The Research On Domain Adaption Classification And Change Detection For Remote Sensing Images

With the continuous development of social economy,the environmental problems caused by human activities are becoming more and more serious.Land use type classification and land cover change detection(Land Use / Land Cover,LULC)using the advantages of remote sensing images can provide the basis for the rational utilization and exploitation of land resources,and is of great significance to ecological protection and sustainable economic development.However,current remote sensing image interpretation methods are affected by the phenomenon of "same land covers may have different spectral signatures,different land covers may have similar spectral signatures" and lack effective methods which are specifically applicable to the dynamic monitoring of remote sensing image,hence,restricting the general adaptation of remote sensing technology in large area,multi-temporal land use / land cover issue.Therefore,it is urgent to explore how to quickly and effectively extract land use / information using image data.This study is mainly carried out from two aspects of land use classification and land cover change detection using remote sensing images,and discusses the domain adaptation classification and the object-based change detection of remote sensing.First of all,the research is aiming at the phenomenon of "same land covers may have different spectral signatures,different land covers may have similar spectral signatures" in the same period of remote sensing image classification.Considering the dynamic monitoring process of remote sensing image interpretation,the pattern classification method cannot adaptively adjust the classifier according to the supplementary sample set,cannot meet the practical application needs of remote sensing image land use land cover interpretation.Therefore,the domain adaptive unimodal subclass decomposition model is established to enhance the description ability of the subclass' features in the same-time image,and to realize the continuous learning in different regions,as well as to improve the accuracy of multi-regional remote sensing image classification.Secondly,aiming at the problem of training sample acquisition in the process of dynamic monitoring of multi-temporal remote sensing images,a method of domain adaptation classification based on weighted extreme learning machine is proposed.Through the domain adaptive learning,the useful samples and effective parameters in source domain are transmitted to the target domain classification task,which significantly reduces the demand for the target image training samples,reduces the manpower and material resources of the target classification task,effectively improves the classification accuracy and integrity,and realizes the timeliness of dynamic monitoring using remote sensing images.Finally,an object-based entropy query-by fuzzy ARTMAP joint classification change detection method is proposed to solve the problem of traditional post-classification comparison method.By using the active learning algorithm to select the training samples which are most conducive to training the neural network,and the influence of the error accumulation problem on the classification result is reduced by the joint classifier.In addition,the result of the object-based change detection based on the superpixel segmentation is obtained,which greatly reduces the change detection excessive evaluation caused by "salt and pepper phenomenon".The dynamic monitoring of Panjin wetland reclamation process is realized by using multi-temporal images,and the applicability of the proposed method is verified.