Skip to Content
MDPIMDPI

Advanced

by
  • Chengzhi Chen 1,†,
  • Haoran Hu 1,† and
  • Sirui Tian 1,*
  • et al.

Reviewer 1: Anonymous Reviewer 2: Anonymous Reviewer 3: Anonymous

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The problem of high-quality ISAR imaging from incomplete observation data is of practical relevance in radar applications, and the idea of combining sparsity and low-rank priors to regularize the ill-posed inverse problem is technically sound. The non-convex formulation to alleviate estimation bias from convex relaxations follows a reasonable research direction, and the experimental results show consistent performance gains over the selected baselines. However, the manuscript currently has several limitations.

  1. The algorithmic framework (ADMM + IRLS + TSVD for rank constraint) is a relatively standard combination in the optimization field. The authors should clearly articulate the unique technical contributions of this work, rather than restating each component separately.
  2. Standard ADMM convergence guarantees only hold for convex problems. For non-convex settings, ADMM may diverge or converge to poor local minima depending on initialization and penalty parameters. The authors should at minimum: (i) prove that the objective value is monotonically non-increasing during iterations; (ii) discuss conditions under which the algorithm converges to a critical point; or (iii) explicitly acknowledge the lack of a global convergence guarantee and support the empirical convergence claim with quantitative evidence (e.g., objective value curves across different initializations); or (iiii) introduce quad-level inshore awareness automatic marine ship surveillance using satellite sar system.
  3. The rank bound r is a critical hyperparameter of the proposed method, but the manuscript never explains how r is chosen in experiments, nor does it analyze its impact on imaging performance. It is necessarily to sea-state-aware conflict rectification under dual-granularity learning framework marine sar panoptic segmentation. For real measured data, it is unclear how the appropriate rank is determined without ground truth.
  4. The physical justification for the low-rank property of the ISAR echo matrix is insufficient. The authors only cite reference [44] and state that "the column correlation of matrix Y is strong", but do not explain the relationship between the rank value and target characteristics (e.g., number of scatterers, motion parameters, CPI length). The proposed method needs to be discussed for reference in ship imaging and language-prompted visual query learning ship-sea-land disambiguation in maritime panoptic segmentation from sar imagery. An ablation study on the rank parameter should be added to show how imaging performance varies with r, and provide practical guidance for parameter selection.
  5. Only three comparison methods are included. Missing comparisons include: (i) other non-convex sparse imaging methods; (ii) non-convex low-rank methods such as truncated nuclear norm minimization; (iii) other joint sparse-low-rank methods with non-convex relaxations. This makes the performance superiority claim unconvincing. Why is it used in tlsa triple-level speckle awareness framework speckle-robust marine ship surveillance using satellite synthetic aperture radar and density knowledge mining quantity-aware marine vessel surveillance using satellite sar data? Multiple figures have mislabeled subplots and typos (e.g., "QRDA", "rangeim" in Figure 4; inconsistent subfigure captions). All figures should be carefully checked and corrected. Axis labels and legends in Figures 5 and 8 are too small and hard to read. Please enlarge the fonts and ensure all legend entries are correct. All ISAR image plots should clearly label the axes as "Range cell" and "Cross-range cell" for clarity. The caption of Table 3 should explicitly state the reference image used for PSNR and SSIM calculation on real data.
  6. The experimental description says missing pulses are "set to zero", which implies an element-wise mask rather than dimensionality reduction. Please align the mathematical model with the experimental setup. How about language-guided semantic field query alignment sar ship-sea-land panoptic segmentation? The claim "establishing it as a state-of-the-art approach" in Section 5.2 is overstated given the limited comparisons. Please tone down the conclusion to match the experimental evidence.
  7. The related work section is poorly structured, with long lists of papers without clear categorization and comparison. It should be reorganized to systematically review sparse-only methods, low-rank-only methods, and joint sparse-low-rank methods, and highlight the research gap that this work addresses. Recent works on non-convex ISAR imaging are missing from the literature review. Please supplement relevant studies and discuss how the proposed method differs from them. Reference formatting is inconsistent and some entries are incomplete. Please follow the Remote Sensing reference style strictly.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The proposed manuscript presents an incomplete observations ISAR imaging method based on ESIIA, a non-convex joint sparse and low-rank optimization method. The framework exploits the sparsity of the problem by using an L1/2 norm sparsity prior and it also explicitly introduces the rank constraint to minimize amplitude shrinkage while maintaining the structure of the target. The reconstruction is evaluated under the noisy and missing data conditions by both simulated and measured experiments.

An ADMM based iterative framework is proposed to solve the optimization problem. The sparse image component is iteratively reweighted least squares (IRWL) and the low rank echo component is recovered by the truncated singular value decomposition (TSVD). The experiments are conducted at various SNRs and sampling rates for comparison with the methods of ESIIA, sparse recovery and joint low-rank–sparse reconstruction. In general, the PSNR, contrast, SSIM and the image entropy of ESIIA are higher than those of the other methods, especially when the SNR is lower and the sampling rate is lower. The measured Yak-42 data gives PSNR of 41.22 dB and SSIM of 0.87 for the chosen attitude of the data.

