Sparse Convolutional Network-Based Atmospheric Correction Of InSAR Coherent Points

Interferometric Synthetic Aperture Radar(InSAR)technology has been widely used for surface deformation monitoring in recent years due to its high spatial resolution and all-day,allweather measurement advantages.Atmospheric delay error is a major error source in InSAR measurements,and the study of atmospheric delay correction is a hot topic of current research.In recent years,researchers have proposed a deep learning Convolutional Neural Network(CNN)based convective layer correction model ARU-Net,but the convolutional neural network used in this model requires interpolation processing when processing InSAR coherent points,which reduces efficiency and introduces new error sources,and has major limitations in producing training model datasets,which largely limits the application of deep learning based convective layer delay correction methods.In order to overcome the above limitations,an atmospheric delay correction method based on deep learning sparse convolution technique is proposed in this paper and applied to the tropospheric atmospheric delay correction at InSAR coherent points.The main contents and conclusions of this study are as follows:(1)To address the problems of low efficiency and poor accuracy of deep learning atmospheric correction methods,an InSAR coherent point atmospheric delay correction model sARU-Net based on deep learning sparse convolution technique is proposed.The proposed sARU-Net model can learn the atmospheric delay characteristics from the interferometric phase of InSAR coherence points in the Hong Kong region and Shangyu region in Zhejiang Province,and can achieve the effective estimation of tropospheric delay at InSAR coherence points.(2)A deep learning model training strategy is proposed to address the problem of difficulty in obtaining training datasets for atmospheric delay correction models.This strategy is used in producing the dataset by combining SVD or Stacking techniques to obtain higher quality data from the available interferograms as initial values to participate in building the dataset,and iteratively updating and optimizing the dataset to achieve model refinement.The proposed deep learning model training strategy can be effectively applied to model training of coastal ground subsidence with good generalizability.(3)The proposed sparse convolutional network model sARU-Net and the training strategy were tested and evaluated using datasets constructed from TerraSAR-X in Hong Kong and Sentinel-1 SAR data in Shangyu,Zhejiang.The results show that: 1)in the Hong Kong region,the accuracy of the sARU-Net model is improved by 45% and the training efficiency is improved by 23%;the atmospheric correction magnitude using the sARU-Net method is improved by about 77.7% compared with GACOS and linear height correction methods;the atmospheric correction accuracy of the obtained coherent points is improved by 13%,and the deformation results are consistent with the measured level results and the DS-InSAR method.2)In the Shangyu area of Zhejiang Province,the training efficiency is basically comparable;the atmospheric correction magnitude using the sARU-Net method is improved by about 68.3%compared with the other two correction methods;the atmospheric correction accuracy of the obtained coherent points is improved by 35.5%,and the deformation results coincide with the measured level results and are basically consistent with the DS-InSAR method results.