| The rapid development of bamboo industry demands the fast and accurate spatial distribution information of moso bamboo supported by remote sensing technology. As the research front and hotspot of RS science, the hyperspectral remote sensing has the characteristic of image-spectrum unity. It can overcome the shortage of multi-spectral RS; improve bamboo classification accuracy and veracity with its advantage of fine spectrum. Currently, there is few available hyperspectral image, especially in the southern area where the image is influenced by the complex terrain, variable climate and other factors, the obtained image quality often difficultly meet the need of vegetation identification and the coverage is too small. However, the multi-spectral data source is so abundant and its image courage is large. In order to take full advantage of multi-spectrum and hyperspectrum, it is practical significant to do the hypersepctral retrieval research based on the multi-spectrum. Meanwhile, all of the spectral features of vegetation are determined by the chemical and morphological characteristics, making the hyperspectral retrieval based on multi-spectrum possible.Focus on the theme of hyperspectral feature index retrieval of Bamboo thematic information, this paper 1) analyzed the dimensionality reduction for hyperspectral data, 2) identified the bamboo hyperspectral feature, then made an extraction, 3) modeled the hyperspectral feature index with hyperspectral feature optimal index, TM bands, various vegetation indices, and topographic factors etc., and 4) extracted the bamboo thematic information based on the hyperspectral feature optimal index from HyperionEO-1 remote sensing data and TM multi-spectral data in Fujian, included Minqing, Dehua, and Yongtai countries. The conclusions are as follow:(1) Through the analysis of hyperspectral remote sensing images, we can see that the bands available sums to 122, in which Band 31~50, 54~68, 75~89, 94~122 contain more information. These bands are good candidates and Band 122 is the best one. In the spectral features, there is large separability between economic forest and other bodies at any band. The DN value difference among bamboo, Chinese fir, broadleaves and economic forest is obvious at Band 1~25, 31~50, 58~66, 74~8, 92~122, 80~85, showing that these bands are good candidates. Nevertheless, they overlap each other terribly at Band 26~30, 51~57, 67~73, 86~91.(2) Five methods are used for dimension reduction including Band Index Method, Adaptive Band Selection Method, Standard Distance Between Means Method, OIF Method, Principal Component Analysis Method. The resulting classification feature combinations are Band Combination 36, 63, 122, Band Combination 63, 111, 122, Band Combination 6, 97, 122, Band Combination 44, 82, 122 and combination of the first component (Y1) at visible light area, the second component (Y2) at near-infrared area, the second component (Y6) at short-wave infrared area.(3) The classification results based on each feature combination show that: the overall accuracy and moso bamboo's accuracy based on PCA are the highest, with these two value of 80.36% and 88.70%; the overall accuracies based on Band Index Method, Adaptive Band Selection Method, Standard Distance Between Means Method, OIF Method are 64.97%, 65.66%, 71.29%, 66.62% respectively; the accuracies of moso bamboo are 71.75%, 72.32%, 82.49%, 72.32% respectively. Above all, the classified features of Y1, Y4 and Y6 selected with PCA are the best ones. In the paper, these three perform the dependent variables for model contruction. (4) Relying on the spectrum difference analysis on multi-spectral images, vegetation indices, terrain factors and tree spices such as moso bamboo, the paper selected TM (1,2,3,4,5,7, etc. bands), vegetation indexs (NDVI, RVI, PVI), elevation, slope and aspect as the independent variable, Y1, Y4, Y6 as the dependent variable,based on the factor screening, stepwise regression is used to construct the index model which can retrieve the hyperspectral characteristics. The models are : where Y1, Y4, Y6 are available classification features for moso bamboo identification in the hyperspectrum, Z6 is the result of Y6 with Box-Cox Transformation, B1, B2, B3, B4 are Band1 to Band4 of TM images, RVI is ratio vegetation index, NDVI is normalized difference vegetation index, Elevation3 is the elevation from 800 to 1200 m, Slight-slope is the slope from 6 to 15°, Abrupt-slope is the slope from 26 to 35°, Steep-slope is the slope from 36 to 45°.(5) According the index model construction, the hyperspectral characteristics retrieval is experimented. It extracted moso bamboo information with the retrieval results. This classification accuracy is higher than the traditional method based on multi-spectrum data. The overall accuracy based on indices from retrieval is 73.18% with the Kappa of 0.6552, which the accuracy based on original TM data is 62.48% with the Kappa of 0.5170. Compared with them, we can see that the former one is 10.70% and 0.1382 higher than the latter one. Meanwhile, the former moso bamboo accuracy reaches 86.52% and the latter one 74.47%, 12.05% higher. The other species'accuracies also have been raised varying degrees, of which Chinese fir raised 43.44%, masson pine 5.73%, broadleaves 1.51% and economic forest 9.41%. |
PDF Full Text Request