Topology-Aware Field Parcel Delineation: Bridging Deep Semantic Features and Geometric Constraints

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Detailed review comments can be found in the uploaded attachment.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.
*Please pay attention to the atteachments. we found that the firstly uploaded file may have problem to for reading, so we submit an 'v2' version along with this session.*

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes an innovative topology-aware field parcel delineation framework (TopoFP), which integrates deep feature learning with topological constraints to effectively address geometric consistency issues such as boundary fragmentation and adhesion. Its "topology-first" vectorization paradigm, boundary width implicit encoding strategy (e.g., DLD/DLE algorithms), and the "robust baseline + explicit constraints" approach possess significant theoretical and practical value. The method outperforms complex SOTA models across multiple test regions, providing a high-precision, topologically correct solution for automated vector mapping. It is a study with a novel methodology and solid validation.
1.  The paper lacks in-depth analysis of the specific details and robustness of the point and line classification algorithm within the Topological Relationship Construction (TRC) module, especially its processing logic at complex junctions. It is recommended to supplement the methods section or appendix with detailed pseudocode or flowcharts, and to discuss potential failure cases and countermeasures for extreme shapes (e.g., acute angles, irregular curves).
2.  The key parameters for the Double-Line Detection (DLD) and Dangling Line Extension (DLE) modules (e.g., width threshold, maximum extension length) appear to rely on empirical or dataset-specific choices. A more systematic parameter sensitivity analysis is recommended to enhance the method's generalizability to images of different resolutions or across different geographical regions.
3.  While comparisons are made with specific field parcel extraction models, recent general-purpose, high-performance instance segmentation or vectorization methods (such as improved Mask R-CNN series or Transformer-based extractors) are not included. It is suggested to add comparative experiments with such general models to more comprehensively position the performance and advantages of the proposed framework within a broader research context.
4.  The "non-cultivated land" category is included in the method, but its role in subsequent topological construction and vectorization is not clearly explained. It is recommended to provide a more detailed explanation of how the features of "non-cultivated land" areas are utilized or excluded from the final parcel polygon generation, and how this affects processing fields with complex backgrounds like woods or buildings.
5.  The TRC heavily relies on a skeletonization algorithm to convert planar boundaries to lines. The document mentions that traditional skeletonization causes topological distortion for double-line structures but does not deeply discuss the degree to which the chosen skeletonization method mitigates this issue, or the potential new problems it might introduce (e.g., skeleton jitter, centerline offset). A more in-depth discussion on this is suggested.
6.  While a code link is provided, the reviewed document does not contain the complete training hyperparameter settings, all data preprocessing details (e.g., the specific implementation of linear stretching), or a full environment description (e.g., PyTorch, CUDA versions) for reproducing the experiments. It is recommended to provide a detailed Reproducibility Checklist in the code repository or the paper's appendix.

Author Response

Please see the attachment.

*Please pay attention to the atteachments. we found that the firstly uploaded file may have problem to for reading, so we submit an 'v2' version along with this session.*

Author Response File: Author Response.docx