LunarPhoto: Lunar Imaging Workflows

• LunarPhoto is a domain defining methods for acquiring and interpreting lunar imagery, emphasizing calibration, photometric benchmarks, and restoration.
• It integrates practical workflows—from consumer webcam mosaics to sophisticated lunar telescopic imaging—for measuring crater diameters and reflectance parameters.
• The methodology converts raw lunar images into actionable data, enabling quantitative assessments in photometry, geometry, and hazard monitoring.

LunarPhoto is a term used across several strands of lunar imaging research to denote, depending on context, a practical workflow for acquiring and analyzing lunar photographs, a photometric benchmark built on real lunar topography, and a broader family of techniques for restoring, calibrating, simulating, and interpreting lunar images. Across these usages, the common objective is to convert lunar imagery into quantitative geometric or photometric products such as crater diameters, BRDF parameter maps, stereo geometry, or event detections (Sato, 2013, Grethen et al., 1 Apr 2026).

1. Scope and meanings

In the cited literature, “LunarPhoto” appears in more than one technical sense. The most formal current usage is as the photometric half of the MoonAnything benchmark, where LunarPhoto is paired with LunarGeo to provide appearance supervision on the same Tycho crater DEM (Grethen et al., 1 Apr 2026). Earlier work instead centers on a practical lunar-imaging workflow built around consumer webcams, telescopes, and manual mosaicking (Sato, 2013). In the supplied literature, the label is also used to frame restoration, simulation, reconstruction, and specialized scientific analysis of lunar photographs. This suggests a broader editorial usage: LunarPhoto as a domain of lunar-image acquisition and inference rather than a single instrument or software package.

Usage	Core task	Representative source
Webcam-CCD workflow	Lunar mosaics and crater measurement	(Sato, 2013)
Photometric benchmark	BRDF supervision and multi-illumination rendering	(Grethen et al., 1 Apr 2026)
Broader methodological usage	Restoration, simulation, recognition, and event analysis	(Roy et al., 2019)

The unifying technical themes are calibration, image formation, illumination geometry, and the conversion of pixel data into physically interpretable quantities. In one setting this means plate scale and crater diameter; in another it means Hapke-family reflectance parameters, shadow visibility terms, or dense stereo geometry.

2. Practical acquisition and metrology

A foundational LunarPhoto workflow is the webcam-CCD laboratory described in “Imaging the Moon II,” where a retail webcam is physically adapted by removing the factory lens and mounting the CCD board in a short tube so the sensor sits at the telescope focal plane (Sato, 2013). The procedure uses modest hardware: a small telescope, a common webcam, and a laptop or netbook running stock webcam software. At KPU, 8-inch instruments were used, although the paper states that they are larger than necessary for lunar work.

The acquisition protocol is deliberately structured to mirror observing practice. Students typically work in groups of four with 50-minute shifts: roughly 30 minutes for orientation, naked-eye and eyepiece familiarization, grid planning, and observing logs, followed by roughly 20 minutes of image capture. The workflow emphasizes overlap between frames, observing discipline, and logistical uncertainty, including weather stand-by for the first clear night in an evening-Moon window. Default auto-exposure is a known failure mode: if blank sky occupies too much of the frame, the Moon becomes overexposed on a gray background, so the practical mitigation is to keep the Moon occupying a large fraction of the frame (Sato, 2013).

The analysis stage assembles still frames into a full lunar mosaic using common office software such as MS PowerPoint rather than specialized astronomical packages. Quantitative metrology then proceeds from plate-scale calibration. The feature-based calibration is

$$s_{\mathrm{arcsec/px}}=\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}},$$

while an instrument-based alternative is

$s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$

A crater diameter measured as $N_{\mathrm{px}}$ pixels is converted through angular size and Earth–Moon distance, with the paper giving the combined feature-based expression

$$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$

Uncertainty propagation is explicit:

$$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$

The baseline workflow does not require frame stacking, although the paper notes the optional refinement $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$ for $N$ independent frames (Sato, 2013).

