The Color of the Zenith Sky Offers a Simple New Way to Tell Clouds from Clear Air

Every day, at ground stations scattered from Tokyo’s polluted outskirts to the pristine ice of Antarctica and the high Arctic, a humble instrument called a sky radiometer stares straight up and measures how bright the sky is at different wavelengths of light. A new study published in Atmospheric Measurement Techniques shows that this upward glance contains a surprisingly powerful secret: the way the zenith sky’s brightness changes with color provides a simple, physically interpretable fingerprint that reliably separates clear skies from cloudy ones. The finding, led by Pradeep Khatri of Soka University and colleagues across Japan, could sharpen the quality of the long-term aerosol and cloud records that underpin climate science worldwide.

Why does sky-state detection matter so much? Ground-based remote sensing networks such as SKYNET retrieve aerosol properties, including aerosol optical thickness, single scattering albedo, and particle size distributions, from measurements of direct solar irradiance and diffuse sky radiance. These retrievals assume the observer is looking through cloud-free air, because clouds violently perturb the radiative field and can inject large biases into aerosol products. Similarly, retrievals of water vapor and ozone depend on uncontaminated clear-sky measurements. If a thin cirrus veil or a dust-laden but genuinely cloud-free sky is misclassified, the error propagates into satellite validation studies, model evaluations, and the decadal climatologies used to quantify aerosol-cloud interactions and climate forcing.

Existing cloud-screening methods rely heavily on statistics: temporal variability of surface irradiance, wavelength-dependent behavior of aerosols, and quality-control tests on retrieved optical thickness. These tools work well much of the time, but they stumble in difficult conditions. Subvisual cirrus can masquerade as clear air, while heavy aerosol loading from dust storms, biomass burning, or urban pollution can inflate turbidity enough that statistical tests wrongly flag a cloud-free sky as cloudy. The new study takes a different route by asking a more fundamental question: what does the spectrum of light scattered down from the zenith actually look like in each situation?

The physics is elegant. Under a clear sky, the light reaching the radiometer from the zenith is dominated by Rayleigh scattering by air molecules, which scales roughly as the inverse fourth power of wavelength, supplemented by aerosol scattering with a somewhat weaker wavelength dependence. The result is a zenith spectrum that is markedly brighter in the blue than in the near-infrared. Cloud droplets, by contrast, are several micrometers across, far larger than visible wavelengths, so they scatter light in the Mie regime, where the wavelength dependence nearly vanishes. A cloudy zenith sky is therefore spectrally flat, its color washed out by multiple scattering inside the cloud layer. The transition from a steep spectral slope to a shallow one is the signature the researchers set out to exploit.

Quantitatively, the team defined a spectral slope from the logarithm of the ratio of measured signals at two wavelengths, divided by the logarithm of their wavelength ratio. Crucially, the slope of the raw measured signal mixes two contributions: the true atmospheric slope and a wavelength-dependent instrumental response arising from the detector’s uneven sensitivity across channels. Using calibration constants obtained by the improved Langley method together with extraterrestrial solar spectral irradiance, the authors separated the observed slope into atmospheric and instrumental parts, isolating a quantity they call the atmospheric spectral slope. Radiative transfer simulations with the SBDART model confirmed the underlying theory: this slope is strongly negative in cloud-free atmospheres, shifts only moderately with rising aerosol loading, and moves systematically toward less negative values whenever clouds are present, whether thin or thick.

To test the idea in the real world, the team analyzed zenith radiance observations from eight SKYNET sites spanning radically different environments: urban-coastal Chiba, the marine background site of Cape Hedo in Okinawa, the continental outflow receptor of Fukue Island, the mid-latitude maritime sites of Miyako and Sendai, tropical Burgos in the Philippines, and the polar extremes of Syowa Station in Antarctica and Ny-Ålesund in the high Arctic, with data extending from 2010 to 2024. Ground truth came from the MODIS instruments on NASA’s Terra and Aqua satellites, using a stringent collocation scheme: only scenes where all nine surrounding one-kilometer pixels agreed on a confident clear or cloudy label with high quality assurance were retained, minimizing cloud-edge and geolocation errors.

