LunarDepthNet: Generation of Digital Elevation Models using Deep Learning and Monocular Satellite Images

Aaranay Aadi; Amitabh; Jai Gopal Singla; Nitant Dube; Praveen Kumar Shukla; Vijaypal Singh Dhaka

arxiv: 2604.22848 · v1 · submitted 2026-04-22 · 💻 cs.CV

LunarDepthNet: Generation of Digital Elevation Models using Deep Learning and Monocular Satellite Images

Aaranay Aadi , Jai Gopal Singla , Amitabh , Nitant Dube , Praveen Kumar Shukla , Vijaypal Singh Dhaka This is my paper

Pith reviewed 2026-05-10 00:47 UTC · model grok-4.3

classification 💻 cs.CV

keywords lunar surfacedigital elevation modelmonocular depth estimationdeep learningUNetChandrayaan-2terrain mapping

0 comments

The pith

A neural network creates lunar elevation maps from single orbital images by reading surface shadows.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents LunarDepthNet, a deep learning model designed to produce digital elevation models of the lunar surface using only one image captured by the Chandrayaan-2 Terrain Mapping Camera. The approach trains a UNet architecture with an EfficientNet encoder to interpret how light and shadow patterns indicate actual terrain heights, using paired image and DTM data for supervision. A combined loss function helps maintain both accuracy in elevation values and smoothness in the resulting maps. Validation shows stable training with 12% loss convergence, and testing yields a normalized root mean square error of 0.437 along with a mean absolute error of 4.5 meters. If reliable, this method would allow creation of elevation data for lunar areas lacking the stereo image pairs traditionally needed for such mapping.

Core claim

LunarDepthNet employs a UNet structure with an EfficientNet encoder and custom layers to learn the relationship between shading in monocular TMC images and corresponding elevation values from DTMs, achieving a mean nRMSE of 0.437 and MAE of 4.5m on held-out test data while maintaining terrain details through a combined loss.

What carries the argument

LunarDepthNet: a UNet-based convolutional network with EfficientNet encoder that maps monocular image intensity patterns to per-pixel elevation estimates.

Load-bearing premise

Shading patterns observed in the training images correspond to elevation in a consistent manner that holds for all other lunar terrains regardless of local lighting or material differences.

What would settle it

Running the model on TMC images from a previously unseen lunar area and comparing the output DEM against an independent high-resolution stereo DTM to verify if the errors stay within 4.5m MAE without large biases.

Figures

Figures reproduced from arXiv: 2604.22848 by Aaranay Aadi, Amitabh, Jai Gopal Singla, Nitant Dube, Praveen Kumar Shukla, Vijaypal Singh Dhaka.

**Figure 2.** Figure 2: Layer-wise architecture of LunarDepthNet: The model uses SE-enhanced decoder blocks and an EfficientNet [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of relative elevation profiles showing the alignment of the predicted DEM (red) with the ground [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of absolute elevation profiles, demonstrating the post-hoc scaling technique that uses the original [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

Recent times have seen an increase in demand of high quality Digital Elevation Models (DEMs) for the lunar surface, because they are highly important for studying the moon and planning future missions. However, there is an evident lack of detailed elevation data on the Moon. To overcome this limitation, this study proposes a novel deep learning method that estimates and generates a surface elevation map directly from monocular images of the surface. The dataset used comprises of the Chandrayaan-2 Terrain Mapping Camera (TMC) images with their corresponding Digital Terrain Models (DTMs). The study proposes LunarDepthNet, which comprises of a UNet architecture to generate DEMS. It incorporates an EfficientNet encoder and custom layers to correctly learn how the light shadows on the surface relate to the actual elevation values. A combined loss function was also utilized to keep the terrain details accurate and smooth. During validation, the model showed a stable loss convergence of 12%. It achieved a mean nRMSE of 0.437 and an MAE of 4.5m in the testing stage. These results prove the model can generate dependable elevation maps from single orbital images, which are quite useful in regions of the moon where stereo-images are not available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Applies UNet with EfficientNet to Chandrayaan-2 monocular DEMs but shows no evidence the mapping generalizes beyond training data.

