A Discrete Grid-Based Approach for Efficient Near-Optimal Coverage Selection in a Large-Scale Remote Sensing Image Dataset

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper addresses the inefficiency of traditional vector topology-based method for regional coverage retrieval in large-scale remote sensing imagery. It proposes a coverage optimization and selection method based on DGGS grid encoding and a greedy strategy, and validates its efficiency, scalability, and robustness in the Beijing, Shaanxi, and Japan study areas. The study offers significant practical value and strong engineering relevance. However, the following issues should be clarified or revised:

 

1、The conclusions regarding the selection of grid levels 6 and 7 should be better clarified across different scenarios.

 

The manuscript states in the Abstract and Conclusions that level 7 provides the best trade-off between accuracy and efficiency. However, the robustness experiment suggests that, in most practical scenarios, level 6 can achieve reliable coverage estimation with substantially lower computational cost. Although these two conclusions are not necessarily inconsistent, the current presentation may cause confusion. The authors should clearly distinguish between different application objectives: level 7 may be more suitable when near-complete coverage selection is required, whereas level 6 may be more practical for rapid coverage estimation or applications under sparse-coverage scenarios.

 

2、Further explanation is needed as to why the DGGS-based method selects fewer images than the vector topology-based method.

 

The manuscript presents two full-coverage experiments. In the Beijing experiment, the vector-based method selects 109 images, whereas DGGS level 7 selects 105 images. In the Shaanxi experiment, the vector-based method selects 1,337 images, whereas DGGS level 7 selects 1,294 images. The authors should provide a more detailed analysis of why fewer images are selected by the DGGS-based method. Specifically, they should compare the spatial locations, coverage contributions, and boundary-area distributions of the images that differ between the results of the two methods.

 

3、A comparison between the DGGS-based method and the vector topology-based method should be added for the Japan study area.

 

In the Beijing and Shaanxi experiments, the manuscript provides comparisons between the DGGS-based method and the vector topology-based method, demonstrating the advantages of the proposed method in terms of efficiency and scalability. However, such a comparison is absent in the Japan experiment, which represents a sparse-coverage scenario. The authors are advised to supplement this experiment to further evaluate whether the DGGS-based method still outperforms the vector topology-based method under sparse-coverage conditions.

 

Overall recommendation:

 

This paper presents a certain degree of innovation, and the experimental results support the efficiency advantage of the DGGS-based method in large-scale coverage retrieval. If the above issues are adequately addressed, the overall quality and persuasiveness of the manuscript would be significantly improved.

Author Response

Dear Reviewer,

We sincerely thank you for your careful review of our manuscript and for your valuable and constructive comments. We have benefited greatly from your suggestions.

We have carefully addressed each of your comments point by point. Please find our detailed responses in the attached document.

Thank you again for your time and insightful feedback.

Sincerely,
Tuo

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

We recommend that corrective measures be introduced to address these two issues identified in the proposed methodology:

1) The proposed DGGS‑based approach entails an intrinsic loss of geometric precision due to spatial discretization, which may affect boundary accuracy in applications requiring strict geometric exactness. In addition, its performance is highly sensitive to the selected grid resolution, with accuracy and computational cost depending on empirically tuned DGGS levels, limiting full automation and generalizability.

2) The use of a heuristic‑guided greedy strategy ensures scalability but does not guarantee global optimality, potentially leading to sub‑optimal image sets in highly redundant or complex scenarios. Finally, the framework adopts a static and predominantly spatial formulation, incorporating temporal and quality factors only as ranking criteria, which restricts its applicability to dynamic or explicitly spatio‑temporal optimization tasks.

Author Response

Dear Reviewer,

We sincerely thank you for your careful review of our manuscript and for your valuable and constructive comments. We have benefited greatly from your suggestions.

We have carefully addressed each of your comments point by point. Please find our detailed responses in the attached document.

Thank you again for your time and insightful feedback.

