Object-level Sparse Representation Classification Of High Resolution Remote Sensing Images Considering Shadow Compensation

With the rapid development of high resolution satellite remote sensing,it has been widely used in multitudinous fields.The new high resolution remote sensing images have dual advantages in spatial resolution and spectral resolution,which could provide richer and more detailed spatial information and ground features,such as image size,texture,shape and the relationship between adjacent objects.However,the classification of high resolution remote sensing image is confronted with two important problems: one is that the shadow in remote sensing image means poor image quality,which increases the difficulty of extracting and recognizing the object information from images;the other is its easily leading to the so-called "dimension disaster" problem due to the characteristics of information overload.According to the two problems mentioned above,shadow detection,shadow compensation and object feature space clustering are studied for new high resolution images in this paper.The main research contents and innovations are as follows:(1)Because the phenomenon that "different body with the same spectrum" or "the same body with different spectrum" exists in remote sensing image,common shadow detection methods based on image features have problems in universality,robustness and other aspects,which easily lead to missed detection and false detection,while the physical model-based shadow detection methods are limited by data sources.Aiming at above problems,taking into account the increasing enrichment of high resolution remote sensing images and basic geographic information,A multi-angle shadow detection method for high-resolution remote sensing images supported by generalized stereo pairs and assisted by multi-source data is proposed from satellite photogrammetry theory and geometric principle of shadow formation.The experimental results show that:(1)With the support of ground control points,the generalized stereo pair can complete three-dimensional reconstruction,and the positioning accuracy of the reconstruction model can meet the standard requirements of the corresponding mapping scale.(2)This method is feasible and accurate.Although the process is relatively complex due to the more rigorous theory,it can avoid the common problems caused by the phenomenon of same object with different spectrum or different object with same spectrum.With the increasing improvement of high-resolution remote sensing image database and all kinds of basic geographic information databases,this method could be widely applied to the accurate shadow detection of all kinds of high resolution remote sensing images,and will have good engineering application prospects.(2)The existing methods of shadow compensation for high resolution remote sensing image have some limitations.In this paper,a new shadow compensation algorithm based on Contourlet transform and HIS color model is proposed for high resolution remote sensing images.This algorithm contains following steps: Firstly,converting images from RGB space to HIS space,and then,playing nonsubsampled contourlet transform to ‘I' component in HIS space,which is very sensitive to human eyes and contains abundant texture information;Secondly,compensating the high frequency and low frequency coefficients from Contourlet transform,adjusting ‘H' and ‘S' components according to the average value and standard deviation of shadow region and its adjacent non-shadow region;Finally,playing color space inverse transformation to I,H and S components,which are compensated.Experimental results show that,compared with common algorithms,such as homomorphic filtering algorithm,separate HIS color model,wavelet transform and HIS color model,the new algorithm could achieve better shadow compensation effectIt can achieve better compensation effect and is more conducive to recognite and extract land-use information from remote sensing images.(3)Image segmentation is the basis and key of object-oriented classification of high-resolution remote sensing images,which directly determines the accuracy and reliability of classification results.Multi-scale segmentation technology can make full use of the spectral and texture features of image to segment step by step,which is a hot topic in image segmentation research.In this paper,two typical high resolution remote sensing images,WorldView-3 and GeoEye-1,are used as experimental data to carry out segmentation experiments with different parameters.The results of non-linear fitting show that for the mainstream high resolution remote sensing image with 0.5 m resolution,the optimum segmentation parameters of scale,shape and compactness in rural areas are 280,0.1 and 0.45 respectively,and the optimumsegmentation parameters of urban areas are 195,0.2 and 0.55 respectively.(4)Aimed at the so-called "feature dimension disaster" problem,the object-oriented classification method based on sparse representation for new high resolution remote sensing images was proposed in this paper,including the steps as followed.First of all,the images were performed for preprocessing,including fusion,enhancement and segmentation.Secondly,the information of spectral features,texture features,shape features,and other features were extracted from high resolution remote sense images whose superiority was fully utilized;and the information of spectral features,texture features,shape features and other features was extracted.Thirdly,sparse dictionary was constructed by the sufficient number of samples.Last but not least,the sparse coefficients to all types of objects were obtained based on the sparse representation theory.Subsequently,the types of ground objects were determined according to the sparse coefficients.As shown by the experimental results,compared with the object oriented classification algorithm in ENVI,the algorithm proposed in this paper could make full use of the spectral features,texture feature and shape features from high resolution remote sensing images;reduce the difficulties in choosing features and determining characteristic parameters;effectively solve the problem of excessive feature dimensions usually occurred when the feature information was extracted from high resolution remote sensing images;improve the classification accuracy;further expand the application field of the sparse representation theory;and have certain advantages and practical value.Finally,the research work of this paper is summarized comprehensively,and some problems to be studied in the future are prospected.