Object-oriented High-resolution Remote Sensing Image Water System Information Extraction

As a vital geographical factor,the water system plays an important role in accurately obtaining its spatial distribution and frequency of change.With the development of modern remote sensing technology,high-resolution remote sensing images provide data support for us to obtain data and use preface technical means.This paper takes the high-resolution remote sensing image of Word View-2 in some districts of Hangzhou in December 2009 as the data source.The image includes 8multi-spectral bands with a spatial resolution of 1.8m and a full-color band with a spatial resolution of 0.5m.An object-oriented approach is used to study the extraction of water system information.Object-oriented image analysis technology is aimed at image segmentation objects rather than pixels in the traditional sense,which can make full use of the rich spectral,shape,texture,spatial relationship and other feature information of high-resolution remote sensing images.The traditional manual visual interpretation is time-consuming and laborious,and can no longer meet the actual needs of production and monitoring.Therefore,how to realize the automatic,rapid and accurate extraction of water system information is the focus of research in recent years.The main content of the paper has the following parts:(1)In the preprocessing of Word View-2 high-resolution remote sensing image,this paper uses ENVI software as the platform to perform image fusion,orthorectification and image cropping processing on the data in turn,and use them separately.Brovey transform fusion,Gram-schmidt Pan Sharpening(GS),NNDiffuse three fusion methods,using information entropy,mean and standard deviation,average gradient as the quality evaluation indicators for comparative analysis,The NNDiffuse fusion method works best.(2)Using e Cognition software as the platform,this paper uses the method of chessboard segmentation,quad-tree segmentation,multi-scale segmentation and spectral difference segmentation to segment the ground objects in the study area and visually interpret the segmentation effect by visual inspection.A method of multi-scale segmentation combined with spectral difference segmentation is proposed,which can optimize the over-segmentation in multi-scale segmentation.At the same time,by setting different segmentation scale parameters,it is found that the parameters of multi-scale segmentation are set to 90 first,and when the spectral difference segmentation parameters are set to 20-30,the segmentation effect of the objects in the study area is the best.(3)In the selection of the band weight,by extracting the mean value,standard deviation,correlation coefficient and other spectral information values of each band of Word View-2 high-resolution remote sensing image,the band selection best index OIF index is used as the index,and the selection is made.Band 2,Band 6,Band 8band combination is more reasonable for the optimal band combination.(4)Through the statistical analysis of the image spectrum,shape and texture information of the study area and the topic index,the separation threshold algorithm SEa TH is used to automatically select the classification features and calculate the threshold value,and select the best feature combination of the water system information to construct Reasonable classification rules.(5)Using the nearest neighbor classification in the object-oriented supervised classification,the deterministic rule method in the rule classification,and the classification strategy method using the combination of the two,and the pixel-based supervised classification method are compared with the research area.Water system and other features are extracted and classified.Accuracy analysis is performed by using user accuracy and drawing accuracy.Comparing the experimental results,the accuracy of producers using the nearest neighbor classification,deterministic rule classification based on improved SEa TH algorithm and combined classification are91.78%,92.99% and 96.61%,respectively,and the user precisions are 91.05%,93.77% and 96.67%,respectively.The producer accuracy and user precision based on the supervised classification of pixels are 86.96% and 88.76%,and the two precisions are less than the water extraction precision using the object-oriented classification method.This shows that the object-oriented classification method has higher precision than the traditional pixel-based classification method;and the method of combining the nearest neighbor classification and deterministic rules can effectively improve the classification accuracy and classification stability.Sexuality is more suitable for the extraction and classification of water system information in high resolution remote sensing images.