Forest Vegetation Classification Using High Resolution Remote Sensing Image

Classification and extraction of forest vegetation is the basis and key link of forest resources monitoring system based on remote sensing technology. The result of forest vegetation classification is not only related to the accuracy of forest area, spatial location, forest classification and other qualitative monitoring index, but also directly affect the quantitative monitoring of the accuracy of accumulation, biomass, carbon storage, biological diversity and ecological service function of forest vegetation.Our series of high resolution remote sensing satellites and series of resource satellite successfully launched greatly enrich the high resolution remote sensing data. Therefore, there is a need to research a series of key techniques involved in the extraction of forest vegetation classification based on the high resolution remote sensing data, such as remote sensing image pre processing process, image segmentation method, the optimal segmentation scale selection and classification model construction way and so on, and build a more complete classification of forest vegetation extraction system.And the multi-seasonal remote sensing data, DEM data, multi-resolution image data effectively comprehensively applied to the extraction of forest vegetation classification research need to implement.These job for the efficient application of domestic high resolution remote sensing data in forest resource monitoring. These work can provide a reference for the high efficiency application of domestic high grade remote sensing data in forest resources monitoring.This study takes Songshan National Natural Protection Area and its surrounding in Yanqing County total of 99 km2 as the research area, takes remote sensing image of GF-2 as data source, and takes forest vegetation classification as the main line to construct the forest vegetation classification system framework which based on object-oriented algorithm and the high resolution remote sensing and Study on the procedure of data processing and the spectrum separability of GF-2 data. A multi-scale segmentation method is improved based on Shift Mean algorithm, region merging algorithm and terrain partition algorithm and an optimal scale selection method is proposed based on prior knowledge. The selection of multi temporal remote sensing data and image feature index is realized according to analysis the results of image object features of different objects under a variety of form. Four hierarchical extraction rules of forest vegetation is obtained using CART decision tree algorithm based on the segmentation results in four different scales. And the results of classification based on the rule set were valuated using the confusion matrix. The main results are as follows:(1) The fusion results obtained based on GF-2 panchromatic and multi spectral image data and PCA, GS, Pansharp and SFIM 4 kinds of fusion methods are analysis by visual and quantitative. The analysis results show that the fusion image got by Pansharp algorithm is superior the other three methods in terms of clarity and spatial identification, and Pansharp algorithm is the best fusion method when it applied to extracted high precision object boundary from GF-2 image.(2) The study comparative analysis the spectral separability of the multispectral remote sensing image of WorldView-2 and GF-2 through J-M distance. The analysis results show that the remote sensing image of GF-2 has better spectral separability in distinguish feature categories, the forest land and other land types have a high degree of discrimination, and the spectral separability of the multi spectral remote sensing image of GF-2 is similar to the multi spectral remote sensing image of worldview-2 composed by four common bands.(3) A multiscale segmentation method was constructed based on Shift Mean algorithm, region merging algorithm and terrain partitioning method in this study. The segmentation result of the study area was obtained by the multiscale segmentation method combined with the optimal scale selection method based on prior knowledge. The result compared with the multi-scale segmentation results based on ESP software and eCogniton scale evaluation tool show that the result obtained based on the method of this paper is more consistent with the application of forest vegetation classification.(4) Four kinds of classification rule set was constructed by CART decision tree algorithm based on the multiscale segmentation results obtained by the method of FNEA, the method of Mean Shift algorithm combined with region merging algorithm, the method of Mean Shift algorithm combined with region merging algorithm take the terrain division as a prerequisite, and the improved hierarchical extraction method for forest vegetation classification. The classification results were obtained according to the classification rule set. And the results of classification based on the rule set were valuated using the confusion matrix. The results showed that the classification accuracy of the improved hierarchical extraction method for forest vegetation classification was the highest, the overall classification accuracy was 82.24%, and the KAPPA coefficient was 82.67%; the method of Mean Shift algorithm combined with region merging algorithm take the terrain division as a prerequisite is second, the overall classification accuracy was 79.51%, and the KAPPA coefficient was 79.99%; the method of FNEA and the method of Mean Shift algorithm combined with region merging algorithm is slightly worse, the overall classification accuracy was 77.32% and 77.05%, and the KAPPA coefficient was 77.84% and 77.38%.