Research On Object Oriented High Resolution Image Information Extraction Based On Edge Information Enhancement

Geographic Information acquisition is becoming increasingly important with the Information Technology development. As an important statistics in Geographic Information, the Census Geographic Conditions should obtain the basic the natural geographical factor. Remote Sensing, the key technique to acquire Geographic Information, has a wide range of applications in Mapping, City Planning, Territorial Resources, Geology and Mineral Exploration, Agro-forestry and Military. With the continuous updates of Remote Sensing,the highest precision of satellites with high resolution applied such as:WorldView, QuickBird, IKONOS, GeoEye and home produced satellite series could reach decimeter degree, which unfolded more abundant geographical information accordingly.Traditional Remote Sensing Image could be classified into supervising, non-supervising, decision tree and so on based on the spectral signature of the surface features which fit more on Medium and Low Resolution. For the High Resolution, the pixel based classification could not utilize fully its ample spatial structure but cause much salt and pepper noise because of little Spectral heterogeneity and lower the precision of classification result. Hence, it is increasingly hard to meet the needs of the information acquisition for the traditional RSI classification with the High Resolution. For the sake of acquiring high resolution image information, people put forward the Object oriented image information extraction technology by the use of HRI texture, space.The object oriented methods for extracting information including:Object Oriented Remote Sensing Image Classification Based on Gram-Schmidt and other fusion methods; Evaluation index of object oriented segmentation based on overall accuracy or Kappa coefficient.Based on the eCognition software, this paper takes the WorldView-III image of Huainan area as an example to make pretreatment, multi-scale segmentation, enhancing the edge information and enhance the edge information of image information extraction and for different classification methods made a precision evaluation. The specific researches are as follows:(1)Taking the high resolution remote sensing image of Huainan area as an example, the original image was fused with Pansharping and supervised classification and unsupervised classification based on the traditional method, and the classification map of remote sensing image was generated under different methods.(2)The object oriented classification method based on eCognition software is mainly studied. Firstly, with the aid of ESP (estimation of scale parameter) evaluation tools, through different experimental and visual discrimination effects determine the optimal segmentation parameters:the scale factor is 110, shape factor is 0.5, the compactness factor is 0.5. After the completion of the image segmentation, the nearest neighbor classification, repeatedly selecting optimal classification samples and sample characteristics (NDVI, Shap-index, Length/Width, mean-Brightness, mean-max.diff), the final generation based on object oriented image classification chart could be achieved.(3)On the high resolution image were edge detection based on Canny algorithm, after the detection of image layers in objects multi-scale segmentation, and set its weight segmentation, edge detection based on multi-scale segmentation image. Based on Getis-Ord local spatial autocorrelation statistical analysis, and on this basis then the edge detection based on Canny algorithm, after the detection of image layers in objects multi-scale segmentation, and set its weight division, based on information enhancement transform multi-scale image segmentation of high resolution remote sensing image is obtained.(4)The classification accuracy of different image information extraction methods are compared and analyzed, and the maximum likelihood classifier in the pixel based classification method is found in this experiment, which overall classification accuracy for 84.9382%, kappa coefficient is 0.8152, object oriented overall classification accuracy reached 92.3076%, kappa coefficient reached 0.9018, enhance the object oriented classification accuracy based on Getis-Ord and Canny edge information enhanced in the experiment is of the highest, the overall classification accuracy for 94.6667%, kappa coefficient for 0.9354.