Classification And Reconstruction Of Microwave And Optical Remote Sensing Imagery

Acquisition and weather conditions lead to missing data in optical remote sensing imagery,which largely impacts the following image analysis and applications.Common missing data types include cloud covers,cloud shadows,stripping noise and other noise.Optical remote sensing image reconstruction methods utilize relevant image data for the missing regions.The methods faciliate the following remote sensing applications such as image analysis,target detection and supervision,landscape classification and change detection.Currently,optical remote sensing techniques can be divided as auxiliary imagery based algorithms and image inpainting based algorithms.The former refers to multi-temporal and multi-spectral reconstruction,while the latter means estimating missing data with existed data in the image which is to be restored.The proposed method is based on image inpainting.One of the limitations of inpainting based techniques is lack of pre-knowledge of missing data.However,considering that the microwave remote sensing imagery is free of daynight and weather condition,especially its ability of penetrating through clouds,the thesis proposed a new method combining microwave imagery as pre-knowledge of missing data.The combination is done by image clustering classification methods to segment microwave images to generate the image structure information.Under the coastline scene,the experiment in the thesis indicated that the proposed method performs well in maintaining the consistency of image structure,and has a better restoration accuracy comparing to several other widely-used reconstruction algorithms.The main work is as follows.1.Based on the well-known Criminisi inpainting method,the proposed method aimed at reconstructing cloud covers.The proposed method replaced the original intensity image with corresponding microwave image,and recalculated the isophote data term.The replacement and recalculation improved the capability of original method in irregular landscape structure.2.Several common clustering image classification methods were applied to microwave remote sensing image segmentation.Segmentation quality,advantanges and disadvantages were analyzed and compared.Aiming at coastline scene,and utilizing the segmentation result image as pre-knowledge,a classfication-and-reconstrcution restoration method was proposed and achived a well restoration accuracy in reconstructing landscape boundaries.Compared with serveral other widely-used optical imagery restoration methods,the result image achived by the proposed method is more similar to the original pure image data than result images achived by other methods.