read the letter

This paper trains a UNet with EfficientNet encoder on Chandrayaan-2 TMC images paired with DTMs to output elevation from single orbital shots. It reports nRMSE 0.437 and MAE 4.5 m after reaching 12% loss convergence on an unspecified test set. The core idea targets a practical gap where stereo coverage is missing on the Moon. That part is straightforward and addresses a real need for mission planning and science in data-sparse regions. The architecture choice and combined loss for preserving terrain details are sensible defaults for this kind of regression task. What stands out is the specific application to TMC data rather than any new architectural insight. The evaluation leaves several gaps. No baseline comparisons appear, whether to classical shape-from-shading, other networks, or existing lunar DEM products. The abstract gives no information on how the train and test scenes were split, their geographic spread, or the range of solar incidence angles involved. Without those, the numbers could reflect dataset-specific correlations instead of a reliable photometric relationship. Generalization to unseen lighting and surface properties is assumed but not tested, which undercuts the claim that the outputs are dependable for new images. The loss figure itself is hard to interpret without knowing the loss formulation or scale. This is the kind of work that might interest planetary remote-sensing groups looking for quick monocular DEM tools. A reader already working on lunar mapping could pull the setup and try it on their own data, but they would have to add the missing controls themselves. I would send it for peer review because the application is timely and the method is reproducible enough to check, even though the current evidence is too thin to stand on its own.

Referee Report

4 major / 2 minor

Summary. The manuscript introduces LunarDepthNet, a UNet architecture with EfficientNet encoder and custom layers, trained on paired Chandrayaan-2 TMC monocular images and DTMs to generate lunar DEMs. It employs a combined loss function for terrain detail and smoothness, reports 12% loss convergence during validation, and achieves mean nRMSE of 0.437 and MAE of 4.5 m on testing, claiming this enables dependable elevation mapping in regions lacking stereo coverage.

Significance. If the performance generalizes beyond the training distribution, the approach could provide a useful alternative for lunar DEM production where stereo imagery is unavailable, supporting planetary science and mission planning. The application of a standard UNet+EfficientNet backbone to real orbital data with a combined loss to preserve details represents a reasonable engineering effort, though the absence of baselines and robustness tests leaves the practical advance provisional.

major comments (4)

[Abstract] Abstract: the reported nRMSE of 0.437 and MAE of 4.5 m are presented without any baseline comparisons (e.g., traditional shape-from-shading, photometric stereo, or prior DL models), without train/test split details, and without error bars or statistical tests, rendering it impossible to judge whether the numbers represent an advance.
[Methods] Methods/Experiments: no information is given on the geographic distribution of TMC scenes, the solar incidence angle ranges in train versus test splits, or any control for albedo/regolith variations, which directly bears on the central claim that the learned shading-to-elevation mapping will generalize to unseen lunar terrain.
[Results] Results: the statement of 'stable loss convergence of 12%' lacks context on the loss components, normalization, or training/validation curves, and no analysis of overfitting or domain-shift risk is provided despite the purely empirical nature of the evaluation.
[Discussion] Discussion: the manuscript contains no experiments or analysis isolating illumination effects, albedo changes, or surface-property differences across lunar regions, leaving the generalization assumption untested.

minor comments (2)

[Abstract] Abstract: 'nRMSE' is used without definition; it should be expanded on first use.
[Methods] Methods: the 'custom layers' added to the UNet are mentioned but not described in sufficient detail for reproducibility; a network diagram or layer equations would help.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects for strengthening the manuscript's claims on generalization and result interpretation. We address each major comment below and commit to revisions that add the requested details, baselines, and analyses where feasible with the available data.

read point-by-point responses

Referee: [Abstract] the reported nRMSE of 0.437 and MAE of 4.5 m are presented without any baseline comparisons (e.g., traditional shape-from-shading, photometric stereo, or prior DL models), without train/test split details, and without error bars or statistical tests, rendering it impossible to judge whether the numbers represent an advance.

Authors: We agree that baseline comparisons and additional statistical context are necessary to properly evaluate the reported metrics. In the revised manuscript, we will add comparisons against shape-from-shading and other relevant traditional or prior DL approaches (where implementations or results are available for the same TMC data). We will also expand the abstract and results to include train/test split details (e.g., number of scenes and images), error bars on nRMSE and MAE, and basic statistical tests for significance. These additions will be supported by new tables or figures. revision: yes
Referee: [Methods] no information is given on the geographic distribution of TMC scenes, the solar incidence angle ranges in train versus test splits, or any control for albedo/regolith variations, which directly bears on the central claim that the learned shading-to-elevation mapping will generalize to unseen lunar terrain.