3. Lunar-surface ultraviolet imaging and photometric calibration

A distinct LunarPhoto lineage concerns optical imaging from the lunar surface itself. The Lunar-based Ultraviolet Telescope, LUT, is described as the first robotic astronomical telescope working on the lunar surface and reported stable photometric calibration during its first six months of operation (Wang et al., 2014). LUT is a Ritchey–Chrétien system with a 150 mm primary, 563 mm focal length, a $1.36\times1.36$ deg² field, and a pixel scale of approximately $4.76$ arcsec pixel$^{-1}$. Calibration is performed directly in the LUT AB system using measured throughput $s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$0 and standard-star spectral energy distributions. The reported magnitude zero points are

$s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$1

for set A and

$s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$2

for set B, with no significant time variation across the first six months (Wang et al., 2014). For reduced data, the counts-to-magnitude relation is

$s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$3

LUCI, by contrast, is a design study for a compact all-spherical near-UV imager intended to exploit the lunar surface as a stable, atmosphere-free platform (Mathew et al., 2016). Its scientific band is 200–320 nm, set by a solar-blind UV filter. The instrument has aperture $s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$4 mm, focal length $s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$5 mm, field of view $s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$6, and pixel size $s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$7m, giving the quoted scale

$s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$8

LUCI is a bright-source instrument: the paper gives a bright limit of approximately AB $s_{\mathrm{arcsec/px}}=\frac{206265\,p}{f}.$9 for the minimum exposure of $N_{\mathrm{px}}$0 s and a limiting magnitude of approximately AB $N_{\mathrm{px}}$1 for representative exposures. Its total response is written through the effective area

$N_{\mathrm{px}}$2

and the reported peak effective area is of order $N_{\mathrm{px}}$3 cm². These two cases illustrate complementary LunarPhoto architectures: calibrated astronomical photometry from an operating lunar telescope, and a low-mass NUV imaging concept tuned to lunar lander constraints (Wang et al., 2014, Mathew et al., 2016).

4. Photometric theory and surface-property inference

A recurrent question in LunarPhoto is how brightness variations over the lunar disk should be interpreted. In “The Sun and the Moon a Riddle in the Sky,” Lachish argues that true full-phase photographs of the Moon and other bodies are nearly uniform across the disk and therefore non-Lambertian (Lachish, 2018). For a Lambertian reflector under full-phase geometry, the radiance would scale as

$N_{\mathrm{px}}$4

which predicts strong center-to-limb darkening. The paper instead proposes a single-scattering backscatter argument in which the angular factors cancel, yielding

$N_{\mathrm{px}}$5

This interpretation is presented as a fundamental-principles explanation of the near-uniformity of full-phase disks. It coexists with more conventional lunar photometric formalisms, including Lommel–Seeliger and Hapke-type models, which remain central to modern reflectance modeling.

Multi-band photometry is also used as a remote-sensing tool. During the total lunar eclipse of 2011-06-15, three-color photometry at 503, 677, and 867 nm mapped the umbra and inferred aerosol extinction along refracted terrestrial limb paths (Ugolnikov et al., 2011). The core observable was relative brightness,

$N_{\mathrm{px}}$6

with aerosol retrieval written as

$N_{\mathrm{px}}$7

The study linked a dark anomaly in the eastern umbra, strongest at 867 nm, to enhanced aerosol loading over eastern China, while western-limb profiles near $N_{\mathrm{px}}$8 matched gaseous-only theory more closely (Ugolnikov et al., 2011).

At much smaller scales, LROC NAC photometry has been used to isolate the photometric behavior of resolved boulder fields. Marshal et al. define the normalized logarithmic phase ratio difference,

$N_{\mathrm{px}}$9

where $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$0 denotes a rock-free reference and $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$1 a rock-rich field (Marshal et al., 2023). Positive NLPRD indicates that the rock-rich field darkens faster with phase angle and is photometrically rougher. A central result is that rock-rich surfaces are not necessarily photometrically rougher than rock-free areas; the observed diversity instead points to sub-mm scale rock roughness and possibly variable rock single-scattering albedo, with spatial clustering that may reflect ejecta asymmetry (Marshal et al., 2023).

