Research On Hierarchical Segmentation Method Of High-resolution Remote Sensing Image Based On Minimum Spanning Tree Model

High resolution remote sensing data is gradually becoming mainstream along with the improvement of remote sensor's resolution and the remote sensing satellites which have high resolution are constantly launched.With the finer spatial identifiability,high-resolution remote sensing images present more abundant ground object details,which manifest the information contents in images at complexity,spatiality,and mass,and are bringing new challenges to traditional remote sensing segmentation methods.In the view of these challenges,finding a more effective segmentation model and parallel processing method are crucial to improve the segmentation accuracy and processing efficiency of large-scale high-resolution remote sensing images.To this end,the study proposes a hierarchical segmentation method of high-resolution remote sensing image based on Minimum Spanning Tree(MST)model and its parallelization.The former makes use of the hierarchical minimum spanning tree model to effectively depict the complex scene information contained in high-resolution remote sensing images,and a hierarchical segmentation model is built by combining with the regional fuzzy clustering segmentation method.The parallel partioning method is realized based on the sub-block tessellation and parallel regionazational fuzzy clustering method for global optimal segmentation.The major contents of this thesis are as follows.(1)The hierarchical MST model of high-resolution remote sensing image depicts its scene information at local,regional and global scales by employing the Image MST Field(IMSTF)model,the Minimum Heterogeneity Rule(MHR)and the regional label field model,respectively.The IMSTF is an image representation model which considers the spatial and spectrum information of pixels at the same time,and has the special effect of the spatial agglomeration which depicts the borders of the objects adaptively.The partitioning of the homogeneous regions is realized by the minmum heterogeneity region partion method,which takes advantage of the merging rule combined the spectrum information and shape information of regions,and has a good effect of restraining the geomatrical noice caused by the trivial objects in the image.The regional MST model depicts the relationship between the heterogeneity sub-regions,which further decreases the redundancy of the neighborhood relationship of the sub-regions in traditional regional segmentation methods,and significantly improves the efficient of the regional segmentation method.(2)Based on the image hierarchical MST model,the image hierarchical fuzzy clustering segmentation model is built by combing the regional hidden Markov Random Field-Fuzzy C-means(RHMRF-FCM)with the object function solved by the partial differential method.In order to verify the effectiveness and viability of the proposed method,the experiments are conducted with World View-3 high-resolution remote sensing images with the emphases on the effects of the partition scale,the similarity weight of spectrum,and the regional compactness and smoothness weight on the final segmentation result,and the comparing experiments by the proposed method,multi-resolution method and the watershed method from e Cognition software with the same data are analyzed.Quantitatively and qualitatively evaluations of the results from the comparing experiments show that the proposed method not only can overcome the widely existed geometric noise in complex scene of high-resolution images,but also have the higher segmentation accuracy than the comparing methods.(3)In order to increase the efficiency of the homogeneous region partition of the large-scale high-resolution remote sensing image,the parallel MHR partition method and parallel sub-block sewing method are proposed based on the serial method analysis.Then,for further improving the efficiency of the segmentation of the serial RHMRF-FCM algorithm,the parallel RHMRF-FCM algorithm with lower data transfer burden is proposed by the master-slave parallel mode.The experiments from the cost and performance of parallel segmentation aspects verify the effectiveness and viability of the proposed parallel segmentation method.The experimental results show that the segmentation accuracy of the proposed parallel is close to the serial methods.The segmentation time is 1 hour for a high-resolution remote sensing image with 67 million pixels,the overall speedup is 2075,and the parallel efficient curve show that the parallel method has high performance on extendibility and equilibrium for parallel scheduling.There are 67 figures,8 tables and 120 references.