Authors: We acknowledge that these details are essential for supporting generalization claims. We will revise the Methods and Dataset sections to explicitly describe the geographic distribution of the Chandrayaan-2 TMC scenes, the solar incidence angle ranges for training versus test splits, and any preprocessing or selection steps used to mitigate albedo and regolith variations. Where possible, we will quantify these factors and discuss their potential impact on the model. revision: yes
Referee: [Results] the statement of 'stable loss convergence of 12%' lacks context on the loss components, normalization, or training/validation curves, and no analysis of overfitting or domain-shift risk is provided despite the purely empirical nature of the evaluation.

Authors: We agree that the loss convergence statement requires more context. In the revised Results section, we will provide the full breakdown of the combined loss function (including weighting of components), normalization details, and plots of training/validation loss curves over epochs. We will also add a dedicated subsection analyzing overfitting risks (e.g., via validation gap metrics) and domain-shift considerations based on the empirical splits. revision: yes
Referee: [Discussion] the manuscript contains no experiments or analysis isolating illumination effects, albedo changes, or surface-property differences across lunar regions, leaving the generalization assumption untested.

Authors: We recognize this as a key gap in validating generalization. We will expand the Discussion to include targeted analysis of illumination effects (e.g., performance stratified by incidence angle), albedo variations, and surface property differences using available metadata from the TMC scenes. This may involve additional subset evaluations or sensitivity visualizations; if full new experiments are needed, we will perform them to the extent supported by the dataset. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML training on paired data yields independent test metrics

full rationale

The paper presents a standard supervised learning pipeline: a UNet+EfficientNet model is trained on paired Chandrayaan-2 TMC images and corresponding DTMs using a combined loss, then evaluated on a held-out test set with nRMSE and MAE. No equations, parameters, or self-citations reduce the reported performance numbers to quantities defined by the fitted weights themselves. The central claim is an empirical generalization result whose validity depends on data distribution and experimental controls, not on any definitional or self-referential reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The performance claim rests on the empirical generalization of a supervised neural network trained on a finite set of image-DTM pairs; no explicit free parameters beyond standard network weights are stated, and no new physical entities are introduced.

free parameters (1)

network weights
All model parameters are fitted during training on the Chandrayaan-2 image-DTM pairs.

axioms (1)

domain assumption Shading and shadow patterns in monocular images contain sufficient information to recover absolute elevation when the network is trained on paired data.
Implicit in the choice of monocular depth estimation for lunar terrain.

pith-pipeline@v0.9.0 · 5538 in / 1185 out tokens · 38432 ms · 2026-05-10T00:47:32.195746+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

and Korzun, A.M., 2024

George, T., Morales, B.E., Munk, M.M., Carson, J.M., Anzalone, E.J., Rubio, M.A., Lawrence, S., Lugo, R.A. and Korzun, A.M., 2024. Safe and Precise Landing at Lunar Sites

work page 2024
[2]

and LROC Science Team, 2012

Burns, K.N., Speyerer, E.J., Robinson, M.S., Tran, T., Rosiek, M.R., Archinal, B.A., Howington-Kraus, E. and LROC Science Team, 2012. Digital elevation models and derived products from LROC NAC stereo observa- tions. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 39, pp.483-488

work page 2012
[3]

and Smith, D.E., 2016

Barker, M.K., Mazarico, E., Neumann, G.A., Zuber, M.T., Haruyama, J. and Smith, D.E., 2016. A new lu- nar digital elevation model from the Lunar Orbiter Laser Altimeter and SELENE Terrain Camera. Icarus, 273, pp.346-355. 8

work page 2016
[4]

