The Study On Adaptive Segmentation Methods For High Spatial Resolution Remotely Sensed Imagery

With the rapid development of remote sensing technology, temporal, spatial and spectralresolutions of Remotely Sensed Imagery (RSI) are higher and higher. The contradictionbetween image data explosion and serious lag in capabilities for dealing with them isincreasingly acute, which challenges traditional image processing techniques. Developing moreefficient theory and technical methods for mining RSI data becomes a tendency to address thechallenge, during which a promising research field, object-oriented RSI analysis was graduallyformed. Image segmentation is one of the key technologies in object-oriented RSI analysis,which is also the hardship of object-oriented image analysis, understanding and application. Sofar, the achieved Remotely Sensed Imagery Segmentation Methods (RSISM) still have variouslimitations and defects. In view of this, the thesis explores adaptive segmentation methods forHigh Spatial Resolution Remotely Sensed Imagery (HSRRSI). Its main research contents andrelated achievements are as follows:(1) Presently available representative achievements in the research field of RSIsegmentation were reviewed and summarized (focusing on various RSISM in mainstreamcategory pedigree), the main current problems (including the poor adaptability of segmentationscale parameter model, the short of utilizing and coupling of object feature information inmulti-scale segmentation, the lack of self-adaptability of algorithms, the inconsistent betweenimage objects by multi-scale segmentation and their corresponding semantic image objects) andpossible researching trend of which were also put forward.(2) The relationship of segmentation effect and fusion in conjunction with filter technologyin HSRRSI preprocessing steps was studied. By taking QuickBird image as sample data in casestudy, respectively verified was the feasibility of integrating advantages of panchromatic andmulti-spectral data of HSRRSI by fusing, and smoothing spectral heterogeneity within semanticimage objects and simultaneously keeping its edge information by filtering. A new idea, whichevaluates and selects image fusion and filter methods based on segmentation effect, has alsobeen explored.(3) By means of output fusion strategy and Canny operator, it proposed new edge detectionalgorithms respectively based on vector, and weighted vector or scalar, which were suitable forHSRRSI. In the multi-colorspace, the algorithms were employed to effectively extract edgeinformation from HSRRSI. Also presented were the differences of spectral responsiveness among various kinds of land cover types and their impacts on edge extraction.(4) Under considering characteristics of segmentation of HSRRSI data in multi-colorspace,the mean shift algorithm was improved, based on which the multi-colorspace,spatial-range-union, and multi-scale mean shift algorithm was implemented. The segmentationresults derived from the algorithm in different colorspace were analyzed and compared witheach other, and which verified the relative advantages of RGB and IHS colorspace.(5) Based on embedded integration strategy, a segmentation algorithm, named AdaptiveIntegration of Canny and Mean Shift Segmentation algorithm (AICMS), of adaptive multi-scale,spatial-range-union, and integrating edge and region was proposed and implemented. Byrespectively taking GeoEye, QuickBird and aerial photography images as sample data, theuniversality was proved. Also proved was the relative advantages of AICMS algorithm inadaptive property superior to the multi-scale segmentation algorithm embedded in eCognition(a commercialized software platform).