Ground-based Differential Interferometry Radar Image Processing Technology Based On Machine Learnin

Ground-Based Differential Interferometric Radar(GB-DIn Rad)has the advantages of all-weather,all-time,flexible observation location and high measurement accuracy.It is now widely used in the deformation monitoring of mountains,mines,buildings,dams and other targets.The selection of permanent scatters(PS)and atmospheric phase compensation are two key technologies in GB-DIn Rad deformation measurement.For PS selection,the conventional amplitude dispersion method and coherence coefficient method have the problem of PS mis-selection in noisy scenes,and some noise points are also selected as PS.Secondly,the number of PS selected by conventional methods is small or the phase quality is not high in the vegetation scene.Although the PS selection method of Sta MPS can select more PS,the phase quality of the selected PS is not high,which is not conducive to high-precision deformation measurement.For atmospheric phase compensation,the conventional compensation methods have the problems of low accuracy of atmospheric phase compensation for nonlinear changes or poor scene adaptability.In view of the problems of the above conventional methods,this paper studies the PS point selection method and the atmospheric phase compensation method in GB-DIn Rad with machine learning method.The main research contents are as follows:(1)Improved PS selection method based on FCMAiming at the problem that the conventional selection method is single and the threshold is set manually,which leads to PS mis-selection in some scenes,a multi-level selection method based on fuzzy C-means clustering(FCM)is proposed.This method makes multi-level selection through the amplitude value,amplitude deviation value and coherence coefficient value of pixel points,and FCM clustering is used to replace the manually set threshold in the coherence coefficient method,then the improved method is verified by a rock slope.The experimental results show that the improved method can better eliminate noise points and select the rock slope as PS,and the phase quality of PS obtained is high,which improves the PS selection accuracy of GB-DIn Rad.(2)Method of selecting PS of vegetation slope based on LSTM networkAiming at the problem that the number of PS selected by GB-DIn Rad under the vegetation slope is small,a PS selection method based on long short-term memory(LSTM)is proposed.This method first removes some obvious non-PS through signal-to-noise ratio and phase standard deviation,then the LSTM network is trained using the characteristic matrix of the candidate PS,and the PS are predicted.Finally,the proposed method is verified by a vegetation slope.The experimental results show that this method can select more PS compared with the amplitude dispersion method;Compared with the coherence coefficient method and Sta MPS method,this method can greatly improve the phase quality of PS;On the premise of ensuring the phase quality,increase the number of PS selected in the vegetation scene.(3)Nonlinear atmospheric phase compensation method based on BP networkAiming at the problem of low precision of nonlinear atmospheric phase compensation by conventional methods,a nonlinear atmospheric phase compensation method based on error back propagation(BP)network is proposed.The method first uses FCM clustering algorithm,slant range-azimuth model and standard deviation-mean threshold method to select stable PS,then the BP network is trained with stable PS,and the atmospheric phase value of PS is predicted and compensated.The verification results of Benxi scene in Liaoning Province with nonlinear atmosphere and deformation show that the BP network method proposed in this paper has higher compensation accuracy than the conventional second-order model method and linear interpolation method.For non-deformable points,the phase value fluctuates around 0rad after compensation.For deformation points,deformation information is well preserved.The compensation accuracy of GB-DIn Rad nonlinear atmospheric phase is improved.