Study On Vegetation Information Extraction Method Based On HJ1A Hyperspectral Data And CCD Data

In recent years, with the constantly developing hyperspectral remote sensing technology, hyperspectral remote sensing data can provide a more-refined spectral resolution for the identification of tree species. However, the processing of hyperspectral data is more difficult due to its large data capacity and redundant information and naturally it becomes a key problem to find a suitable reduction method for hyperspectral data. The study area of this thesis is Yunxiao County, located in Zhangzhou, Fujian Province. The purpose of this research is to extract vegetation information from hyperspectral data based on HJlA-HSI and HJ1A-CCD data. In this study, segmented principal component analysis (SPCA) along with Optimum Index Factor method (OIF) was used to propose three improved hyperspectral band selection methods, that is, the improved SPCA and OIF hyperspectral band selection methods based on eigenvalue、accumulating contribution rate and the dually weighted contribution rate. The main contents of this study were listed as follows:(1) This study developed a corresponding flow of data preprocessing based on HJ1A-HIS hyperspectral data and CCD data. HSI data preprocessing includes stripe noise removal, atmospheric correction based on FLAASH model, geometric precision correction, etc, while CCD data preprocessing includes atmospheric correction based on FLAASH model, geometric precision correction, image mosaic and cropping, etc. The NDVI value calculated from the preprocessed CCD image was used to determine the vegetation area in order to achieve the extraction of the vegetation area from HJ1A-HSI data in the study area.(2) The study proposed three improved dimension reduction methods for hyperspectral data based on the combination of SPCA approach and OIF method, including improved SPCA and OIF wave band selections based on the eigenvalue, the accumulating contribution rate and the dually weighted contribution rate. According to the correlation between the bands of original image, the bands are divided into several subspaces. After converting subspace images through PC A respectively in order to make sure that the first m PCA bands include nearly all information of the original image, the paper calculates the contribution rate for each sub-band with the former m PCA bands, then sorted sub-bands according to its contribution rate to the first principal component band, cumulative contribution rate to the first m principal component bands and the weighted contribution rate of the former m principal components band for the original image with the contribution rate of original bands for the former m principal components band in descending order, and filter out the maximum amount of information containing the lowest correlation subspace subset of bands with the descending band set under inter-band correlation coefficient, lastly select out the original bands that can represent the original image according to the OIF value to achieve the effective dimensionality reduction of hyperspectral images(3) Based on the wave band selection results from SPCA, OIF and the three improved SPCA and OIF wave band selection approaches, the study extracted the vegetation information for 5 different forest types from vegetation area of HJ1A-HSI data in the study area through three different classification approaches, including maximum likelihood method (ML), support vector machine (SVM) and neural network (NN). The 5 different forest types were eucalyptus, economic forest, broad leaved forest, Chinese fir and masson pine. The results showed:ML method had the highest classification accuracy, and also had a better extraction result of vegetation information of hyperspectral data than SVM and NN. The improved SPCA and OIF wave band selection based on the accumulating contribution rate approach had the best ML classification result, with an overall accuracy of 72.13% and a Kappa coefficient of 0.6452, and the improved approach based on the dually weighted construction rate ranked the second, whose overall accuracy was 71.25% and Kappa coefficient was 0.6345. The SVM method had the lowest overall classification accuracy, while the improved approach based on the eigenvalue had the highest accuracy (overall accuracy=56.62%, Kappa coefficient=0.4224). On consideration of the wave band selection approaches, the three improved band selection approaches all had a higher overall classification accuracy than the approaches directly based on SPCA and OIF except the low ML classification accuracy of the improved approach based on the eigenvalue. In conclusion:the three improved band selection approaches wholely had a higher overall classification accuracy than the approaches directly based on SPCA and OIF and the optimal method of information extraction was the combination of the improved SPCA and OIF wave band selection based on the accumulating construction rate approach and the maximum likelihood classification method.