Research On Multi-scale Segmentation Method Of High Spatial Resolution Remote Sensing Image Based On Sparse Representation

Remote Sensing Image Segmentation(RSIS)is the process of dividing the image into segmented objects with different characteristics and non-overlapping.The better segmentation result is help to improve the accuracy of image feature information extraction and target recognition.As the acquisition technology of remote sensing image data becomes more and more developed,the spatial resolution of remote sensing images continues to increase.In addition to spectral information,High Spatial Resolution Remote Sensing Images(HSRRSI)has rich texture,structure and shape features.The current High Spatial Resolution Remote Sensing Image Segmentation(HSRRSIS)has two problems: on the one hand,due to the complex and diverse features,it is difficult to accurately segment all features in the image with a single-scale segmentation method;on the other hand,the traditional RSIS method cannot use all the features to segment the HSRRSI,make full use of these features is conducive to improving the segmentation accuracy of remote sensing images.However,with the increasing number of the added feature dimensions of HSRRSI,it will cause the problem of "dimensionality disaster".In order to solve the above problems,this paper proposes a multi-scale segmentation(MSS)method of HSRRSI based on sparse representation(SR)for feature selection.The main research contents are as follows:(1)Selection of MSS parameters.In this study,the multi-scale pre-segmentation of HSRRSI is first performed.This method first needs to set the scale,the weight of shape factor and compactness factor.The RMAS index method is used to select the scale of various features in MSS;the maximum area method is used to quantitatively select the best shape factor and compactness factor of various features.Finally,multiscale segmentation is performed on the eCognition 9.0 platform based on the optimal MSS parameters.(2)HSRRSI feature selection based on SR.HSRRSI has abundant spectral information,texture,structure and shape features.Making full use of these features for segmentation will cause "dimensional disaster" while improving segmentation accuracy.SR judges the feature importance of the training samples based on the sparse coefficient value,and performs SR feature selection on the training samples to reduce the dimension,which is an effective supervised feature selection algorithm.Therefore,this paper uses the L1 regularized SR method based on the Scikit-learn library to study and analyze the HSRRSI feature selection,which can achieve the purpose of removing redundant and unrelated features.(3)Feature space clustering segmentation method based on mean shift algorithm.The problem of remote sensing image segmentation can be regarded as a clustering problem of remote sensing image features,so feature space clustering is an effective method of HSRRSIS.In this paper,on the basis of HSRRSI feature selection based on SR,a feature space clustering segmentation method based on mean shift algorithm is studied.Then,the connected objects are spatially separated based on mathematical morphology,and then the clustering segmentation results are further refined.This paper takes the two research areas R1 and R2 with different types of features in Zhanggong District as examples.The R1 area with simple features is mainly distributed by factory buildings and residential land followed by transportation land and other features.The R2 area with more complex features includes six types of features: arable,woodland,residence,transportation,water and others.This paper takes Geoeye-1 HSRRSI as the data source,and studies a method of MSS of HSRRSI based on SR.The mean pixel accuracy(MPA)and mean intersection over union(MIOU)are used as the accuracy measurement indicators to comprehensively compare and analyze the accuracy of the segmentation algorithm proposed in this paper under different sparsity and different scales.For the R1 area,when the sparsity of other land types is 1000,MPA reaches 89.99% and MIOU reaches 78.66%.When the sparsity of the two types of residence and transportation is 10,000,MPA reaches 91.42% and MIOU reaches 76.87%.At this time,the effect of cluster segmentation of various types of land is better.For the R2 area,when the sparsity of the two types of residence and anable is 1.0,MPA reaches 79.44% and MIOU reaches 69.88%.When the sparsity of others is 10,MPA reaches 88.18% and MIOU reaches 81.88%.When the sparsity of transportation land is 1.0,MPA reaches 82.59% and MIOU reaches 77.97%.When the sparsity of woodland land is 100,MPA reaches 91.26% and MIOU reaches 79.41%.When the sparsity of water land is 100,MPA reaches 91.57% and MIOU reaches 80.95%.At this time,the effect of cluster segmentation of various types of land is better.