Multi-scale Segmentation Of High Resolution Seismic Remote Sensing Image Based On Data Field

China is an earthquake-prone country,earthquake has bring huge damages to safety of people's lives and property.At present,domestic and foreign scholars make use of remote sensing and computer image processing technologies to extract seismic information,among which object-oriented approaches that are based on image segmentation have drawn much attention.Seismic events are usually complex,while the obtained images provide insufficient information for identification of the seismic targets automatically.Damage information extraction based on remote sensing image remains to be done.It has a lot of problems to be solved,mainly in the following issues: The first is the scale complexity of landscapes.The small-scale segmentation results in too broken regions while the large-scale segmentation fails to capture it's structure.The scale problems remains challengeable,e.g.,the optimal segmentation scale.The second is the characteristics,such as the spectral average,shape and other characteristics cannot effectively distinguish seismic objects.It is difficult to extract the buildings,landslides and other damage bodies correctly.Therefore it is not conducive to further damage information extraction.To address these problems,this paper focuses on the following four aspects:Firstly,an approach based on the gravitational field for the multiscale segmentation of high-resolution remote sensing images(HRRSI)is proposed in this paper.In order to obtain the initial segmentation results of the image,the image is transformed into the gravitational field,which are generated by local contextual information.In this field,pixels travels due to the attraction of neighboring pixels.During travelling,local similar pixels get grouped.Therefore,image is initially segmented to semi-homogenous regions with correct boundary localization.Secondly,this paper addresses the optimized seismic image segmentation problem in a dynamic region-merging style.Starting from the over-segmented regions,the region's features are constructed to represent each region.Then,we use the homogeneity that combines the distance and shape measure to conduct the merge criterion.Neighbor regions are dynamically merged following the graph theory and breadth-first or deepth-first strategy.Thirdly,this paper proposed a semi-automatic and interactive object oriented seismic targets extraction approach.We obtain the initial objects by Mean Shift image segmentation algorithm.Then,we mark the targets on the segmented image.we measure the similarity between objects using spectral histogram.The targets are extracted by region merging procedure following the minimum cost criteria,which fuses the marker and similarity measurement.Fourthly,we design and developed the first version of high-resolution remote sensing image analysis software.Based on IDL and C / C ++ language,the remote sensing image analysis software is initially created,including data read and write,image data visualization,image segmentation,image classification and many other functions.To test the accuracy and effectiveness,qualitative and quantitative analysis are adopted.Experiments are conducted using the earthquake images,including collapsed buildings and seismic secondary geological disaster.Comparative experiments with eCognition's MRS show that the proposed approaches produced more reasonable segmentation and reduced much over-segmentation.