Skylight: Diverse Technical Perspectives

• Skylight is a polysemous term that defines a diffuse natural light field and serves as a cue in atmospheric optics, astronomical calibration, planetary pit access, and as a namesake for advanced systems.
• Applications include polarization-based navigation, spectral calibration in astronomy, and sensor fusion with GNSS/inertial data for improved orientation and attitude estimation.
• Modern implementations range from maritime monitoring platforms to relativistic radiative-transfer codes and 3D photonic tensor-core architectures, showcasing practical integration of theory and technology.

Skylight is a polysemous technical term whose meaning depends strongly on disciplinary context. In recent research literature it denotes natural diffuse atmospheric illumination and its polarization field, atmospheric sky emission in spectroscopy, light collected from the sky by astronomical instruments, collapsed-ceiling entrances to subsurface pits on the Moon and Mars, a fine-grained vehicle attribute, and several named systems including a maritime monitoring platform, a general-relativistic radiative-transfer code, and a 3D photonic tensor-core architecture (Soga et al., 2018, Wu et al., 2021, Rombach et al., 18 Jun 2026, Morad et al., 2019, Wang et al., 2019, Beukema et al., 2023, Pelle et al., 2022, Zhang et al., 22 Feb 2026).

1. Terminological range

In current arXiv usage, “skylight” appears in at least four distinct technical senses. First, it refers to the naturally illuminated sky and, more specifically, to the partially polarized light produced by atmospheric scattering. Second, in astronomy and spectroscopy it denotes either useful celestial light entering an instrument or unwanted atmospheric sky background superposed on spectra. Third, in planetary robotics it denotes a collapsed-ceiling opening into a pit or lava tube. Fourth, as a proper name, “Skylight” or “SKYLIGHT” identifies deployed or proposed systems in maritime monitoring, relativistic radiative transfer, and photonic AI acceleration (Galindo et al., 2018, Wu et al., 2021, Morad et al., 2019, Beukema et al., 2023, Pelle et al., 2022, Zhang et al., 22 Feb 2026).

Sense	Domain	Representative use
Polarized skylight	optics, navigation	Rayleigh-scattered atmospheric light
Sky light / skylight background	astronomy, spectroscopy	wavelength recalibration and sky subtraction
Skylight entrance	planetary exploration	collapsed-ceiling pit opening
Skylight / SKYLIGHT as proper name	AI, astrophysics, conservation	platform, code, tensor-core architecture

A plausible unifying interpretation is that the term preserves a relation to the sky either as illumination source, observational domain, or access aperture. The differences arise from whether the sky is being measured, used as a cue, treated as contamination, or abstracted into a system name.

2. Atmospheric optics and polarization

In atmospheric optics, skylight is the diffuse light field generated when sunlight is scattered by the atmosphere. The standard physical explanation in the cited literature is Rayleigh scattering, with a wavelength dependence approximately proportional to $1/\lambda^4$, which accounts for the blue appearance of the sky and for the stronger scattering of shorter wavelengths (Soga et al., 2018). The same paper gives the Rayleigh-scattering intensity as

$$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$

linking spectral dependence and angular dependence in a single expression (Soga et al., 2018).

The polarization geometry of skylight has long been a subject of both physical and observational interest. Historical observations summarized in the literature report that the plane of polarization is tangential to circles centered on the Sun, that the effect is strongest at $90^\circ$ from the Sun, and that there are directions where polarization vanishes: the Arago, Babinet, and Brewster neutral points (Galindo et al., 2018). A historically significant but unsuccessful attempt to exploit this geometry was William Kingdon Clifford’s 1870 eclipse expedition in Sicily, where he sought a change in the plane of polarization of skylight as a possible empirical test of gravitation-as-curvature; the observation was clouded over and revealed no anomaly beyond the known polarization pattern (Galindo et al., 2018).

Modern treatments emphasize both measurement and visualization. A simple demonstration uses partially polarized blue skylight, a birefringent plastic object, and a polarizer; the birefringent element changes polarization differently across position and wavelength, and the analyzer converts those differences into colored patterns (Soga et al., 2018). The same source notes that skylight polarization is not perfect, citing a maximum of about 85% polarization, which explains why brightness-only detection with a rotating polarizer can be subtle (Soga et al., 2018).

