| With the rapid development of earth observation technology,imaging sensors with different high temporal and spatial resolutions have been produced,and the amount of remote sensing image data has increased exponentially,bringing about great challenges for efficient storage and processing technology of remote sensing data.The technology of distributed clusters provides technical support and computational power for the realtime processing and analysis of massive large-scale remote sensing data.It makes the synthesis of image processing algorithms and distributed clusters for large-scale image processing a trend to expand the computing power of the existing image processing algorithms to meet the diverse image processing requirements.However,a vast of object-oriented image segmentation algorithms are still facing many issues when they are directly transplanted to the distributed environment,mainly due to the problems of the border incomplete segment objects,the data redundancy and computational redundancy,and the data skew of Hash partition method.Focusing on the above issues,this thesis mainly focuses on the following four aspects based on the Apache Spark platform:(1)Research on distributed image segmentation methods using repeated calculations.Aiming at the issues of the incomplete segment objects generated by the object-oriented image segmentation algorithms at the image tile boundary,this thesis proposed an information carrier for intermediate calculation results based on three auxiliary bands(boundary,category,direction),which overcame the problems of broadcast communication description between adjacent image tiles in various computing nodes under distributed environment.Then this thesis constructed a repair method for the border incomplete segment objects according to the theory of masked objects extraction,pixels restoration and segmentation,and attribute information replacement,which achieved the repairing of border incomplete segment objects of image tiles.Two types of image segmentation algorithms were employed to evaluate the proposed method in multiple study areas.The experimental results showed that the image processing results of proposed method were highly similar with the ground truth image.(2)Research on distributed image segmentation methods using auxiliary information integration.Aiming at the data redundancy issues of the repeated calculation method proposed in(1),this thesis proposed an automatic buffer detection mechanism based on the border segment objects,which solved the problem of setting buffer size relying on artificial experience.Then this thesis built an auxiliary information integration method aiming at simplifying the number and range of auxiliary bands,which achieved the reduction of data redundancy generated by the repeated calculation image segmentation method.Two seed point image segmentation algorithms were employed to evaluate the proposed method in multiple study areas.The results demonstrated that the distributed method using auxiliary information integration could improve the task implementation efficiently by reducing the data redundancy generated by that using repeated calculation.(3)Research on distributed image segmentation methods using neighborhood region prior.Aiming at the data redundancy and computational redundancy problems of the distributed image segmentation methods in(1)and(2),this thesis proposed a one-way accurate transmission scheme of border objects according to the indexed oddeven parity division of image tiles,which overcame the problem of data redundancy during data shuffle stage.Then this thesis established a repairing method of incomplete segment objects based on the theory of border even tile object sharing and its re-growth in adjacent buffered odd tile,which solved the problem of computational redundancy during the implementation of the image segmentation method using auxiliary information integration.The superpixel algorithm was employed to evaluate the proposed method in multiple large-scale study areas.The results showed that the distributed image segmentation method using neighborhood region prior achieved higher efficiency than that using auxiliary information integration.(4)Research on Apache Spark partition methods using clustering principle.Aiming at the data skew problem of Hash partition method used in experiment in(1),(2)and(3),this thesis transformed the partition problem into a uniform and compact clustering problem by regarding the image tiles as the image pixels without spectral and texture information.The equal area conversion of image layout was used to plan the seed points of partition and aggregate the image tiles,which achieved the initial partition of image tiles.Then this thesis constructed the vertical and horizontal direction adjustment scheme of image tiles in adjacent partitions,which achieved the even data distribution of image tiles among partitions.The evaluation results showed that the proposed partition method solved the data skew problem of Hash partition method,and achieved an approximate monotone linear relationship between the elapsed time and parallelism.Faced with the accuracy and efficiency issues when transforming object-oriented image segmentation algorithms to the distributed platforms.The experimental results showed that the proposed distributed image segmentation methods can solve the problem of the border incomplete segment objects,and the distributed image segmentation method using neighborhood region prior also solved the problems of the data redundancy and computational redundancy existed in that using repeated calculations and auxiliary information integration.The proposed partition method solved the data skew problem of Hash partition method.It is promising to provide theoretical basis for implementing the object-oriented image processing algorithms efficiently in the Apache Spark distributed environment. |