5. Restoration, simulation, and benchmark datasets

LunarPhoto increasingly includes learned restoration of degraded imagery. A U-Net-based inpainting study addressed grayscale Kaguya Multiband Imager mosaics in which vertical black stripes occupy less than 2% of the image (Roy et al., 2019). Training data consisted of 340 clean crater images and 53 corrupted crater images, resized to $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$2 pixels; realistic stripe masks were extracted from corrupted images and randomly overlaid on clean samples to generate 10,000 training pairs and 5,000 validation/testing pairs. The network was trained in PyTorch on an NVIDIA TITAN-V GPU for approximately 12 hours using L2 loss. Reported PSNR improvements on example patches were substantial, including $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$3 dB to $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$4 dB and $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$5 dB to $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$6 dB, with the restored pixel values described as natural relative to the surrounding terrain (Roy et al., 2019).

The most explicit benchmark use of the name is the LunarPhoto sub-dataset of MoonAnything (Grethen et al., 1 Apr 2026). LunarPhoto covers the Tycho crater region, approximately $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$7 km $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$8 $$D_{\mathrm{crater}}=N_{\mathrm{px}}\times\left(\frac{\theta_{\mathrm{Moon,arcsec}}}{N_{\mathrm{Moon}}}\right)\times\left(\frac{\pi}{648000}\right)\times\Delta_{\mathrm{Moon}}.$$9 km in the master DEM, and contains 84,000 samples split into 67,000 train, 8,500 validation, and 8,500 test. Each sample is a $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$0 DEM crop at $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$1 m/px with pixel-aligned assets including dem.tif, real_image.tif, depth.tif, normal.tif, brdf_map.tif, and metadata.json, together with nine rendered images under SPICE-derived solar configurations. The rendering pipeline uses SurRender and provides both a canonical Hapke reflectance model and a learned SVBRDF, allowing supervision of reflectance estimation, inverse rendering, relighting, and illumination-robust recognition (Grethen et al., 1 Apr 2026). The image formation model is written as

$$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$2

with $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$3 the visibility term.

Open-source synthetic-image generation remains heterogeneous. An evaluation of ABRAM, CORTO, Blender, QGIS 3D, and an in-house Python renderer concludes that ABRAM is strongest for explicit lunar photometry, Blender and CORTO yield the most convincing shadows and visual realism, and QGIS is most effective for DEM preparation, alignment, and tiling (Singla et al., 24 Apr 2026). This suggests a division of labor between physically interpretable reflectance simulation and photorealistic rendering, especially when large DEMs, sensor models, and low-Sun geometries must be combined.

6. Geometry, recognition, and 3D reconstruction

Geometric inference from lunar photographs is difficult because the regolith is nearly monochromatic, repetitive, and often observed under extreme illumination. For single-image crater recognition, Christian et al. give a mathematically rigorous treatment of the lost-in-space problem by modeling crater rims as conics under perspective projection (Christian et al., 2020). A central result is that there are no nonconstant rational invariants for the projection of arbitrary conics in $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$4 from a single image, but complete invariant sets do exist for structured configurations: $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$5 algebraically independent invariants for $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$6 coplanar conics and $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$7 for $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$8 conics on a common nondegenerate quadric. This enables searchable descriptor indices and pose recovery from crater-rim observations rather than heuristic center-point matching (Christian et al., 2020).

Stereo and dense reconstruction require domain adaptation. LunarStereo, also referred to in the paper title as StereoLunar, provides over 50,000 stereo pairs rendered with ESA/Airbus’s SurRender over LOLA South Pole DEM data at 5 m/px, spanning altitudes from 3.5 km to 30.5 km, a $$\left(\frac{\sigma_D}{D}\right)^2= \left(\frac{\sigma_N}{N}\right)^2+ \left(\frac{\sigma_s}{s}\right)^2+ \left(\frac{\sigma_\Delta}{\Delta}\right)^2.$$9 field of view, three trajectory families, and three illumination instances (Grethen et al., 20 Oct 2025). Fine-tuning MASt3R on approximately 31,000 lunar pairs for 25 epochs with AdamW at learning rate $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$0 and batch size 2 improved both pose and surface metrics. In the Dynamic setting, for example, relative translation accuracy reached $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$1, $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$2, $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$3, and $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$4, while Chamfer error dropped to 111 m / 0.41% (Grethen et al., 20 Oct 2025). The paper’s practical point is that terrestrial stereo priors transfer poorly unless the model is retrained on lunar reflectance and viewpoint regimes.