The manuscript considers an ISAR reconstruction problem that is relevant and introduces a potentially useful non-convex formulation. But there are some significant issues that are missing in the convergence guarantees, parameter-selection fairness, component-level ablation, statistical reliability, and computational efficiency in practice. The problems should be addressed with more precise mathematical arguments and further experiments. Thus, it is necessary to make a significant revision, as the manuscript is then eligible for publication.

For Major and Monor Comments: See the Word File

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript proposes an enhanced sparse ISAR imaging algorithm, termed ESIIA, for incomplete echo data. The proposed method jointly exploits the sparsity of the ISAR image and the low-rank property of the echo matrix. Specifically, the conventional -norm regularization is replaced by an -norm regularizer, while nuclear-norm relaxation is replaced by an explicit rank constraint. The resulting optimization problem is solved using an ADMM framework combined with iterative reweighting and truncated singular value decomposition. Simulated aircraft point-scatterer data and measured Yak-42 data are used for validation, and the authors report improved performance over RDA, L1SA, and JLRSA, particularly under low-SNR and low-sampling-rate conditions.

The research problem is relevant, the manuscript is generally well organized, and the numerical results indicate that the proposed method may have potential. However, several important issues remain regarding the validity of the model assumptions, the novelty of the contribution, the convergence of the non-convex algorithm, the fairness and reproducibility of the experiments, and the interpretation of the measured-data results. In particular, the relationship between the low-rank constraint and the sparse imaging model has not been sufficiently clarified, and the current experiments do not adequately support strong claims such as “state-of-the-art,” “unbiased estimation,” and “high computational efficiency.” Therefore, substantial revisions are required before the manuscript can be reconsidered.

1.The claim of computational efficiency is not supported by runtime experiments

The manuscript claims that ESIIA is computationally efficient because its updates are available in closed form and no inner iterations are required. However, each iteration still requires truncated SVD, which is often the most computationally expensive part of the algorithm. No actual runtime results or hardware specifications are provided.

The authors should report, on the same hardware and software platform:
1) total runtime for each method;
2) average time per iteration;
3) total number of iterations;
4) runtime scaling with increasing matrix dimensions;
5) whether parameter grid-search time is included in the reported runtime.
without such results, claims such as “high computational efficiency” and “improved speed” should be removed or substantially weakened.

2.The reference image used for PSNR and SSIM in the measured-data experiment is unclear

The definitions of PSNR and SSIM require a reference image Sref. However, a true scattering-coefficient image is not available for the measured Yak-42 data. Nevertheless, Table 3 reports PSNR and SSIM values without clearly defining the reference image.

The authors should clarify:
1) whether the full-aperture RDA image is used as the reference;
2) whether the reference and reconstructed images are normalized, thresholded, or subjected to dynamic-range truncation;
3) why the full-aperture RDA image can be treated as a ground-truth reference;
If the full-aperture RDA image is used as the reference, then PSNR and SSIM measure similarity to the RDA result rather than reconstruction error relative to the true target scattering distribution. The corresponding conclusions should therefore be stated more cautiously.

3.The parameter settings are incomplete, and the experiments are not sufficiently reproducible
The proposed algorithm involves several important parameters, including
                                                             λ,r,ρ1​,ρ2​,ε,κ,Kiter​.

However, Algorithm 1 lists only some of these parameters as inputs. The regularization parameter λ, the IRLS stabilization parameter ε, and the maximum number of iterations are not clearly specified.
The experimental section states only that a grid search was used to identify parameter combinations that optimize image quality. The search ranges, step sizes, evaluation criteria, and final parameter values are not reported. This raises several concerns:
1)Were ground-truth PSNR or SSIM values used to select parameters in the simulations?
2)How were parameters selected for the measured data, where the true scattering image is unavailable?
3)How would the algorithm be used when the noise level is unknown?

4.There are many grammar and expression issues throughout the text, such as subject verb inconsistency, misuse of articles, and unnatural technical expressions. It is recommended to conduct a comprehensive English polishing.

5.Please unify M and N to represent distance dimension and azimuth dimension respectively. The definitions of different positions in the text seem to be interchangeable

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

A great deal of work has been done.


Minor Comments:
1. For the revised manuscript, the main method used for establishing convergence of the non-convex ADMM scheme is still monotonic objective decrease, which is used in section 3.5 – Convergence Analysis. Please strengthen your theory grounds or at least ensure that you report objective evolution along with primal/dual residuals and behavior of the stopping conditions under representative experimental conditions.

2. Sections 4.1–4.2 – Experimental Validation: repeated-trial statistics and an ablation of the L1/2 prior and rank constraint is still missing from the manuscript. Please report mean±SD for several random noise/sampling realizations and baseline, L Different variants of 1/2, rank-only and full ESIIA under the same conditions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have addressed all questions excellently, and I would like to recommend acceptance.

Author Response

We thank you very much for your comprehension and recognition of our work. We also appreciate you for your previous professional comments on our article. These comments are very helpful for improving the quality of our work. On behalf of all the authors of this paper, we thank you again for your friendly recognition.