the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 01 Sep 2025
A New Hybrid Framework for Digital Terrain Modeling Using Sector-Based Neighbor Selection and Neural Network Blending
Abstract. This paper presents a new hybrid framework for digital terrain modeling that combines directional sector-based neighbor selection (DSNS), artificial neural networks (ANN), and gradient-based weighted blending. The framework addresses the spatial imbalance and ripple artifacts commonly seen in interpolation-based terrain models. In the first stage, 12 sector-divided neighbors are selected around each query location to ensure directional balance. Next, ANN models are trained on reference terrains using either expert-adjusted or natural interpolated surfaces, depending on the test region. Finally, a gradient-based weighting mechanism blends ANN outputs with those of linear interpolation to create a coherent and smooth elevation surface. The proposed method is validated on three real-world terrains of varying size and complexity. Results show that the model significantly improves topographic continuity, numerical stability, and generalization across different landscapes. Compared to conventional interpolation methods, the proposed method reduces oscillations, maintains terrain flow, and eliminates the need for manual adjustments. The framework offers a scalable, automated, and accurate approach for terrain surface reconstruction in both regular and anisotropic datasets.
- Preprint
(2560 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 20 Nov 2025)
-
RC1: 'Comment on egusphere-2025-2539', Anonymous Referee #1, 03 Sep 2025
reply
-
AC1: 'Reply on RC1', Kadir Akgol, 11 Sep 2025
reply
General Comments
We thank the referee for the constructive feedback. We have revised the manuscript accordingly. Specifically, we inserted a concise uncertainty/significance summary immediately below Table 3 in the Results and Discussion section, and we added a short pointer in the Conclusion that acknowledges these statistical checks without repeating numerical details. In addition, we harmonized the capitalization of method names as a technical correction. These edits improve clarity and traceability while leaving the scientific content and results unchanged.Specific Comments
Results & Discussion – addition below Table 3:
To quantify uncertainty and test significance, we computed 95% bootstrap confidence intervals (CIs) for RMSE/MAE and performed cell-wise paired comparisons. The CIs indicate that MATLAB linear attains the lowest errors (RMSE 0.062–0.112; MAE 0.0053–0.0123), the Proposed ANN+Blending ranks second (RMSE 0.085–0.117; MAE 0.046–0.053), NetCAD is comparable to the Proposed method in RMSE but worse in MAE (RMSE 0.117–0.164; MAE 0.017–0.028), and Natural is the least accurate (RMSE 0.176–0.206; MAE 0.094–0.107). Cell-wise paired tests confirmed that inter-method differences are statistically significant—for example, Proposed vs Linear (MAE: t-test p=4.58×10⁻²⁰⁹; Wilcoxon p<1×10⁻³⁰⁰; squared-error: t-test p=0.0031; Wilcoxon p<1×10⁻³⁰⁰)—corroborating the qualitative conclusions above.Conclusion – brief pointer (no repetition of numbers):
Complementary statistical analyses further support these findings. Bootstrap confidence intervals confirmed the robustness of the RMSE and MAE values, while paired tests demonstrated that inter-method differences were statistically significant. In particular, although MATLAB linear interpolation achieved the lowest error magnitudes, the Proposed ANN + Blending method offered a more balanced trade-off between accuracy and geomorphological realism, with its advantages consistently validated across all sites.Technical Corrections
Terminology consistency: We harmonized method names to MATLAB linear and MATLAB natural throughout the manuscript (text, figure captions, and tables), removing mixed capitalization variants (e.g., “Linear/Natural”).Citation: https://doi.org/10.5194/egusphere-2025-2539-AC1
-
AC1: 'Reply on RC1', Kadir Akgol, 11 Sep 2025
reply
-
CEC1: 'Comment on egusphere-2025-2539 - No compliance with the policy of the journal', Juan Antonio Añel, 11 Oct 2025
reply
Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".
https://www.geoscientific-model-development.net/policies/code_and_data_policy.htmlFirst, you have not shared the implementation of the ANN that you have developed for your work, neither the NetCAD implementation. Secondly, you do provide information on the NetCAD and Matlab versions that you have used. We need all this information to assure the replicability of your work. Therefore, the current situation with your manuscript is irregular. Please, publish your code in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible, as we can not accept manuscripts in Discussions that do not comply with our policy. Also, please, clarify the versions of the software that you have use it, and how to obtain it. Regarding Matlab, it would be good if you could confirm if the M code developed runs in GNU Octave, the free M Language interpreter.
Also, you must include a modified 'Code and Data Availability' section in a potentially reviewed manuscript, containing the information of the new repositories and version numbers for software.
I must note that if you do not fix this problem, we cannot accept your manuscript for publication in our journal.
