bands. The support vector machine is used to classify the images of thereduction band combination and full-band images. The classification accuracy basedthe images of the reduction band is94.13%, and the classification accuracy based thefull-band image is94.17%. The classification accuracy difference is less than1%.The experimental results show that the proposed dimension reduction of hyperspectral image data algorithm can effectively reduce the band redundancy anddecrease classification time under keeping unchanged retrieval and classificationaccuracy.