3. Polarized-skylight navigation, datasets, and inference

Bioinspired navigation research treats polarized skylight as a compass cue analogous to the sensing used by insects such as desert ants and locusts. In this literature, polarization imagers usually measure four analyzer orientations, $I_0$, $I_{45}$, $I_{90}$, and $I_{135}$, from which the Stokes quantities are formed as

$$S_0=I_0+I_{45}+I_{90}+I_{135}, \qquad S_1=I_0-I_{90}, \qquad S_2=I_{45}-I_{135}, \qquad S_3=0.$$

The degree of polarization and angle of polarization are then computed in-network or in preprocessing as

$$\mathrm{DOP}=\frac{\sqrt{S_1^2+S_2^2}}{S_0}, \qquad \mathrm{AOP}=-\arctan\frac{S_2}{S_1}$$

in the formulation used for neural orientation determination (Liang et al., 2021).

The principal open benchmark in this area is the Polarized Skylight Navigation Simulation dataset, which was introduced as the first open polarized-skylight navigation simulation dataset and contains about 138,000 images (Liang et al., 2020). It combines a Sun position model, a Berry sky model, and a Hosek sky model to generate light intensity (LI), degree of polarization (DOP), angle of polarization (AOP), and raw black-and-white intensity images at $0^\circ$, $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$0, $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$1, and $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$2 analyzer angles (Liang et al., 2020). This emphasis on raw analyzer-direction images is methodologically important because real polarization imagers do not observe DOP or AOP directly.

Neural methods built on this formulation can estimate orientation accurately under the dataset’s simulated clear-sky conditions. One artificial neural network with specific dilated convolution, in-network DOP/AOP extraction, and exponential-function orientation encoding reported MAE $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$3, ME $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$4, and RMSE $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$5 on the public PSNS dataset, outperforming direct angle regression and alternative encodings (Liang et al., 2021). That same work also identified a random $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$6 error mode when solar altitude is very small, which it interprets as a physical ambiguity in the sky polarization pattern rather than only a learning failure (Liang et al., 2021).

The limits of the cue are equally prominent in the literature. A richer Berry + Hosek sky model was shown in simulation to contain 3D attitude information, but the accuracy of three-Euler-angle estimation drops significantly under measurement noise or model error, and field experiments made 3D attitude estimation very difficult in practice (Liang et al., 2020). This limitation motivates integrated systems. A PSNS/GNSS/SINS architecture therefore treats the skylight-derived observation as a bi-direction solar vector and fuses it with inertial and GNSS information inside a Kalman filter, implemented on a DSP + FPGA dual-core platform (Liang et al., 2020). Sensor pose is also critical: tilt studies showed that Zenith, SM-ASM, Symmetry, and Least-square approaches have highly consistent tilt-induced error behavior, with errors governed by tilt magnitude, solar altitude, and the relative position between the Sun and the sensor (Liang et al., 2020).

4. Astronomical and spectroscopic instrumentation

In astronomy, skylight appears in two nearly opposite operational roles: as desired astronomical signal collected from the sky and as atmospheric background that must be modeled and removed. MOSAIC for the Extremely Large Telescope illustrates the first use. Its roughly 300 robotic positioners pick off skylight from the ELT focal surface and relay it to visible and near-infrared spectrographs spanning roughly 390–1800 nm (Rombach et al., 18 Jun 2026). The focal-plane concept is modular and robotic, with each positioner built from a POS SCARA patrol mechanism and a POS ADC local atmospheric-dispersion corrector (Rombach et al., 18 Jun 2026).

The MOSAIC optical train departs from conventional single-fiber pickoff. Light from the ELT focal surface is relayed through four mirrors, M1–M4, then reimaged onto one of two fixed fiber bundles located 600 mm behind the ELT focal plane (Rombach et al., 18 Jun 2026). This design is imposed by three ELT-scale constraints: the beam cannot be focused into a single fiber at the focal plane, local telecentricity must point toward the pupil center 37.868 m away, and a full-instrument ADC covering the more than 2 m focal plane was judged too risky, too large, and too expensive (Rombach et al., 18 Jun 2026). The result is a positioner architecture that must patrol a donut-like region, maintain telecentric pointing, and perform local atmospheric-dispersion correction on each robot.

