Research On Object-based Classification Of Remote Sensing Imagery Based On MSRC

The Remote sensing image data, with the advantages of largecoverage, information of objective truth, low cost and easy access, hasbeen widely applied in various fields. For the transportation sector, usinghigh-resolution remote sensing technology, combined with the existingmeans of traffic information collection, can provide effective technicalmeans for comprehensive decision service, such as urban trafficmonitoring, road network planning and construction, road traffic runningstate identification, industry management, leadership decision-making etal. However, with the continuous improvement of remote sensing imageresolution, feature information extraction technology is lagging behindthe development of remote sensing technology and the application ofremote sensing technology in the transportation sector is still in itsinfancy.High resolution remote sensing image, relative to low resolutionremote sensing image has more abundant spatial information, textureinformation and geometric information of ground surface. With thedevelopment of remote sensing technology and the increasing resolutionof remote sensing images, the image data information extraction andclassification technology is facing new problems and challenges. Thetraditional classification method based on pixel due to its limitation oflow classification accuracy has been unable to meet the needs of thedevelopment of remote sensing technology, so the object-orientedclassification technique emerge as the times require. When theobject-oriented classification technology in the processing of remotesensing image, the minimum information extraction unit is no longer asingle image element, rather than the spectrum and texture feature similar"homogeneous object ", so we can make full use of other structures’information including spectral features, which greatly improves theclassification accuracy and efficiency.On the object-based classification technology,①We use themarker based watershed algorithm to obtain the image object. Puttingforward a kind of overflow water model, and modify the mark generation method, at the same time by using edge detection methods, improve theability to extract weak edge and restrain the over segmentationphenomenon.②We extract objects’ features in this thesis, includespectral information, texture information and edge geometry information,etc. The different combination of features experiments prove that, thedifferent features have different abilities in object recognition, so it isnecessary to combine useful features after selection for the specific object.③We use adaptive weighted Multiple Sparse RepresentationClassification approach(MSRC) as the classification method in this thesis.Meanwhile, MSRC adaptively adjust the weight of each feature, whichsolves the dimension disaster caused by over multiple features.Quantitative and qualitative evaluations of these experiments can be seenthat the object-based classification method based on MSRC can makebetter use of the synergy of different characteristics and both the overallaccuracy and Kappa has been significantly improved.