Research On Feature Modeling And Change Detection For Remote Sensing Image

With the continuous development of multi-sensor satellite remote sensing（RS）technology,it is convenient to acquire massive remote sensing image data in real time,and the acquired RS image shows high temporal resolution,high spatial resolution and hyperspectral characteristics.Due to the continuous accumulation and high quality of RS data,the research and technology development of RS images are developing rapidly.Remote sensing images change detection（RSCD）is of great significance and application value in urban planning,resource exploration,disaster warning and military activities.In this thesis,feature modeling and deep learning RSCD methods for synthetic aperture radar（SAR）and optical RS images are studied.The main research contents are as follows:（1）The feature of water in SAR are studied,and a water feature model based on gray histogram neighborhood minimum is constructed,which lays a foundation for subsequent scene change detection based on SAR images（2）A feature modeling CD method based on the combination of contour and morphology was proposed.In this method,the geographical concept contour line is introduced into the image processing,and the Water in SAR is extracted.In order to reduce the influence of the inherent speckle noise in the SAR,the complete continuous water is extracted by the method of morphology and regional growth,and the CD is completed.In this thesis,tests were carried out on San Francisco data set,farmland data set,Omodeo Lake and Tirso River data set and Qori Kalis data set respectively.Meanwhile,the CD results were compared with the other four methods.The experimental results show that the overall effect of this method is optimal,which also indicates that the algorithm can efficiently extract the changing water with low consumption of computing power resources.It has some generalization ability.（3）A RSCD method of semantic feature modeling based on Transformer is proposed.In this method,the improved Res Net18 is used to obtain the feature map,and the differentiated semantic tokens of the feature map are extracted for Transformer which cannot process the image directly,then the output of Transformer is classified by simple network,and the RSCD of urban buildings is realized.In this thesis,comparative tests were carried out on LEVIR-CD and DSIFN-CD data sets with five algorithms respectively.Objective indicators show that the method has the best number of indicators,and some indicators are significantly improved.At the same time,the algorithm can obtain more accurate detection effect in less training time.