The Forest Type Identify Research Based On MODIS Unmixing Methods

Mixed pixel is widely existed in remote sensing images, especially low spatial resolution remote sensing image, and ground surface features distributed more complex areas. Using traditional classification methods for low spatial resolution of the image is more difficult to classify, with great uncertainty, therefore, mixed pixel decomposition must be used to classify. Currently linear and nonlinear unmixing model are more mature approach.Unmixing is one of the multi-spectral image processing technology research hotspot and it is difficult to solve, the mixed pixel element contributes to the problem of categories and sub-pixel precision target detection, and then it provides technical support to more fully utilize hyperspectral images.Because of high spectral and temporal resolutions, large coverage, and low cost, MODIS (Moderate Resolution Imaging Spectroradiometer) data has been widely used to quickly extract information of forest types at regional, national and global scales. However, its coarse spatial resolution often leads to mixed pixels and low classification accuracy of forest types. Using spectral unmixing can, to some extent, increase the accuracy of classification. But, how to accurately extract pure endmembers for a study area often is a great challenge. The selection of linear or non-linear spectral unmixing algorithm is another challenge. In this study, a method to extract endmembers from MODIS images was developed. In this method, the time series of MODIS derived vegetation index was first derived and the phenological variation of forest types were analyzed. Decision tree classification was then conducted and the obtained results were used to extract endmembers. With the endmembers, linear spectral unmixing of MODIS images with and without constraints, and nonlinear spectral unmixing were finally carried out and the classification results were compared. The classification accuracy of the land cover types using MODIS images was assessed using the data from forest inventory plots and the area data of land cover classes from forest inventory across Hunan, and the classification results using Landsat Thematic Mapper TM images for Zhuzhou City of Hunan, respectively. The main purpose is to find the best way to quickly extract forest types information based on MODIS data.This research was supported by objective results and achieved the desired goal,it provids an important technical support for the MODIS data to extract forest types information quickly. The main conclusions are as follows:(1) Using MODIS remote sensing data with high accuracy spectral radiation, can achieve full coverage of large-scale regional forest type information extract quickly, thus this study provids a new method useful attempt for monitor large-scale forest resource area, and provide important information for the forestry sector policy-making.(2) the author analyzed the vegetation index NDVI/EVI MODIS time series, find a suitable threshold decision tree classification, and appropriate decision tree classification model for the vegetation index data,the NDVI were2th,8th,9th,15th,16th,18th, the EVI were9th,11th,16th.(3)Using decision tree classification model to purify endmember.Decision tree classification method reduces the dependence on training samples, and avoid the influence of "synonyms spectrum, with spectral foreign body" phenomenon, and the method was higher than the accuracy of supervised classification and unsupervised classification.Using decision tree classification method to purify endmember has a very good results, can improve the accuracy of the final land cover classification, the overall accuracy of the classification decision tree model is85.1%.(4) The classification accuracy of the land cover types using MODIS images was assessed using the data from forest inventory plots and the area data of land cover classes from forest inventory across Hunan, and the classification results using Landsat Thematic MapperTM images for Zhuzhou City of Hunan, respectively.The results showed that the overall accuracies for three kinds of assessment were85.8%,87.4%and85.9%for linear spectral unmixing without constraints,85.1%,88.4%and84.7%for linear spectral unmixing with constraints,64.2%,67.5%and64.7%for nonlinear spectral unmixing, and72.7%,79.7%and73.8%for maximum likelihood classifier.(5)The results showed that: these implied that linear spectral unmixing was the best regardless of with and without constraints, then maximum likelihood classification and non-linear spectral unmixing.Without the constraints of linear decomposition and the maximum linear constrained RMS decomposition were0.2685,0.2637, RMS minimum0,0.0038respectively, indicating a linear spectral mixture more successful. Regardless of the accuracy of testing methods and results accuracy of each classification are very similar, so the results of this study with high reliability. (6) For the MODIS image, the mix of different types of surface features consisting of pixels after decomposition, although a single element like decomposition accuracy is not high, but compared with the traditional method of unsupervised and supervised classification, the classification into the pixel internally, improves the classification accuracy low-resolution remote sensing data, the entire image classification accuracy is higher than conventional classification methods, and the linear unmixing accuracy significantly is better than the non-linear mixed pixel in unmixing, this study wil lbe useful to improve the utilization of MODIS data.