Abstract:Deep neural networks based on unrolled iterative algorithms have achieved remarkable success in sparse reconstruction applications, such as synthetic aperture radar (SAR) tomographic inversion (TomoSAR). However, the currently available deep learning-based TomoSAR algorithms are limited to three-dimensional (3D) reconstruction. The extension of deep learning-based algorithms to four-dimensional (4D) imaging, i.e., differential TomoSAR (D-TomoSAR) applications, is impeded mainly due to the high-dimensional weight matrices required by the network designed for D-TomoSAR inversion, which typically contain millions of freely trainable parameters. Learning such huge number of weights requires an enormous number of training samples, resulting in a large memory burden and excessive time consumption. To tackle this issue, we propose an efficient and accurate algorithm called HyperLISTA-ABT. The weights in HyperLISTA-ABT are determined in an analytical way according to a minimum coherence criterion, trimming the model down to an ultra-light one with only three hyperparameters. Additionally, HyperLISTA-ABT improves the global thresholding by utilizing an adaptive blockwise thresholding scheme, which applies block-coordinate techniques and conducts thresholding in local blocks, so that weak expressions and local features can be retained in the shrinkage step layer by layer. Simulations were performed and demonstrated the effectiveness of our approach, showing that HyperLISTA-ABT achieves superior computational efficiency and with no significant performance degradation compared to state-of-the-art methods. Real data experiments showed that a high-quality 4D point cloud could be reconstructed over a large area by the proposed HyperLISTA-ABT with affordable computational resources and in a fast time.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the high - dimensional weight matrix learning problem encountered when applying deep - learning algorithms for four - dimensional (4D) reconstruction in differential synthetic aperture radar tomography (D - TomoSAR). Specifically, the existing TomoSAR algorithms based on deep learning are mainly limited to three - dimensional (3D) reconstruction. When extended to four - dimensional imaging that includes temporal variations (such as seasonal movement or linear subsidence), the dimension of the weight matrix that the network needs to learn is very high, usually containing millions of free parameters. This not only requires a large number of training samples, resulting in a huge memory burden and excessive time consumption, but is also difficult to implement in practical applications. To solve these problems, the paper proposes an efficient and accurate algorithm named HyperLISTA - ABT. ### Main Contributions 1. **Proposing the HyperLISTA - ABT algorithm**: This is the first time that a deep neural network has been applied to D - TomoSAR and multi - component D - TomoSAR inversion. 2. **Introducing the Adaptive Block Thresholding (ABT) scheme**: Through the block - coordinate technique, a local thresholding method is used instead of global thresholding, thereby preserving the local characteristics of weak signals. 3. **System performance evaluation**: Through experiments on simulated data and real data, the competitiveness and superior computational efficiency of HyperLISTA - ABT are verified, showing its application potential in large - scale D - TomoSAR processing, such as 4D point - cloud reconstruction in Las Vegas. ### Method Overview - **HyperLISTA - ABT**: Based on HyperLISTA, the weight matrix is determined by analyzing the optimization method, and an adaptive block thresholding scheme is adopted. The determination of the weight matrix is based on the minimum coherence criterion and optimized by an analytical method, reducing the training complexity. The adaptive block thresholding scheme allows thresholding within each local block, thereby better preserving the local characteristics of weak signals. - **Update rules**: \[ \gamma_{k + 1}^{ip}=\eta_{\theta_k^{ip}}(\gamma_k^{ip}+W_{ip}^T(y - R_{ip}\gamma_k^{ip})+\beta_k^{ip}(\gamma_k^{ip}-\gamma_{k - 1}^{ip})) \] where: \[ \theta_k^{ip}=c_1\|R_{ip}^+(R_{ip}\gamma_k^{ip}-g)\|_1 \] \[ \beta_k^{ip}=c_2\|\gamma_k^{ip}\|_0 \] ### Experimental Results - **Effective detection rate**: Under different signal - to - noise ratio (SNR) conditions, HyperLISTA - ABT shows a higher effective detection rate compared to the original HyperLISTA and other existing methods, especially in the case of high amplitude ratios. - **Computational efficiency**: HyperLISTA - ABT shows a significant computational efficiency advantage when processing large - scale data. Compared with existing deep - learning methods and traditional CS methods, the required training time and computational resources are greatly reduced. In conclusion, through proposing the HyperLISTA - ABT algorithm, this paper successfully solves the problem of high - dimensional weight matrix learning in D - TomoSAR inversion, providing an efficient and accurate solution for large - scale D - TomoSAR processing.

HyperLISTA-ABT: An Ultra-light Unfolded Network for Accurate Multi-component Differential Tomographic SAR Inversion

TomoSAR-ALISTA: Efficient TomoSAR Imaging via Deep Unfolded Network

ATASI-Net: An Efficient Sparse Reconstruction Network for Tomographic SAR Imaging with Adaptive Threshold

AETomo-Net: A Novel Deep Learning Network for Tomographic SAR Imaging Based on Multi-dimensional Features

High Resolution SAR Tomography 3-D Imaging Via Sparse Recovery Deep Learning Network

Improved Analytic Learned Iterative Shrinkage Thresholding Algorithm and Its Application to Tomographic Synthetic Aperture Radar Building Object Height Inversion

Super-Resolving Sar Tomography Using Deep Learning.

Deep SAR Tomography: A Model-Inspired Approach With Learned Sparse Regularizer

An Efficient Sparse Imaging Method of TomoSAR Via Unfolded Deep Network

Adatomo-Net: a Novel Deep Learning Approach for SAR Tomography Imaging and Autofocusing

Efficient SAR Tomographic Inversion Via Sparse Bayesian Learning

gamma-Net: Superresolving SAR Tomographic Inversion via Deep Learning

γ-Net: Superresolving SAR Tomographic Inversion via Deep Learning

Single Target SAR 3D Reconstruction Based on Deep Learning

$\boldsymbolγ$-Net: Superresolving SAR Tomographic Inversion via Deep Learning

URBAN 3D RECONSTRUCTION OF VHR SAR IMAGES USING ITERATIVE OPTIMIZATION ALGORITHM AND LAYOVER FIXED-ORDER MODEL

An Iterative Multidimensional Feature Reconstruction Network for Tomographic SAR Imaging

Gridless Tomographic SAR Imaging Based on Accelerated Atomic Norm Minimization with Efficiency

TomoSAR Sparse 3-D Imaging Via DEM-Aided Surface Projection.

Sar Tomography Based on Reweighted Atomic Norm Minimization