Research On Object-oriented Classification Technology Based On GF-2 Image

With the improvement of the resolution of remote sensing images,information in remote sensing images is continuously enriched.The traditional pixel-based classification only uses spectral information to extract information,and it is difficult to meet the requirement for extracting high-resolution image information.At present,the method of object-oriented classification is one of the main methods for extracting information from high-resolution remote sensing image.In object-oriented c:lassification,the selection of features and the dimensions of image segmentation are generally determined manually based on empirical knowledge,and it is subjective and blindness.Focusing on the process of object-oriented classification,this paper uses GF-2 image to study pretreatment of image,image segmentation,selection of features,classification and accuracy evaluation.The tasks have done,including:(1)Based on GF-2 remote sensing image,the pretreatment of remote sensing images is systematically analyzed.Radiometric calibration,atmospheric correction and orthorectification were mainly performed on the images.In this paper,HSV fusion,Brovey fusion,Gram-Schmidt fusion,principal component transformation fusion and NNDiffuse fusion were used to test the GF-2 remote sensing image.Through analysis,NNDiffuse fusion is more suitable for the GF-2 remote sensing image,it not only maintains the band information of the image,but also improves the spatial resolution of the image.(2)Research of multi-scale image segmentation is in progress based on eCognition.The results of multiple image segmentation methods were compared and analyzed to determine the effectiveness of the segmentation based on multi-scale.Because of the lack of scientific basis for the choice of multi-scale segmentation optimal scale,it is difficult to grasp.In this paper,the optimal segmentation scale of all kinds of features is determined by RMAS(Ratio of Mean Difference to Neighbors to Standard Deviation).It shows that the segmentation algorithm based on optimal scale can effectively segment high resolution remote sensing images.And it has certain application value for image segmentation.(3)The combination model of ReliefF algorithm based on MATLAB and J-M distance is proposed to perform feature selection of object-oriented classification.The ReliefF algorithm is used to remove the unrelated features in the initial feature space and obtain the relevant feature sets.Then the J-M distance is used to calculate the degree of separation between the object categories and remove the redundancy in the relevant feature sets.Obtain feature sets that correspond to all types of features.The combined model can quickly obtain the effective feature set of the extracted features,improve the efficiency of object-oriented classification,and have certain research significance for object-oriented classification.(4)Based on the optimal segmentation scale determined by RMAS model and feature set based on ReliefF algorithm and J-M distance optimization,object-oriented classification of the study area image is performed.Compared with pixel-based classification accuracy,the overall accuracy of object-oriented classification results is 88.72%,and the overall accuracy based on pixel classification results is 78.32%-The results show that:for high resolution remote sensing image information extraction,object-oriented classification method has more advantages...