Typical Vegetation Types Classification Using HJ-1A HSI Hyperspectral Remote Sensing Data In The Huangshui River Basin

In this thesis,taking the Huangshui river basin of Qinghai Province as the study area,the HJ-1 A HSI hyperspectral remote sensing images,field measured vegetation canopy spectrum and other auxiliary classification data were used as data sources,and then the identification and classification for the typical vegetation types from hyperspectral remote sensing data were carried out with the help of hyperspectral remote sensing technology,After data preprocessing,some spectral data transformation forms,hyperspectral image reduction and classification feature were compared and selected,five kinds of classification methods including support vector machine(SVM),BP neural network(BPNN),maximum likelihood method(MLC),CART decision tree(CART-DT)and spectral angle mapping(SAM)are used to classify typical vegetation types.Because of the complexity of terrain,trivial types of patches and the diversity of vegetation types in the study area,some auxiliary classification feature data such as terrain data(including DEM,slope and aspect)and spectral characteristic parameters(including green peak amplitude,red valleys amplitude and red edge normalization vegetation index)were determined to help classification.The objectives of this thesis are to compare five kinds different classification methods and assess their accuracies,obtaining their advantages and disadvantages,finally choose the optimal and effective classification method for complex terrain regions from hyperspectral remote sensing data,at the same time,different spectral data transformation forms and dimensional reduction methods for hyperspectral remote sensing data are compared,and the optimal transformation form and dimensional reduction method were selected for this thesis.Finally,the research results provide effective classification method support for terrain regions from hyperspectral remote sensing data.The main research conclusions are as follows:(1)Five kinds of classifiers are used in this thesis,among which the SVM has the best classification accuracy,the optimal overall accuracy is 89.21%,and the Kappa coefficient is 0.88.The BPNN classification accuracy is slightly lower to the SVM classification accuracy,the overall accuracy is 88.7% and the Kappa coefficient is 0.87.The classification accuracy of the CART-DT is slightly lower than the above two classifiers,the optimal overall accuracy and the Kappa coefficient are respectively is 87.69% and 0.85,but the efficiency of the CART-DT is significantly higher than that of the SVM and BPNN classifier.The overall classification effect of the SAM is lower than that of the above three classifiers,and the overall accuracy of the optimal classification results is 85.5% and the Kappa coefficient is 0.83.The SAM classifier is widely used for hyperspectral remote sensing classification,which is suitable for hyperspectral image classification.However,the selection of the endmember spectrum for SAM algorithm is more time-consuming,and the selection of accurate endmember spectrum also has great influence on the precision of the SAM.The MLC is the worst,and the optimal overall accuracy is 77.91% and the Kappa coefficient is 0.74.(2)For the 8 typical vegetation types in the study area,the producer accuracy of the cold-temperate evergreen coniferous forest is 86.05%,and the five classifiers have higher classification accuracy for this vegetation type.The optimal producer accuracy of the temperate deciduous broad-leaved forest is 88.00%,and the classification accuracy of SVM and BPNN is higher than that of the CART-DT and SAM,and MLC for the recognition accuracy of the temperate deciduous broad-leaved forest is lowest,the producer accuracy is only 72.00%.The optimal producer accuracy of shrub vegetation is 87.00%,and the recognition accuracy of CART-DT is lower than that of the other 4 kinds of classifiers,and the other 4 kinds of classifiers can obtain the producer accuracy of about 86.00% for this vegetation type.The optimal producer accuracy of grassland reaches 88.15%,the recognition accuracy of SVM is the highest,while the precision of the SAM is the lowest,and the producer accuracy of the other three classifiers is about 85.00%.For cultivated land vegetation,the highest producer accuracy is 91.67%,and the precision sorting of five kinds for classifiers on cultivated vegetation is as follows: SVM>BPNN> CART-DT >SAM>MLC.The highest producer accuracy of the alpine vegetation is 94.00%,and MLC is the worst among the 5 classifiers.The optimal producer accuracy of the vegetation-free area is 93.85%,and the five classifiers can attain high recognition precision for this type,in which the classification accuracy of the MLC and the SAM is slightly lower than that of other classification methods.The optimal producer accuracy of the water body reaches 93.33%,and the classification result of the SAM is slightly less than the other 4 classifiers.(3)In this thesis the dimensionality result of three types of data with auxiliary information are compared,their classification results indicated that the classification accuracy of MNF reduction results is better than thoes of the other two,and the average overall accuracy is 86%.The average overall accuracy of PCA dimension reduction results is 84%,and the average overall accuracy of ICA reduction results is 81%.Because of adopting the same auxiliary classification data,it can be concluded that the MNF effect is the best among the three hyperspectral image reduction methods,and the ICA effect is the worst.(4)Four kinds of hyperspectral data forms(including R,1/R,log(R),CR)are obtained by spectral data transformation in this thesis,and the typical vegetation types are classified based on the four kinds of data forms.Through the comparison of the accuracy showed that the classification accuracy of SAM from CR data is the highest with the overall accuracy of 85.5% and the Kappa coefficient of 0.83;The overall accuracy of SAM from log(R)data is 84.82% and the Kappa coefficient is 0.83;The overall accuracy of SAM from R data is 82.29% and the Kappa coefficient is 0.79;The overall classification accuracy of SAM using 1/R data is 80.61% and the Kappa coefficient is 0.83.The above results show that the better hyperspectral data form can improve the recognition accuracy of the classifier,and the classification accuracy of four kinds of hyperspectral data for typical vegetation is as follows: CR > log(R)> R >1/R.