Dimensionality Reduction and Compression Method for Time Series InSAR Images Based on Tensor Decomposition

Objectives: With the increase of SAR images, the processing capacity of time-series InSAR data is exponential growth, which brings new challenges to the monitoring of wide Area long term surface deformation monitoring. Especially in distributed scatters InSAR (DS-InSAR) technology, all interferograms are involved in computation, which requires high computing and storage resources, which to some extent limits its development and application promotion. However, time-series InSAR images have redundant information in both temporal and spatial dimensions, dimensionality reduction compression is one of the effective methods for removing redundant information. Methods: Due to the inability of matrices to meet data processing requirements, a tensor with unique advantages in storing and processing high-dimensional data is introduced, and a temporal InSAR image dimensionality reduction compression method based on tensor decomposition is proposed. According to the similar statistical properties between the pixels in the small subspace, the covariance matrix in the small subspace is expressed as a third-order tensor form. The Tucker decomposition algorithm is used to realize the time dimension and space dimension reduction and compression processing at the same time. Results: Taking Sentinel-1A image data as an example for experimental verification and analysis, the results show that: (1) The deformation spatial position obtained after image compression is highly consistent with that before image compression and the PS InSAR method, and the deformation rate value is similar, indicating that the proposed compression method is feasible and has high reliability. (2) When the compressed subspace is 2×3 and 2×5, the computational efficiency is improved by about 24 times and 40 times respectively, and it can meet the monitoring accuracy requirements. When the compressed subspace is 2×10 and 2×15, the information loss increased, but the deformation position could still be identified and its computational efficiency increased by about 80 times and 120 times, respectively. Therefore, in practical applications, the selection of compressed subspace size should be based on computational efficiency and monitoring refinement. Conclusions: The research results provide a new data processing method for wide-area and long-term surface deformation monitoring, which can effectively improve the computational efficiency of time-series InSAR.