Research And Implementation Of InSAR Phase Unwrapping Method Based On Deep Neural Network

Phase unwrapping is the key step of InSAR technology.The performance of phase unwrapping algorithm directly affects the accuracy of digital elevation model.In recent years,the deep neural network technology has developed rapidly,and the neural network algorithm has good robustness and anti noise.This paper mainly studies how to apply the neural network to the phase unwrapping calculation to improve the accuracy and anti noise performance of phase unwrapping.This paper involves two fields: neural network and InSAR phase unwrapping.Therefore,firstly,this paper introduces the technical principle and data processing process of InSAR,analyzes the principle of traditional phase unwrapping algorithm,and then introduces the common calculation layer principle and training process of neural network.On this basis,a neural network structure for end-to-end phase unwrapping is designed;Two kinds of neural networks are designed to calculate the phase ambiguity gradient,and the phase unwrapping calculation is completed combined with L1 norm optimization;Finally,the accelerated calculation method of neural network is studied.The main research work of this paper is as follows:(1)A neural network for end-to-end phase unwrapping is designed.By adding a senet module to the network,the network can pay attention to the more important features and improve the accuracy of network calculation of unwrapping phase.Experimental results show that the unwrapping phase can still be accurately calculated when the signal-to-noise ratio is 0d B,and the calculation time is shorter than other unwrapping algorithms.(2)For the algorithm based on the combination of neural network and L1 norm optimization,FGANet and Res DANet networks are designed respectively to calculate the phase ambiguity gradient.FGANet uses pyramid attention mechanism and global attention upsampling mechanism to improve the accuracy of network calculation,and extracts spatial features of different scales through dilated convolution,which effectively reduces the amount of network parameters.Res DANet combines residual structure,jump connection structure and two attention mechanisms.The residual structure is used for down sampling to continuously extract spatial features and prevent network degradation caused by network deepening.Two attention mechanisms are used to improve the accuracy of network calculation.The features obtained from down sampling are directly sent to the corresponding up sampling module through jump connection structure to strengthen feature fusion.The phase unwrapping result is obtained after L1 norm optimization.Through experimental tests,both FGANet and Res DANet have achieved accurate phase unwrapping effect,and have good generalization and anti noise performance.When the signal-to-noise ratio is 0d B,they can still accurately calculate the unwrapping phase,better balance the calculation time and accuracy,and have a certain practical value.(3)The accelerated calculation methods of the two neural network designed in this paper are studied.The neural network designed in this paper is optimized by using Tensor RT library.Through experimental verification,several times of accelerated calculation effect is achieved without loss of accuracy.