Phase Classification Method Of Ground-based SAR Slope Monitoring Based On Transformer Framework

Ground Based Synthetic Aperture Radar(GBSAR)system has the advantages of high deformation measurement accuracy,large observation area and all-weather non-contact monitoring,which is one of the main technical means for sub-millimeter deformation monitoring of open-pit mines.However,ground-based SAR is inevitably affected by the environment or system changes in the long-term continuous differential interferometric measurement process,resulting in the error phase change contained in the differential interferometric phase sequence is misaligned and identified as deformation,resulting in deformation false alarm problem.In this paper,the problem is transformed into the classification of phase sequences,and the difference of time series characteristics corresponding to different phase change types is studied.A time series feature expression and classification method of ground-based SAR differential interferometry based on Transformer framework is proposed.The main work of this paper is as follows :(1)In this paper,the main components of differential interferometric phase in slope scene are analyzed,including slope shape change phase,echo multipath effect phase,offset phase caused by long-term environmental change,and residual error phase after preliminary correction.According to the temporal variation characteristics of the phase sequence,three types of feature sequence coding methods,namely,difference rate,standard deviation and least square fitting,are proposed,which can calculate and reflect the differences in temporal variability,volatility and change direction of different types of phase sequence changes.(2)Based on the original Transformer framework,this paper constructs the temporal multi-heading self-attention mechanism,the phase category self-attention mechanism,and the temporal-category multi-heading attention mechanism to learn the temporal variation relationship of various phases.The monitoring data set of open-pit mine slope stability is made and the Transformer training and testing are completed,and four kinds of differential interference phase sequence classification are realized.The experimental results of phase classification show that the differential interferometric phase feature coding can enhance the discrimination between different types of phases,but the failure of relative position coding makes the frame insensitive to the correlation between time series,resulting in the classification results of deformation phase and error phase are still inseparable.(3)In order to solve the problem that the relative position coding method in the original Transformer framework is difficult to reflect the temporal correlation,this paper adds Gaussian distribution to improve the relative position coding when calculating the temporal attention.Compared with the previous Transformer framework,it is proved that the improved relative position coding method based on Gaussian distribution can further distinguish the deformation phase and the error phase,and effectively avoid the error phase interference in deformation inversion.