| Hyperspectral remote sensing is a new type of remote sensing technology that has emerged after conventional remote sensing technology.It breaks through the shortcomings of traditional remote sensing,such as low spectral resolution,low number of bands,and difficulty in expressing the detailed spectral characteristics of objects.It has outstanding advantages such as high spectral resolution,large amount of information,and continuous spectral curve.Therefore,hyperspectral remote sensing is widely used in agriculture,forestry,atmosphere,geology and many other fields.Due to the numerous bands of hyperspectral images and large amount of information,it provides advantages for the fine recognition of the ground objects.However it also brings many problems such as large amount of data,large correlation between bands,and reduction of processing accuracy and efficiency.Moreover,in the remote sensing classification the narrower and more channels are used,the classification accuracy will not the better,Therefore,it is necessary to explore the selection scheme of hyperspectral band,reduce the dimension of image,and find the best balance between spectral resolution and object recognition.Based on this study,this study combines multispectral data,changes the band width of spectral data,then merged a number of narrow bands into wide bands,next discuss the classification and recognition of forest types in the study area,the purpose is to find the suitable spectral range of type recognition.The main research results and conclusions of this paper are as follows: 1.It is effective and feasible to simulate the spectral response function of hyperspectral data by Gauss function.By comparing with the original spectral response function of Landsat8 data,the trend is basically consistent,futhermore it can simulate the real spectral imaging process to a large extent.2.The wavelet analysis method can be used to extract useful components from the signal through stretching and translation.As for extracting spectral features,compared with the traditional spectral features processing based on the overall shape of the spectral features,the wavelet decomposition can selectively highlight the target features and suppress the influence of non-related features.The wavelet transform has a good ability of reconstruction.In the process of decomposition,not only the information loss is reduced but also the redundancy is reduced.Therefore,the maximum separability of feature features can be achieved.3.Support vector machine(SVM)has good classification effect on image recognition results generated by the two methods,and the accuracy of each image is higher than 80%.When using Gauss function to simulate spectral response function,the classification precision of 10 nm image is 85.68%,the classification precision of 20 nm image is 85.19%,the classification precision of 30 nm image is 84.92%,the classification precision of 40 nm image is 84.79%,the classification precision of landsat8 data simulated by hyperion data is 82.60%,and the classification precision of landsat8 data is 79.51%.As for using wavelet analysis was to generate two kinds of scale images,the classification accuracy of scale 2 is 84.79%,and the classification accuracy of scale 3 is 83.94%.For the first method,with the increaseing of band width,the accuracy of image classification is reduced.The classification results of 10 nm and 20 nm images have little difference.In the same wavelength range,the classification accuracy of the simulated Landsat8 data is higher than that of Landsat8 data,so the forest tree recognition ability of EO-1Hyperion data is better than the OLI data of landsat8.When SVM is used to classify and recognize forest type in all cases,the wavelet analysis is slightly higher than the method of simulating the spectral response function by Gauss function,which shows that the wavelet analysis can improve the classification accuracy to a certain extent. |
PDF Full Text Request