Sincerely,
Tuo

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The core idea of this manuscript—reducing vector topological operations to set operations on DGGS grid symbol sequences—presents a meaningful direction for large-scale remote sensing data processing. In addition, the experiments are generally well-designed, as they are not limited to a single case study but instead include datasets with diverse scales and coverage conditions. Nevertheless, the current manuscript would benefit from further refinement in several aspects, including (1) the lack of a clearly defined criterion for the “greedy” algorithm, (2) potentially overstated use of the term “optimal” and interpretation of the reported acceleration factors, (3) questionable interpretation of the Experiment IV results, (4) insufficient reproducibility of the baseline implementation, and (5) several editorial and organizational issues.

 

Title; Lines 33–38: The expression “Optimal coverage selection” appears somewhat overstated in the current context. Since the set cover problem is NP-hard and the proposed framework relies on a heuristic approximation strategy, repeated use of the term “optimal” throughout the title and manuscript may unintentionally give the impression of theoretical optimality. It may therefore be more appropriate to consistently adopt expressions such as “near-optimal” or “heuristically selected.” Likewise, revising the title to something like “Efficient Near-Optimal Coverage Selection” could improve terminological precision while preserving the work's contribution.

 

Lines 19–20: The reported acceleration factors (1.57×, 10.9×, and 11×) would benefit from clearer and more consistent interpretation. While the abstract and conclusion emphasize “up to 11×,” Experiment I (Beijing dataset) reports only approximately 1.57× improvement (6109 ms → 3882 ms), whereas the 10.9× acceleration is achieved only in Experiment III (Shanxi dataset). The conclusion appropriately presents the range (“1.5× to up to 11×”), but the abstract highlights only the upper bound, which may unintentionally overstate the overall performance trend. In addition, the manuscript does not currently provide a theoretical explanation for why the same algorithm exhibits substantially different acceleration ratios across datasets of different scales. Including a complexity analysis with respect to the number of candidate images n and the ROI grid cells ∣Z∣, together with a regression-based scalability analysis, would strengthen the claim of near-linear scalability.

 

Line 162: Although H3 is an open-source project developed by Uber, reference [27] currently cites only a website URL. In this field, it is generally preferable to cite additional academic or technical references that describe the H3 framework itself, such as the OGC DGGS Abstract Specification (OGC 15-104r5) or related scholarly publications, to provide a more rigorous methodological foundation.

 

Lines 189–194: The recommendation that “the ratio between the image coverage area and the grid cell area falls within the range of 10–100” is potentially useful, but the manuscript does not explain how this guideline was derived. It would be helpful to provide a dedicated ablation experiment—for example, sweeping DGGS levels while varying image size over the same ROI—to quantitatively justify this recommendation.

 

Lines 196–224: The proposed heuristic determines the “current optimal image” through a lexicographic comparison based on (i) acquisition time, (ii) spatial resolution, and (iii) cloud coverage. However, the procedure does not appear to explicitly incorporate the contribution of uncovered cells into the evaluation criterion. As a consequence, a large image covering many uncovered regions could potentially be ranked behind a smaller image simply because of slightly inferior temporal or spatial characteristics, thereby increasing the total number of selected images. This issue directly affects both the validity of the heuristic and the appropriateness of describing it as “greedy.” It would therefore be valuable either (i) to explicitly incorporate uncovered-cell contribution as a primary criterion or tie-breaker, or (ii) to explain why it was intentionally excluded and compare the resulting behavior empirically.

In addition, the practical effectiveness of the lexicographic heuristic itself remains somewhat unclear. In datasets such as the Beijing experiment (2023-08-03 to 2023-08-31), where acquisition times are concentrated within a narrow temporal window and spatial resolutions are also highly similar (0.5–0.8 m), many candidates are likely to remain tied after the first comparison stages. Under such conditions, the final selection may effectively depend on the arbitrary rule that “the first image is selected.” Furthermore, comparisons against alternative decision strategies—such as weighted scoring schemes or Pareto-optimal selection—are currently absent. Since the manuscript emphasizes “flexibility,” it would strengthen the discussion to report the actual priority settings used in the experiments and to include sensitivity analysis under different priority orders or weighting schemes.

 

Lines 250–253: The statement that “A short-circuit evaluation strategy is employed to terminate the process once a valid grid cell is identified” is somewhat difficult to interpret operationally. It may improve clarity if the pseudocode explicitly indicates which loop is terminated and under what exact condition the short-circuit operation is triggered.

