Change Moniroting Technology Based High Resolution Remote Sensing Images For Coastal Zone

With the development of Chinese satellite remote sensing technology and the successive launch of Gao Fen remote sensing satellites,the rapidly growing multi-temporal remote sensing data provide data support for continuous monitoring of coastal zones.On the one hand,long-term serial satellite remote sensing data record the process of land cover change,reflecting its spatial-temporal variation pattern.Existing multi-temporal change detection methods have been difficult to meet the satellite processing requirements;on the other hand,“continuous” coastal zone change type's research is insufficient,and rarely consider the information of time dimension.This paper makes use of the new Gao Fen remote sensing monitoring technology to study the remote sensing image change detection and classification methods that can effectively us e the multi-temporal context information to realize the dynamic monitoring of the coastal zone.In this paper,we mainly study using the Gao Fen 1 Wide Field of View multi-temporal remote sensing image to monitor the changes of the coastal zone.The main means of monitoring include change detection and multi-temporal classification.The research includes the following three aspects:First,for the current time-series trajectory analysis,the change detection method needs to presuppose the changing model and find the parameters of the model,our study the unsupervised multi-temporal change detection method.An unsupervised slow feature analysis network is proposed,which can mine abstract feature representations of dynamically changing input and detect change s using automatic threshold segmentation on learned slow features.This method is insensitive to light,atmospheric,etc.The experiment compares the proposed method with several classic change detection methods.The results show that the proposed slow feature network has better performance and can effectively improve the separability of features.Then,in order to determine the type of change further,multi-temporal images need to be classified.It is difficult to obtain the labels of remote sensing images,the number of label samples is small,the distribution is uneven,and the types of coastal land features are complex.The research uses multi-kernel support vector machines to classify coastal zones land cover.Support vector machines(SVMs)are widely used due to their ability to process small samples and have a good classification effect.However,common single-kernel SVM methods have disadvantages in dealing with complex multi-class problems.Different kernel have different advantages in processing dat a types.Multi-kernel learning methods combine different kernel functions and have better generalization capabilities.The multi-kernel support vector and single-kernel SVM classification methods,and the classification effects on the original features and the extracted slow features were compared experimentally.The results show that the multi-kernel SVM and the extracted slow features can improve the classification performance.Finally,for traditional statistical classification methods such as multi-kernel support vector machines,each temporal remote sensing image is considered as an independent individual,ignoring the problem of time-dimension information,and it is possible to use the temporal classification model of multi-temporal context information to distinguish the feature changes.LSTM has a “memory” function that can use historical information to predict and classify.In this paper,on the extracted slow features,the LSTM model is used to identify the network classification of multi-temporal remote sensing images and analyze the changes of the coastal zone.This method and multi-kernel SVM were compared,and the influence of the temporal number on the recognition effect of LSTM was analyzed.The experimental results show that the SVM has advanta ges in small samples.With the increase of the number of samples,the LSTM recognition effect is better.Increasing the temporal number has a positive effect on the identification of LSTM.