Study On Feature Extraction And Classification Of Radar Remote Sensing Image Sequence Based On Transformer Framework

As a means to detect and science research by using the information of natural changes in the scenes,Synthetic Aperture Radar is complementary to optical sensors and artificial contact measurement technology.After several times of development of imaging technology,the high-resolution Synthetic Aperture Radar remote sensing data with different phases,systems and bands will greatly enhance its application ability in the field of comprehensive utilization of large scene resources and investigation and monitoring.However,it is time consuming and need great effort to interpret the collected and focused remote sensing images,and it is difficult to achieve the full automation and unified standard of remote sensing image thematic information extraction.Therefore,some scholars at home and abroad began to use neural network to extract effective features,and through the combination of computer artificial neural network and artificial interpretation technology to mine a database reflecting the characteristics of land information variability.In order to simplify the process of further extracting sequence features or learning structured information,understand the representation of the relationship between various types of data at a new level,and design the required model structure according to data,tasks or intermediate results.For the application of deformation monitoring,this paper focuses on the study of Permanent Scatterers feature extraction and classification method in the ground-based radar remote sensing image sequence.The model based on Transformer framework is used to represent and learn the input remote sensing data,which can focus on the important and useful information in the input data in line with the current output environment,so as to reduce the input dimension,and improve the credibility of the predict graph.It is verified by experiments in this paper,the model based on Transformer framework can be effectively used in radar remote sensing image sequence feature extraction and sequence feature research.The main work of this paper is as follows:(1)Based on the analysis of the limitations of traditional sequence modeling neural network framework and the development of Transformer framework components,by fusing the features of radar Permanent Scatterers points and deformation information closely related to the task,the processing flow of classification algorithm based on Transformer framework is given.(2)For the application of deformation monitoring,aiming at the existing multi threshold method of Permanent Scatterers extraction from ground-based radar remote sensing image sequence,the threshold feature is adjusted to the time series feature which can be used by classification network.The amplitude deviation is extended to amplitude first difference to mirror the features of amplitude information;The coherence coefficient is used to dig the interference phase and correlation curve between master and slave images to express the sequence variability;The mean filtering phase is developed from the noise phase filtering to observe the noise filtering;And from the regional characteristics of the real landform,it is found that the distribution density of Permanent Scatterers is used to represent the effective regional range of Permanent Scatterers,so as to integrate the differences and similarities in spatial position.(3)This paper proposes an algorithm for Permanent Scatterers points selection based on Transformer framework model processing sequence radar data.It uses spatial attention and feature stitching fusion instead of traditional multi threshold method for Permanent Scatterers points selection.Each feature is used as an attention head separately.The modeling of position coding is a Gaussian distribution,and a gating mechanism is added,Mask,sum of attention distribution and other components are used in different positions of the model.Then it compares with the Permanent Scatterers points selection and classification method based on recurrent neural network,long-term memory network and convolution network,and then compares with the calculation method of additive attention.Finally,the ablation experiment is carried out to verify and predict the classification of the whole scene.The experimental results show that the real-time performance and accuracy are better.