Characterization of lunar polar illumination from a power system perspective

Fincannon, J., 2008, January. Characterization of lunar polar illumination from a power system perspective. In 46th AIAA Aerospace Sciences Meeting and Exhibit (p. 447)

work page 2008
[5]

and Saputra, M.R.U., 2025

Madani, A.I., Kuswati, R.A., Lechner, A.M. and Saputra, M.R.U., 2025. Digital Elevation Model Estimation from RGB Satellite Imagery using Generative Deep Learning. arXiv preprint arXiv:2511.21985

work page arXiv 2025
[6]

and Hu, D., 2024

Zhao, Y ., Liu, Y ., Gao, S., Liu, G., Wan, Z. and Hu, D., 2024. Deep Learning-Based Digital Surface Model Reconstruction of ZY-3 Satellite Imagery. Remote Sensing, 16(14), p.2567

work page 2024
[7]

and Zhang, D., 2024

Yang, L., Zhu, Z., Sun, L. and Zhang, D., 2024. Global attention-based DEM: A planet surface digital elevation model-generation method combined with a global attention mechanism. Aerospace, 11(7), p.529

work page 2024
[8]

and Oberst, J., 2022

Chen, H., Hu, X. and Oberst, J., 2022. Pixel-resolution DEM generation for the lunar surface based on a combined deep learning and shape-from-shading (SFS) approach. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 3, pp.511-516

work page 2022
[9]

and Liu, Z., 2022

Liu, Y ., Wang, Y ., Di, K., Peng, M., Wan, W. and Liu, Z., 2022. A generative adversarial network for pixel- scale lunar DEM generation from high-resolution monocular imagery and low-resolution DEM. Remote Sensing, 14(21), p.5420

work page 2022
[10]

and Barriot, J.P., 2023

Zhong, J., Yan, J., Li, M. and Barriot, J.P., 2023. A deep learning-based local feature extraction method for improved image matching and surface reconstruction from Yutu-2 PCAM images on the Moon. ISPRS Journal of Photogrammetry and Remote Sensing, 206, pp.16-29

work page 2023
[11]

and Willner, K., 2025

Chen, H., Oberst, J., Gl ¨aser, P. and Willner, K., 2025. Review on Deep Learning Techniques in Planetary Topo- graphic Modeling. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 48, pp.275-280

work page 2025
[12]

and Oberst, J., 2024

Chen, H., Gl ¨aser, P., Hu, X., Willner, K., Zheng, Y ., Damme, F., Bruzzone, L. and Oberst, J., 2024. ELunarDTM- Net: Efficient reconstruction of high-resolution lunar DEM from single-view orbiter images. IEEE Transactions on Geoscience and Remote Sensing

work page 2024
[13]

and Cui, H., 2024

Tian, P., Yao, M., Xiao, X., Zheng, B., Cao, T., Xi, Y ., Liu, H. and Cui, H., 2024. 3-D semantic terrain reconstruc- tion of monocular close-up images of Martian terrains. IEEE Transactions on Geoscience and Remote Sensing, 62, pp.1-16

work page 2024
[14]

and Oberst, J., 2022

Chen, H., Hu, X., Gl ¨aser, P., Xiao, H., Ye, Z., Zhang, H., Tong, X. and Oberst, J., 2022. CNN-based large area pixel-resolution topography retrieval from single-view LROC NAC images constrained with SLDEM. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15, pp.9398-9416

work page 2022
[15]

and Chen, S., 2025

Chen, T., Wang, Y ., Nan, J., Zhao, C., Wang, B., Xie, B., Liu, W.C., Di, K., Liu, B. and Chen, S., 2025. A Generative Adversarial Network for Pixel-Scale Lunar DEM Generation from Single High-Resolution Image and Low-Resolution DEM Based on Terrain Self-Similarity Constraint. Remote Sensing, 17(17), p.3097

work page 2025
[16]

https://pradan.issdc.gov.in/ch2/ 9

work page

[1] [1]

and Korzun, A.M., 2024

George, T., Morales, B.E., Munk, M.M., Carson, J.M., Anzalone, E.J., Rubio, M.A., Lawrence, S., Lugo, R.A. and Korzun, A.M., 2024. Safe and Precise Landing at Lunar Sites

work page 2024

[2] [2]

and LROC Science Team, 2012

Burns, K.N., Speyerer, E.J., Robinson, M.S., Tran, T., Rosiek, M.R., Archinal, B.A., Howington-Kraus, E. and LROC Science Team, 2012. Digital elevation models and derived products from LROC NAC stereo observa- tions. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 39, pp.483-488

work page 2012

[3] [3]

and Smith, D.E., 2016

Barker, M.K., Mazarico, E., Neumann, G.A., Zuber, M.T., Haruyama, J. and Smith, D.E., 2016. A new lu- nar digital elevation model from the Lunar Orbiter Laser Altimeter and SELENE Terrain Camera. Icarus, 273, pp.346-355. 8

work page 2016

[4] [4]

