Identification Of The InSAR Persistent Scatterers Based On Deep Learning

The interferometric synthetic aperture radar(InSAR)is an all-weather technique to monitor surface deformation(subsidence and uplift)in a high-resolution and precision.The persistent scatterers InSAR(PS-InSAR)algorithm is one of the representative InSAR techniques,assessing surface deformation by analyzing the persistent scatterers(PS)in the time-series SAR dataset.However,the algorithm’s success depends on widely distributed and high-density PS points.Otherwise,selecting an adequate number of PS is challenging due to missing and misclassified PS points,time-consuming,and inefficiency.Then,the application of the PS-InSAR technique is adversely affected.On the other hand,deep learning has evolved rapidly in recent years.The learning has a robust,expressive ability and high efficiency.The deep neural network can efficiently extract the semantic features of different levels of data and learn complex mapping relationships and has been widely used in various fields.In this study,the PS-InSAR technology coupled with deep learning has been carried out.The deep learning algorithm was integrated into the PS-InSAR analysis to identify the persistent scatterers.The primary accomplishments include the followings.(1)A unsupervised network model based on the kernel principal component analysis(PCA)was proposed to identify PS in SAR images,called the KPCA-PSNet.Analyses were performed using acquired SAR data.The results show that compared with the Sta MPS algorithm,the KPCA-PSNet increased the PS density,and the identified PS points have high coherence values.Thus,the KPCA-PSNet is effective in delineating the PS versus non-PS points.(2)A network model based on the fusion of the deep convolutional neural network(CNN)with residual block structure and attention mechanism and the recurrent neural network(RNN)was proposed,called the DCR-PSNet.The net extracted SAR image features and temporal interferometric phase information,and it learned the features from training images of different terrains and landscapes.The analysis results show that the PS points identified by the DCR-PSNet have high density and are uniformly(dispersed)distributed in an area of interest.Compared with the Sta MPS algorithm,the DCRPSNet effectively improves the spatial distribution density of the PS in non-urban areas,and the PS misclassification rate is low.(3)The performance of Sta MPS,KPCA-PSNet,and DCR-PSNet in different regions was evaluated.The results show that,in urban areas,the KPCA-PSNet and DCR-PSNet effectively improve the PS density with high reliability.In non-urban areas,the DCR-PSNet identifies more PS points with a low misclassification rate.Thus,the two deep learning-based algorithms efficiently and reliably identify PS points,delineating abundant permanent scatterers and improving the PS density effectively.Nevertheless,the computational efficiency of DCR-PSNet and KPCA-PSNet algorithms is much higher than that of Sta MPS.(4)The applicability of the two deep learning algorithms was analyzed.The KPCA-PSNet performed better in urban areas because it can identify more PS points in built-up areas but has a slightly high misclassification rate in non-urban areas.The DCR-PSNet has a good performance in both urban and non-urban areas.The DCRPSNet increases the PS density effectively,and the misclassification rate is low.However,the DCA-PSNet is more complicated than the KPCA-PSNet structure-wise,and it takes longer to train the former than the latter.