Juan A. Añel
Geosci. Model Dev. Executive EditorCitation: https://doi.org/10.5194/egusphere-2025-2539-CEC1 -
AC2: 'Reply on CEC1', Kadir Akgol, 14 Oct 2025
reply
Dear Dr. Añel,
Thank you for your guidance. We have brought the submission into full compliance with GMD’s Code and Data Policy. We deposited the code and the data as two separate public archives, each updated to Version 2. The code archive (DOI: https://doi.org/10.6084/m9.figshare.29279717.v2) contains the MATLAB and Python implementations used in the study—DTM_DSNS_ANN_Blending_main.m, DTM_Performance_Comparison_and_DiffMaps.m, DTM_Contour_From_XYZ.m, and dtm_diffmaps_from_mat.py—with added usage notes and inline comments to make them public-ready without altering the algorithmic structure or defaults. The data archive (DOI: https://doi.org/10.6084/m9.figshare.29279729.v2) includes the NetCAD NCZ project files (source TINs), NetCAD-exported XLS point clouds for all sites, an extended grid XLS for Alemdar-1 to mirror NetCAD contours closely in MATLAB, and the evaluation matrices comparison.mat and alemdar.mat used for the results and visualizations. If any step remains unclear during verification, we will be happy to assist to ensure full reproducibility.
Regarding the NetCAD baseline, we used NetCAD GIS 8.5.5 with the NETSURF module (commercial, en.netcad.com) to build and, where necessary, edit TIN surfaces. As in Civil 3D, triangle-based surface modeling permits expert connectivity edits that legitimately affect the resulting surface; to make this auditable and reproducible, the NCZ projects are shared so readers can open the TINs in NETSURF, regenerate 1-m contours, and export grid points to XLS for downstream comparisons. Outside NetCAD, our MATLAB contouring uses griddata(...,'linear'), which we found to provide close parity with the NetCAD TIN for cross-platform visualization; this choice is for figure consistency and does not replace NetCAD in the scientific comparisons.
The modeling and ANN training were performed in MATLAB R2023b (Neural Network Toolbox). We also tested GNU Octave 10.3.0; however, we encountered blocking incompatibilities in two of the scripts—most notably the unavailability of griddata('natural'), which is critical for a key stage of the terrain modeling pipeline—so we could not proceed further in Octave. We provide fallbacks such as xlsread for data loading, but full reproducibility requires MATLAB. The accompanying Python script is used only for publication-quality visualizations from MATLAB-produced grids and does not affect the modeling results. Visual Studio Code (free) was used solely as an editor.
We have revised the manuscript’s Code Availability and Data Availability sections accordingly, inserting the separate DOIs referenced above. Please let us know if any additional metadata or adjustments would be helpful.
Sincerely,
Kadir Akgöl (on behalf of the authors)Citation: https://doi.org/10.5194/egusphere-2025-2539-AC2 -
CEC2: 'Reply on AC2', Juan Antonio Añel, 15 Oct 2025
reply
Dear authors,
Many thanks for your reply. We can consider now the current version of your manuscript in compliance with the Code and Data policy of the journal.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Citation: https://doi.org/10.5194/egusphere-2025-2539-CEC2
-
CEC2: 'Reply on AC2', Juan Antonio Añel, 15 Oct 2025
reply
-
AC2: 'Reply on CEC1', Kadir Akgol, 14 Oct 2025
reply
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,492 | 52 | 14 | 1,558 | 7 | 8 |
- HTML: 1,492
- PDF: 52
- XML: 14
- Total: 1,558
- BibTeX: 7
- EndNote: 8
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript presents a well-motivated and carefully executed contribution to digital terrain modeling by combining a directional sector–based neighbor selection with an ANN regressor and a gradient-based blending strategy. The evaluation over three real sites against established baselines (NetCAD TIN, MATLAB linear, MATLAB natural) is convincing, and the writing is clear. In my view, the paper is suitable for publication in its current form.
Two minor, optional suggestions could further strengthen the presentation without being prerequisites for acceptance. First, in the Results and Discussion section, directly below Table 3, adding a single, concise sentence that summarizes the uncertainty and significance checks (e.g., that 95% bootstrap CIs were computed for RMSE/MAE and paired comparisons showed inter-method differences to be statistically significant) would make the quantitative conclusions more explicit. If desired, a brief, number-free nod to these checks can also be included in the Conclusion to close the loop, but this is optional.
Second, for polish, I recommend a quick pass for terminology consistency limited to method names—use a single convention for MATLAB linear and MATLAB natural throughout (avoid mixing “Linear/Natural” vs “linear/natural”). This small edit will improve stylistic uniformity.
Overall, I enjoyed reading the paper. The method is practical, the validation is sound, and the manuscript is well aligned with GMD’s scope. The optional tweaks above would only enhance clarity; the work is acceptable as is.