 

Lines 265–269, 323–330: The implementation details of the “vector-based topological method,” which serves as the primary baseline, are currently insufficiently specified. Important details such as the geometry library used (e.g., JTS, GEOS, Shapely, Java Topology Suite, OGC Simple Features), whether spatial indexing structures (R-tree, STR-tree, etc.) were enabled, coordinate systems, the order of intersection/union/difference operations, and whether geometry simplification was applied are not described. As a result, it is difficult to determine whether the reported acceleration—particularly the 10.9× case—originates primarily from the proposed algorithmic formulation or from differences in implementation efficiency. To improve reproducibility and fairness, it would be beneficial to (i) provide pseudocode and implementation details of the baseline in an appendix, and (ii) additionally compare against an optimized vector baseline equipped with spatial indexing.

 

Lines 326, 336–339: Figure 6 contains six subplots ((a)–(f)), yet the main text (Line 326) refers only to panels (b), (e), and Table 1. Since panels (c), (d), and (f) are discussed later in Section 4.3 (Parameter sensitivity), it may improve readability either to explicitly mention this transition in the text or to separate the figure into multiple figures according to experimental purpose.

 

Lines 358, Table 2: When increasing the DGGS level from 7 to 8, computation time increases by approximately 15.8×, whereas the corresponding accuracy improvement is only 0.01 percentage points. Given that the number of H3 cells increases approximately 7-fold per level, the observed runtime growth appears strongly nonlinear. Additional profiling or analysis identifying the dominant bottleneck—such as memory access overhead, hash operations, or image-grid mapping stages—would help clarify the source of this nonlinearity.

 

Lines 416–420; Table 4: The four subsets containing 210, 563, 1033, and 1464 images are not clearly described. Specifically, it is unclear whether these subsets are (i) randomly sampled from the 1464-image dataset, (ii) cumulatively expanded in temporal order, or (iii) independently collected datasets. Since the subset construction directly affects the interpretation of “initial coverage ratio,” the generation procedure should be explicitly documented. If random sampling was used, reporting the random seed would further improve reproducibility.

 

Lines 421–429; Table 4: Table 4 indicates that, across all four datasets, the DGGS-based estimates at levels 6 and 7 are consistently smaller than the original coverage ratios. For example: original 18.2667% → L6 18.2130%, L7 18.2632%; 50.0458% → L6 49.9114%, L7 50.0350%; 80.5484% → L6 80.2331%, L7 80.5270%; and 95.4946% → L6 94.9573%, L7 95.4666%. All eight comparisons, therefore, exhibit the same sign direction. This pattern appears more naturally interpretable as a systematic underestimation arising from discretization effects near image boundaries (or from ROI cells being evaluated as outside image regions). However, Lines 426–427 state that “the estimated coverage remains stable without any systematic bias.” This interpretation may therefore require reconsideration. It would strengthen the analysis to (i) statistically evaluate the sign consistency (e.g., using a sign test), and/or (ii) explicitly define the inclusion rule for partially intersected cells (e.g., centroid-based inclusion, area thresholding, etc.) and discuss the resulting discretization bias more transparently.

 

Minor Issues

Figure 5 (Line 257): The visual difference between “(a) Before filtering” and “(b) After filtering” is currently rather subtle. Readability may improve if the removed images are highlighted in a distinct color or with shading in an auxiliary panel.

 

Lines 167, 173: The expressions “as shown in Figures 2” and “as shown in Figures 3” refer to single figures and should therefore be revised to “Figure 2” and “Figure 3,” respectively.

 

Lines 447, 519, 552: The section numbering and organization appear inconsistent. After “4. Experiments,” the manuscript proceeds to “4. Discussion,” followed by “5. Conclusions” and “6. Patents” (Line 552). Revising the numbering hierarchy and removing empty or unnecessary sections would improve editorial consistency and readability.

Author Response

Dear Reviewer,

We sincerely thank you for your careful review of our manuscript and for your valuable and constructive comments. We have benefited greatly from your suggestions.

We have carefully addressed each of your comments point by point. Please find our detailed responses in the attached document.

Thank you again for your time and insightful feedback.

Sincerely,
Tuo

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

.