At rover and analog scales, the POLAR Traverse Dataset provides 3,960 stereo pairs across 24 traverses under simulated south-polar lighting, with an approximately 0.40 m stereo baseline, 11 positions per traverse, and 15 exposure settings from 1 to 500 ms (Hansen et al., 2024). Light sources at approximately 2–3° above the terrain reproduce long shadows and high dynamic range. Geometric calibration used a $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$5 checkerboard with 60 mm squares and achieved an average reprojection error of 0.111 pixels. The dataset is designed for stereo or monocular visual odometry, multi-view stereo, and exposure-robust perception under lighting conditions close to those expected near the lunar south pole (Hansen et al., 2024).

7. Event detection, hazards, and specialized targets

LunarPhoto techniques are increasingly used for scientifically and operationally specialized targets. In permanently shadowed regions, laboratory simulations showed that passive UV/VIS ratio imaging can separate water ice from regolith because ice is darker than regolith near Lyman-$\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$6 yet brighter in the visible (Godin et al., 2020). With co-registered UV and VIS images, the ratio

$\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$7

improved the ice/regolith discrimination SNR by 36% relative to single-band imaging in the top-down geometry. The experiment used a cryo-vacuum chamber at 95 K and approximately $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$8 Torr, a Resonance Ltd. Krypton line-source VUV lamp, and JSC-1A regolith simulant (Godin et al., 2020).

Impact monitoring is another major application. An end-to-end simulator for lunar far-side impact flashes viewed from Earth–Moon L2 models flash temporal radiation, lunar background emission, telescope optics, and CCD/CMOS detection in Python 3.9 (Song et al., 2024). In its example instrument, the field of view is $\mathrm{SNR}_N\approx\sqrt{N}\,\mathrm{SNR}_1$9, the Moon subtends approximately $N$0 at L2, and predicted peak SNRs for a representative Flash 2 reach 383 in R and 525 in I at phase 0.5. The paper also estimates that a minimum aperture of approximately 537 mm is required for marginal detection of Flash 1 with SNR $N$1 from L2 (Song et al., 2024). Post-event LRO imaging can then close the loop between flash and crater: the 2013-09-11 event produced a crater of $N$2 m diameter and $N$3 m depth, ejecta over more than 2 km with area approximately $N$4 m², and a mean 16.54% increase in spectral slope between 321 nm and 643 nm in the central ejecta region, which the paper identifies as the first reported detection of impact-induced color changes on the Moon (Rizos et al., 18 Nov 2025).

Historical imagery can also be reinterpreted quantitatively. Apollo landing videos, analyzed with modern photogrammetry, yielded dust-ejection angles of 1–3 degrees, lofted particle densities of $N$5–$N$6 particles/m³, and evidence for ejection of 10–15 cm objects (Immer et al., 2021). In that workflow, camera pose is recovered from crater and hardware geometry, while brightness changes between clear and dusty frames are linked to particle density by Beer–Lambert attenuation. This is a different use of LunarPhoto from photometric mapping or reflectance learning, but it relies on the same principle: extracting physically meaningful state variables from calibrated lunar imagery (Immer et al., 2021).

LunarPhoto therefore spans a continuum from accessible lunar mosaicking to BRDF-supervised benchmarks, from ultraviolet photometry on the lunar surface to event forensics and hazard monitoring. The common technical denominator is not the sensor class or wavelength range, but the insistence that lunar photographs be treated as measurable radiometric and geometric data rather than as purely illustrative images.