The opposite usage appears in the LAMOST MRS-N pipeline, where skylight is the terrestrial atmospheric emission contaminating nebular spectra. The workflow explicitly includes fitting sky light emission lines, wavelength recalibration from those lines, and subtracting skylight before nebular parameter measurement (Wu et al., 2021). In the red band from 6300–6800 Å at resolving power $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$7, seven skylines at 6287, 6300, 6363, 6498, 6533, 6544, and 6553 Å are fitted with Gaussian functions and used for a quadratic recalibration function $$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$8 (Wu et al., 2021). For sky subtraction around H$$I(\theta)=\frac{8\pi^4 a^6}{R^2\lambda^4}\left|\frac{m^2-1}{m^2+2}\right|^2(1+\cos^2\theta),$$9, the pipeline uses the empirical relation

$90^\circ$0

constructs the geocoronal H$90^\circ$1 component from the OH 6554 Å line, and reports that about 90% of the sky light mixed in nebular H$90^\circ$2 emission lines can be reduced (Wu et al., 2021).

5. Skylight as a maritime monitoring platform

“Skylight” is also the name of a deployed real-time maritime intelligence and ocean-conservation platform built around satellite imagery and machine learning (Beukema et al., 2023). Its purpose is to detect and monitor vessel activity associated with illegal, unreported, and unregulated fishing and other suspicious maritime behavior. The platform ingests near-real-time satellite data, runs specialized computer-vision models, correlates detections with Automatic Identification System (AIS) broadcasts, and delivers results through a GUI and API (Beukema et al., 2023). It is provided free of charge worldwide and, at the time reported, served over 308 organizations in more than 60 countries (Beukema et al., 2023).

A central architectural claim is that no single sensor is sufficient for global maritime monitoring. Skylight therefore combines specialized services for VIIRS nighttime lights, Sentinel-1 SAR, and Sentinel-2 optical imagery, with AIS/GPS correlation as an auxiliary signal (Beukema et al., 2023). VIIRS offers approximately 2.5 hours latency and about two observations per night but only 750 m spatial resolution, so its detector is deliberately hybrid: unsupervised source extraction, deterministic removal of lightning, gas flares, moonlit clouds, auroras, South Atlantic Anomaly artifacts, scanline false positives, and edge noise, followed by a small 2D CNN using nanowatts, land-water masks, moonlight, and clouds as inputs (Beukema et al., 2023). Sentinel-1 uses a Faster R-CNN with a customized 13-layer fully convolutional backbone plus FPN and aligned historical images to suppress persistent structures (Beukema et al., 2023). Sentinel-2 uses Faster R-CNN with a Swin Transformer backbone and cloud suppression via s2cloudless; the paper reports that pretraining on SatlasPretrain improved performance (Beukema et al., 2023).

Operationally, the platform is organized as a streaming pipeline from satellite acquisition to user-facing output. AIS correlation is formulated as a minimum-weight bipartite matching problem using haversine distance and the Jonker-Volgenant algorithm (Beukema et al., 2023). Reported service characteristics as of 10/2023 were latency $90^\circ$3 hrs, revisit $90^\circ$4, and vessel count 145,063 for VIIRS; latency 5 hrs, revisit 14 days, and vessel count 182,234 for Sentinel-1; and latency 5 hrs, revisit 5 days, and vessel count 430,467 for Sentinel-2 (Beukema et al., 2023). Research-stage model evaluation uses held-out F1, with Sentinel-1 improving from 70.1% to 82.7% relative to a previous xView3-submitted version and Sentinel-2 reaching F1 $90^\circ$5 (Beukema et al., 2023).

6. Skylight and SKYLIGHT as computational infrastructures

In relativistic astrophysics, Skylight is a numerical code for general-relativistic ray tracing and radiative transfer in arbitrary asymptotically flat spacetimes and coordinate systems (Pelle et al., 2022). It computes images, spectra, light curves, phase-resolved spectra, sky maps, and animations from models of compact objects as seen by distant observers (Pelle et al., 2022). The code implements both emitter-to-observer Monte Carlo radiative transfer and observer-to-emitter backward ray tracing, integrating the first-order geodesic system

$90^\circ$6

and the covariant transfer equation in Lorentz-invariant form (Pelle et al., 2022). Validation includes conservation of Kerr constants of motion, agreement with a semianalytic Schwarzschild deflection benchmark, and astrophysical tests involving thin accretion disks, orbiting hot spots, and neutron-star hot-spot pulse profiles (Pelle et al., 2022).

