Information Extraction Of Typical Wetlands In Hunan Based On GF-1 Remote Sensing Image

Wetlands are one of the important ecosystems of Earth’s surface. And it has both characteristics of aquatic ecosystems and terrestrial ecosystems. Remote sensing techniques which can offer timely, up to date and relatively accurate information of wetland becomes the important tool for the research of surface ecological environment.My research which takes East Dongting Lake area as the research district have the purpose to explore the ability and to find the optimal algorithm of extract wetland information for GF-1 remote sensing data. First of all, the best band combination that applies to GF-1 images wetland classification was analyzed by standard deviation, entropy and the Optimum Index Factor (OIF). And then by comparative analysis the segmentation result of conventional software ENVI segmentation algorithm, mean-shift segmentation algorithm and watershed segmentation algorithm it find the best segmentation algorithm of GF-1 remote sensing data. After that, integrated NDVI mean shift algorithm was used to multi-scale segment the sample images which were chosen as the typical wetlands. So the optimal segmentation scale for each wetland was found to carry out Object-oriented classification. Three classification methods include single scale rules classification, hierarchical classification and supervised classification combined with the effective characteristics of image were analyzed by the accuracy of classification. Based on the optimal classification algorithm and the data of GF-1 remote sensing, maps of wetland resources in Hunan can be obtained. Main findings are as follows:(1) Analysis of the best band combinationAfter comprehensive considerate the spectral characteristics and the amount of information include the standard deviation, the information entropy and the Optimum Index Factor quantitative evaluation and considerate the visual effect, the best band combination was determined as 432 combinations.(2) The research of image segmentation algorithmIntegrated NDVI edge information mean shift segmentation is the optimal segmentation algorithm in wetland areas. Watershed segmentation algorithm does not applied to the wetland areas whose remote sensing image will appear a lot of weak edges. In the three watershed segmentation algorithm, the distance transform and control tag transform watershed segmentation algorithm perform owe serious, and watershed transformation gradient segmentation algorithm produces too finely segmented result. The single mean shift algorithm is slightly better than watershed segmentation algorithm and ENVI software algorithms. Considering when adding other auxiliary data can optimize the division of weak edges, NDVI, NDWI and PC1 were selected to compare. By analyze the average number of regions, the evaluation function of experience and local variance, integrated NDVI edge information segmentation was better than integrated NDVI or PC1 edge information segmentation.(3) The analysis of optimal segmentation scaleDifferent wetland types correspond to different optimal segmentation scale. The optimal segmentation scale can be determined by the Global score which synthesized analysis the heterogeneity and homogeneity between objects. And the size of the optimal segmentation scale has a greater degree of its distribution. In this wetland area, the most concentrated distribution sections are water and mudflats, so its optimal segmentation scale is 120. While the distribution of reeds and sedges is more dispersed than water and mudflats, so its optimal segmentation scale is 80.In addition, the distribution of Polygonum hydropiper+Mud and weeds are most crushing, and its optimal segmentation scale is 50.(4) Comparison of wetland information extraction algorithmThe rule classifications with expert knowledge have a better performance than computer classification. By analyzing the spectrum and texture characteristic of each band of GF-1 multispectral data, NDVI and CIWI index, the effective characteristics of image can be chosen to build classification rules. In the three classifications, the highest accuracy method is hierarchical classification, and its overall accuracy up to 89.62%. At the same time, the accuracy of rules classification only get to 84.43%. It is indicate that the segmentation result has a direct impact on the classification accuracy.(5) Wetland information extraction of Hunan provinceBy extracting wetland information of Hunan province, it was found that wetland vegetation majority concentrated in the Dongting Lake area in Hunan, and appeared gradient distribution for environmental feature. In addition, the rate of wetlands in Hunan is not high, and the distribution is relatively concentrated. What’s more, most of the wetlands urgent need to be protected and ecological recovery.