The observational spectra matched the simulations beautifully. Across every site, cloudy conditions produced flatter zenith spectra than clear conditions, and the atmospheric slope computed for four wavelength pairs spanning the visible and near-infrared, 0.400 to 0.675, 0.400 to 0.870, 0.500 to 0.675, and 0.500 to 0.870 micrometers, showed clean separation between the two sky states. Under clear skies the slope depended strongly on solar zenith angle and aerosol optical thickness, becoming more negative at long slant paths and less negative as aerosol loading increased. Under clouds, the slope was governed mainly by cloud optical thickness and quickly saturated for optically thick cloud, consistent with the physics of multiple scattering. Importantly, no plausible amount of aerosol alone could reproduce the cloudy-sky signature, so the method resists the classic confusion between hazy and cloudy air.

Turning the physics into an operational classifier, the authors used receiver operating characteristic analysis over pooled observations to find optimal decision thresholds, adding a symmetric buffer zone that labels borderline observations as undetermined rather than forcing a guess. All four wavelength pairs performed impressively, with areas under the ROC curve between 0.978 and 0.984 and sensitivity and specificity each above roughly 0.94. The 0.500 to 0.675 micrometer pair emerged as the best overall performer, achieving the highest balanced accuracy of 95.3 percent across sites with the smallest site-to-site variability, a practical advantage since those channels also show stable long-term calibration behavior in the POM-02 instruments used by SKYNET.

Two independent validations bolstered confidence. First, collocated direct normal irradiance measurements from pyrheliometers, which track the unscattered solar beam and are acutely sensitive to cloud obstruction, showed exactly the expected pattern: cases the slope method called clear had high, smoothly declining irradiance, cases called cloudy showed strongly attenuated or near-zero beam radiation, and the undetermined class sat neatly in between, functioning as an honest buffer for thin, broken, or transitional skies. Second, when the team applied the new screening to aerosol retrievals and compared the resulting aerosol optical thickness and Angstrom exponent distributions with those from the established SR-CEReS cloud-screening package, the agreement was close across all eight sites and all seasons, from the very clean polar atmospheres with Angstrom exponents above two to the heavier mid-latitude and tropical loadings. Notably, the new method often retained slightly more observations without shifting the mean aerosol properties, suggesting it recovers usable clear-sky data that statistical screens discard.

The implications reach beyond SKYNET. Because the method rests on the physical spectral behavior of zenith radiance rather than network-specific processing, the authors note it could be extended to other ground networks such as AERONET that acquire comparable measurements. As a computationally light preprocessing step, it could screen cloud contamination before aerosol retrieval and serve as a first filter before cloud-property retrievals from the very same zenith radiance data. The authors caution that their evaluation is statistical rather than event-by-event, and that specific dust outbreaks, smoke plumes, and thin cirrus cases would benefit from confirmation with lidar or other ancillary observations. Still, the core message stands: a decades-old instrument staring at the color of the zenith sky, once its own spectral quirks are accounted for, can reliably tell the difference between a cloudy sky and a clear one, and that simple act of discrimination quietly strengthens the entire observational foundation of aerosol and cloud climate science.

Subject of Research: Spectral detection of clear- and cloudy-sky conditions using zenith sky radiance from SKYNET sky radiometers

Article Title: Spectral signatures of zenith sky radiance from surface-based sky radiometers: implications for clear- and cloudy-sky detection

Article References: Khatri, P., Irie, H., Kobayashi, H., Morino, I., Jin, Y., Hayasaka, T., Iwabuchi, H., Takamura, T., & Takayama, Y. (2026). Spectral signatures of zenith sky radiance from surface-based sky radiometers: implications for clear- and cloudy-sky detection. Atmospheric Measurement Techniques, 19(19), 6447-6467. https://doi.org/10.5194/amt-19-6447-2026

Image Credits: AI Generated

DOI: 10.5194/amt-19-6447-2026

Keywords: sky radiometer, SKYNET, zenith radiance, cloud screening, aerosol optical thickness, spectral slope, Rayleigh scattering, Mie scattering, MODIS cloud mask, direct normal irradiance, atmospheric remote sensing, polar aerosols