Deep Neural Network InSAR Key Data Processing For High Accuracy Deformation Monitoring

The Interferometric Synthetic Aperture Radar(InSAR)is an important mean to realize wide-area deformation monitoring and terrain inversion.In the process of InSAR data processing,phase filtering and phase unwinding are the key steps that affect the quality of data inversion.An efficient and high-precision filtering and unwinding method is crucial for InSAR to accurately monitor surface deformation.Traditional phase filtering and phase unwinding methods based on physical models are difficult to balance performance and efficiency in areas with low coherence or very dense interference fringes.The deep neural network model has strong feature extraction and nonlinear fitting ability,and is introduced into InSAR phase filtering and phase unwinding method,which can greatly alleviate the contradiction between performance and efficiency.In this paper,the phase simulation method based on mathematical fractal method is constructed,and the phase filtering and phase unwinding model data set is obtained.Then,the Multi-scale Convolutional Neural network(MSCNN)InSAR phase filter model is constructed to solve the problem of resolution and detail information loss.Based on this,the SE module of the channel attention mechanism is added.The MSFFDCNN phase filtering model is constructed to further improve the detail preserving ability and noise suppression ability of InSAR phase filtering.Then,the UnwrapR2 AUnet phase unwrapping model based on improved U-Net is constructed to realize efficient and high-precision InSAR phase unwrapping.The Unwrap-cGAN phase unwrapping model is constructed based on conditional generation of the adversarynetwork model to improve the absolute phase recovery ability and detail retention ability of the low coherence region.Finally,simulation and real data are used to verify the performance of InSAR phase filtering and phase unwrapping models proposed.The main conclusions of this study are as follows:(1)An interference phase simulation method based on mathematical fractal method is proposed in this paper,and a deep learning phase filtering or unwinding model data set is constructed.This method uses the rhomboid square method to simulate the terrain model close to the real terrain features,and synthesizes the data set required for model training and testing based on InSAR imaging geometry.It avoids the errors carried by the existing digital elevation model itself,and is more stable in the process of model training.Moreover,it is more accurate and objective to use simulation data for quantitative evaluation of InSAR phase filtering and phase unwinding.(2)InSAR phase filtering model of MSCNN is constructed by adopting multi-path and multi-level feature fusion strategy.On this basis,SE module is embedded after sine and cosine feature fusion and multi-scale feature fusion,and InSAR phase filtering model of MSFF-DCNN is constructed.This model can dynamically adjust multi-scale feature weight distribution.Make the model strengthen the important features and weaken the secondary features during training;The InSAR phase filtering models proposed by MSCNN and MSFF-DCNN are verified based on simulated and real data.It is found that the proposed MSFF-DCNN phase filtering models are superior to the traditional phase filtering models in terms of performance and efficiency,especially in the phase detail recovery in the low coherence region.Finally,the proposed new model is applied to InSAR deformation monitoring,and it is found that the time series settlement is more complete in space,and the time series settlement is closer to the level data.(3)Based on the U-Net model infrastructure,combined with the residual cyclic convolution module and the attention module,the Unwrap-R2 AUnet phase unwinding model based on the improved U-Net is constructed.The model made full use of the detail information retention ability of U-Net,and improves the convergence speed and anti-noise ability of the model.In addition,based on the main frame of UnwrapR2 AUnet model as generator and Patch GAN as discriminator,the Unwrap-cGAN phase unwrapping model is constructed.Simulation and real data are used to verify the performance of the proposed model.Experimental results show that the proposed Unwrap-cGAN phase unwrapping model obtains the best phase unwrapping results,and the shape variables closest to the level data are obtained in InSAR deformation monitoring.