Change Detection Of Hyperspectral Remote Sensing Images Based On Spectral Similarity Measure

With the emergence of global environmental change research,multi-temporal remote sensing images processing and changing detection become an active research direction in remote sensing field,and the change detection with hyperspectral remote sensing images is one of its important branches.Hyperspectral remotely sensed images which are characterized by continuous spectrum,and combination of both spatial and spectral information,record more detailed land cover category information,and have been widely applied to land use,environment monitoring,military reconnaissance,etc.One of the hot topics in remote sensing field is how to take advantage of these resources to automatically detect and rapidly extract land surface change information.Aiming at the key issues,based on current research achievements,this thesis explores the new way of how to automatically extract change information from multi-temporal hyperspectral remote sensing images by improving the algorithm of 2D Otsu and the change detection algorithm based on multiple spectral similarity measure.The major contents and results of this thesis are as follows:(1)The general principles,development background of change detection are systematically stated.Existing problems of hyper-spectral remote sensing images change detection approaches are induced and the ideas and methods for these problems are proposed.(2)EO-1 Hyperion remote sensing image data is the main experimental data of this study.Firstly we need remove the zeroed bands and strongest water vapor bands;secondly we need recalibrate the data,because the data were scaled by 40 for the VNIR and 80 for the SWIR;thirdly we need fix the bad pixels,remove the dark vertical strips,reduce the smile effect,and atmospheric correction;lastly geometric correction is used for further change detection analysis.(3)A 2D Otsu threshold segmentation algorithm based on decomposition has been proposed.To address the problem caused by huge amount of calculation and that only considering grey information,the thesis is decomposed into two 1D Otsu threshold methods by utilizing the thought of decomposing.Then based on analyzing 1D Otsu algorithm,it presents a new threshold recognition function through integrating inter-class variance with intra-class variance.Experiments show the effectiveness for the method.(4)A spectral similarity model for extracting hyperspectral images change information is proposed.The experimental result proves that the model further improves the precision of hyperspectral remote sensing images change detection by fusing the differences of the spectral curve characteristics,spectral intensity characteristics,and spectral information entropy characteristics of two temporal hyperspectral images at an identical geographical area.