Characterization of lunar polar illumination from a power system perspective

Fincannon, J., 2008, January. Characterization of lunar polar illumination from a power system perspective. In 46th AIAA Aerospace Sciences Meeting and Exhibit (p. 447)

work page 2008

[5] [5]

and Saputra, M.R.U., 2025

Madani, A.I., Kuswati, R.A., Lechner, A.M. and Saputra, M.R.U., 2025. Digital Elevation Model Estimation from RGB Satellite Imagery using Generative Deep Learning. arXiv preprint arXiv:2511.21985

work page arXiv 2025

[6] [6]

and Hu, D., 2024

Zhao, Y ., Liu, Y ., Gao, S., Liu, G., Wan, Z. and Hu, D., 2024. Deep Learning-Based Digital Surface Model Reconstruction of ZY-3 Satellite Imagery. Remote Sensing, 16(14), p.2567

work page 2024

[7] [7]

and Zhang, D., 2024

Yang, L., Zhu, Z., Sun, L. and Zhang, D., 2024. Global attention-based DEM: A planet surface digital elevation model-generation method combined with a global attention mechanism. Aerospace, 11(7), p.529

work page 2024

[8] [8]

and Oberst, J., 2022

Chen, H., Hu, X. and Oberst, J., 2022. Pixel-resolution DEM generation for the lunar surface based on a combined deep learning and shape-from-shading (SFS) approach. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 3, pp.511-516

work page 2022

[9] [9]

and Liu, Z., 2022

Liu, Y ., Wang, Y ., Di, K., Peng, M., Wan, W. and Liu, Z., 2022. A generative adversarial network for pixel- scale lunar DEM generation from high-resolution monocular imagery and low-resolution DEM. Remote Sensing, 14(21), p.5420

work page 2022

[10] [10]

and Barriot, J.P., 2023

Zhong, J., Yan, J., Li, M. and Barriot, J.P., 2023. A deep learning-based local feature extraction method for improved image matching and surface reconstruction from Yutu-2 PCAM images on the Moon. ISPRS Journal of Photogrammetry and Remote Sensing, 206, pp.16-29

work page 2023

[11] [11]

and Willner, K., 2025

Chen, H., Oberst, J., Gl ¨aser, P. and Willner, K., 2025. Review on Deep Learning Techniques in Planetary Topo- graphic Modeling. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 48, pp.275-280

work page 2025

[12] [12]

and Oberst, J., 2024

Chen, H., Gl ¨aser, P., Hu, X., Willner, K., Zheng, Y ., Damme, F., Bruzzone, L. and Oberst, J., 2024. ELunarDTM- Net: Efficient reconstruction of high-resolution lunar DEM from single-view orbiter images. IEEE Transactions on Geoscience and Remote Sensing

work page 2024

[13] [13]

and Cui, H., 2024

Tian, P., Yao, M., Xiao, X., Zheng, B., Cao, T., Xi, Y ., Liu, H. and Cui, H., 2024. 3-D semantic terrain reconstruc- tion of monocular close-up images of Martian terrains. IEEE Transactions on Geoscience and Remote Sensing, 62, pp.1-16

work page 2024

[14] [14]

and Oberst, J., 2022

Chen, H., Hu, X., Gl ¨aser, P., Xiao, H., Ye, Z., Zhang, H., Tong, X. and Oberst, J., 2022. CNN-based large area pixel-resolution topography retrieval from single-view LROC NAC images constrained with SLDEM. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15, pp.9398-9416

work page 2022

[15] [15]

and Chen, S., 2025

Chen, T., Wang, Y ., Nan, J., Zhao, C., Wang, B., Xie, B., Liu, W.C., Di, K., Liu, B. and Chen, S., 2025. A Generative Adversarial Network for Pixel-Scale Lunar DEM Generation from Single High-Resolution Image and Low-Resolution DEM Based on Terrain Self-Similarity Constraint. Remote Sensing, 17(17), p.3097

work page 2025

[16] [16]

https://pradan.issdc.gov.in/ch2/ 9

work page