The code has subsequently been used to model observables in boson-star spacetimes (Rosa et al., 2024). In that application, Skylight ray-traces thermal emission from a Novikov-Thorne thin disk and relativistically broadened line emission from a lamppost corona around quartic and solitonic boson stars (Rosa et al., 2024). The study found that quartic-potential boson stars differ strongly from Schwarzschild black holes because stable circular orbits exist at all radii, whereas compact solitonic boson stars can mimic black holes closely; in particular, SBS3 and Schwarzschild share $90^\circ$7 at $90^\circ$8 in the reported disk-temperature profiles (Rosa et al., 2024).

In photonic computing, SKYLIGHT denotes a scalable hundred-channel 3D photonic in-memory tensor-core architecture for real-time AI inference (Zhang et al., 22 Feb 2026). The architecture combines a low-loss 3D Si/SiN crossbar topology, a thermally robust non-MRR-based WDM component implemented as a Bragg grating-assisted wavelength-selective coupler, hierarchical signal accumulation using a multi-port Ge photodetector, and optically programmed non-volatile PCM weights (Zhang et al., 22 Feb 2026). It is explicitly designed to overcome scaling barriers associated with planar routing loss, resonant thermal sensitivity, accumulation bottlenecks, and electrical PCM programming complexity (Zhang et al., 22 Feb 2026).

The reported system-level results are unusually concrete. A single $90^\circ$9 SKYLIGHT core is described as feasible within a single reticle and as delivering 342.1 TOPS at 23.7 TOPS/W, enabling ResNet-50 inference at 1212 FPS with 27 mJ per image and achieving 84.17 FPS/W end-to-end, which the paper states is $I_0$0 higher than an NVIDIA RTX PRO 6000 Blackwell GPU under the same workload (Zhang et al., 22 Feb 2026). The architecture also supports in-situ weight updates and label-free, layer-local learning such as forward-forward local updates, extending its intended scope beyond inference alone (Zhang et al., 22 Feb 2026).

7. Other technical uses: planetary openings, reflected skylight, and fine-grained attributes

In planetary robotics, a skylight is a naturally formed collapsed-ceiling entrance into a lunar or Martian pit or lava tube (Morad et al., 2019). Such openings are scientifically attractive because they can provide access to environments shielded from solar UV and ionizing radiation, buffered from extreme diurnal temperature swings, protected from micrometeorite impacts, potentially rich in water ice on the Moon, and potentially preservative environments for signs of past life on Mars (Morad et al., 2019). The SPEER concept addresses these sites by launching disposable spherical microbots through skylights into pits. The paper gives a 1 kg bot buildable from commercial parts for under 500 USD, with 769 g available for science payload, and reports a Mare Tranquillitatis feasibility study using a pit depth of about 80 m, a standoff distance of 5 m, a $I_0$1 launch angle, required launch speed $I_0$2 m/s, and spring compression 4.49 cm for $I_0$3 N/m (Morad et al., 2019).

In field polarimetry, reflected skylight is exploited as a diagnostic signal. A UAV-deployable Polarimetric Imaging-based Mirror Soiling method uses polarization images of reflected skylight/sunlight to infer heliostat soiling in concentrated solar power plants (Tian et al., 3 Jan 2025). The physical basis is that clean mirrors preserve polarization better than soiled ones, while particle scattering reduces the Degree of Linear Polarization. The field workflow pre-calculates the sky polarization pattern, chooses high-DoLP viewing geometry, captures polarization images with a compact camera, and maps DoLP to relative reflectance and soiling level (Tian et al., 3 Jan 2025). Reported errors are 1.41%–2.77% for controlled sample measurements and under 3.1% MAE overall in field deployment (Tian et al., 3 Jan 2025).

In aerial vehicle re-identification, “skylight” is yet another specialized term, referring not to the sky but to a vehicle roof opening annotated as a binary attribute (Wang et al., 2019). The VRAI dataset contains 137,613 images of 13,022 vehicle instances captured by two DJI Phantom 4 UAVs at 11 location pairs and altitudes from 15 m to 80 m (Wang et al., 2019). Skylight is one of four binary attributes, appears among the manually annotated discriminative parts, and achieved 90.16% attribute-classification accuracy in the reported experiments (Wang et al., 2019). In the final system, weighted pooling over detected discriminative regions improved the model from mAP 78.31% and CMC-1 80.05% for average feature aggregation to mAP 78.63% and CMC-1 80.30% for weighted features, supporting the claim that skylight can function as a high-value fine-grained identity cue in aerial imagery (Wang et al., 2019).

Across these usages, the term does not denote a single object class. It instead indexes a family of technically specific concepts tied to the sky as illumination source, observational background, geometric opening, or naming metaphor.