Study On High Spatial Resolution Remote Sensing Image Classification Based On Region Sparse Representation

In recent years, the high resolution remote sensing technology is developing rapidly. High resolution remote sensing image had become the main source of geospatial information now, and it also hasbeen widely used in surveying and mapping, urban design, disaster monitoring, traffic construction, resources and environment, fine agriculture, national defense, social and public service, etc.High resolution remote sensing can clearly show the shape, spatial distribution and structural characteristics of objects on earth, but it alsobringsnew problems and challenges to the intelligent processing for remote sensing image.Feature information of objects on earth are presentingmuch higher details with the improvement of spatial resolution of remote sensing image, and rich high-frequency details intensifiyspectral changeswithin same classes and reduce spectral differences among different classes, "same object with different spectrums" and "same spectrum with different objects" had largely happened.High resolution remote sensing imagehas not only the spectral information,it also hasrich texture, structure and shape characteristics,and it has always been the research focus of scholars both at home and abroad that using those characteristicsimproved the mode separability of spectral feature space.But the related studies show that the increasing of feature vector dimension cantrigger a so-called "dimension disaster" and reduce classification accuracy of the surface features in the absence of abundant training samples.In order to overcome these problems and use of these characteristics better, this paper puts forward a kind of remote sensing classification algorithmsuitable for high-dimensional feature,combining with the object-oriented thought and sparse representation theory.Based on the solid theoretical foundation of sparse representation and object-oriented method,sparse representation and object-oriented methodhas showed the superior performancein some areas, such as signal processing, classificationof image, target recognition and so on.Therefore, this paper bringsparse representation theory to the use inhigh resolution remote sensing image classification,conducting deeper research and discussion.Combing the object oriented method, this paper puts forward the new methods and new ideas.The main contents were as follows:1. Combining with the characteristics of high resolution remote sensing image,this paper analyses the problems faced with the high resolution remote sensing image processing, statesthe advantages of the sparse representation and object-oriented technology,andintroduces basic principle and method of the sparse representation and the object-orientedsystematically.2. This paper reviews the application of sparse representationtheory in image classification,giving the sparse representation used in classification of hyperspectral remote sensing image, and this paper extracts the characteristics of high resolution remote sensing image(texture, shape, structure and spatial relation), combing obvious geometrical structureand texture information of high resolution remote sensing image features, making up for the deficiency of the spectral characteristics,andmakesthe classification processingon the basis of the sparse representationtheoryfinally.3. This paper also puts forward the sparse representation algorithm based on region.It risesthe processing primitives of the sparse representation theory from image element to regional object hierarchies,and image analysisin higher level could obtain more abundant information and contribute to image processing.This articlemakesthe sparse representation algorithm based on region to theuse of the image classification;it is the region segmentation of high-resolution remote sensing image first, then using image element characteristics in the region to consist a set of featuresbased on the characteristicdictionary,so a new representative regional featureisconcluded.Finally,we can obtain the regional category according to the theory of sparse representation classification.In this classification experience, we choice 3 high-resolution remote sense images, which are QuickBirdsatellite image, HYDICE aerial imageand IKONOS satellite image. In order to contrast the sparse representation algorithm based on region,wediscuss7 additional classification methods, such as Minimum Distance to Means, Maximum Likelihood Classification, Back Propagation, Support Vector Machine, Region-based Back Propagation, Region-based Support Vector Machine and Sparse Representation Classification. Theresults prove that Region-based Sparse Representation algorithmhas higher classification accuracy, and can effectively control the "dimension disaster".4. This paper puts forward the sparse representation algorithm based on multi-scale area.Due to thecomplexityof high resolution remote sensing image feature shape and spatial distribution and its characteristic of strongmulti-scale, so this article make regional sparse representation of a single scaleexpanded to multi-scaledfurtherly, putting forward a new kind of multi-scale area sparse representation classification schemes: we use single scale area sparse representation method to get representative features in each scale and region, and according tothe Local Moranindex and regional variance in each region,we can get a regional heterogeneity index representing the weight of each dimension,combining the theory of weighted sparse representation tomake classification finally.Experiments show that sparse representation algorithm based on multi-scale area have has higher classification accuracy then single-scaling classification algorithm.