The technical foundation of the work is transfer learning, a technique that repurposes neural networks already trained on vast general image collections rather than building one from scratch. The team first experimented with GoogLeNet, a comparatively compact network of about seven million parameters, before settling on InceptionV3, a larger architecture with roughly 24 million parameters that performed most evenly across all four classes. The final classification layers of the pretrained network were replaced so the output became a probability vector over the four sky-condition categories. Training images were rescaled to the network’s required 299 by 299 pixel input and augmented with random horizontal and vertical flips and rotations of up to ten degrees, a standard practice that prevents the model from simply memorising its training examples. On a desktop computer equipped with an NVIDIA GeForce RTX 4090 graphics card, training the final network took about twelve minutes, and classifying the entire decade of images took just 14.2 hours, or roughly six milliseconds per image.

What makes the study unusual is not the network itself but the painstaking construction of its ground truth. Supervised learning lives or dies by the quality of its labelled examples, and auroral images are notoriously ambiguous. Faint aurora blends into thin cloud, moonlit clouds mimic red emission bands, and frost on the camera dome can masquerade as a celestial structure. Partamies labelled thousands of images by hand, then ran several iterative rounds in which an unfinished classifier proposed labels, random samples were re-checked by the human expert, and any image whose content could not be judged unambiguously was banished to an ‘Unclear’ category and discarded. Masking out the lowest ten degrees of elevation near the horizon, where mountains, instrument domes and metadata clutter the frame, proved essential for producing clean labels. The resulting training set of carefully vetted images, spanning every winter season from 2016 to 2025, is itself published as a reference dataset for future method comparisons.

Validation was equally rigorous. The team divided the labelled data into training, validation and independent test sets, and additionally drew 2,000 random images per class from the full classified archive to estimate real-world performance. ClearAurora and CloudyNoAurora images were recognised with 97 percent success, CloudyAurora with 93 percent, and CloudyNoAurora with 90 percent in testing. The most common confusions occurred near class boundaries, as expected: images with barely half-sky cloud cover, or aurora so faint it barely registers. A separate cross-check with the observatory’s older Nikon DSLR camera system, which overlapped with the current Sony setup in early 2016, showed that the classifier transfers to different hardware with match rates between 92 and 97 percent, suggesting the discrepancies stem from instrument properties such as white balance rather than changing sky conditions.

Because clouds dominate the dataset so thoroughly, the authors validated their cloud statistics against a completely independent instrument: a co-located cloud sensor operated with University College London, which infers sky clarity from the temperature difference between the sky and the sensor itself. Monthly values from the two methods correlated with a coefficient of 0.86, a strong agreement given how differently the two measurements are produced. This matters because the cloudiness statistics are not a side product but a central scientific result. Svalbard sits under some of the cloudiest winter skies in the Arctic, and previous long-term meteorological studies had reported winter cloudiness around 60 percent with an increasing trend between 1981 and 2010. The new image-based statistics confirm the high cloud fraction and add, for the first time, a decade of consistent nighttime cloud measurements from a location where the Sun does not rise for months.

The auroral findings are equally consequential. Auroral occurrence over Svalbard averages about 25 percent of imaging time and shows no correlation with the solar cycle, despite the dataset spanning the rise from solar minimum toward the recent maximum. In 2019, a year of very low sunspot activity, auroral occurrence peaked at 33 percent; in 2023, a year of higher solar activity, it fell to just 16 percent. The controlling variable in both cases was cloudiness, which was exceptionally low in 2019 and exceptionally high, at 78 percent, in 2023. The result echoes earlier Finnish studies based on visual inspection of all-sky film archives, which likewise found no obvious link between auroral occurrence and solar cycle evolution at auroral oval latitudes, even though the complexity and intensity of auroral structures do vary with geomagnetic activity.

The diurnal pattern is equally distinctive. Expressed in Magnetic Local Time, auroral occurrence over Svalbard peaks in the late morning hours around 07:00 to 11:00 MLT, with a secondary maximum in the pre-midnight sector around 17:00 to 21:00 MLT, and minima in the early afternoon and around magnetic midnight. This double-peaked distribution held steady across all ten years. January and December emerge as the months with the highest auroral occurrence and the clearest skies, while November is decisively the cloudiest month, making it the worst time to hunt for the lights over Svalbard. The observatory’s unique position near the poleward boundary of the auroral oval, combined with the polar night during which optical observations are possible around the clock, means the dataset also captures the dayside aurora, a phenomenon largely inaccessible at lower latitudes.

The practical payoff extends well beyond statistics. KHOnet2026 now runs in real time at the Kjell Henriksen Observatory, classifying each new image as it arrives and continuously expanding the labelled archive. For researchers studying the fine structural evolution of aurora, the ability to prune away the roughly two-thirds of images obscured by cloud saves enormous amounts of computing time and human attention. The classified sky conditions also serve as ready-made metadata for the dozens of other optical instruments hosted at the observatory, and the occurrence climatology will help planners of sounding rocket campaigns and EISCAT radar experiments schedule observations when skies are statistically most likely to cooperate. The team notes that future refinements could employ self-supervised learning to expand the labelled set, or Vision Transformers as an alternative architecture, and that adding labelled images from the older DSLR archive could extend the classified time series by another seven years.

There are honest caveats. All manual labelling was performed by a single expert, introducing an unavoidable subjective element into the class boundaries, and twilight and moonlit conditions, which the authors deliberately retained in all classes, add a layer of ambiguity that a human cannot always resolve either. Tests excluding twilight and bright-moon images shifted the occurrence statistics by several percentage points but preserved the overall diurnal behaviour, and the published data files include solar and lunar elevation angles so that future users can filter the dataset to their own standards. Even so, the study stands as a demonstration of how machine learning, applied with unusual care to its weakest link, the training labels, can transform a decade of raw camera data into a coherent scientific record. For a phenomenon as famously unpredictable as the aurora, the message from Svalbard is unexpectedly mundane: above all, it is the weather that decides what you see.

Subject of Research: Automatic classification of auroral and cloud occurrence in a decade of all-sky camera images from Svalbard using convolutional neural networks

Article Title: High-latitude auroral and cloudiness occurrence from automatic image classification

Article References: High-latitude auroral and cloudiness occurrence from automatic image classification. (n.d.). https://doi.org/10.5194/angeo-44-1003-2026

Image Credits: AI Generated

DOI: 10.5194/angeo-44-1003-2026

Keywords: aurora, all-sky camera, Svalbard, machine learning, convolutional neural network, cloud occurrence, auroral occurrence, transfer learning, polar night, solar cycle, Kjell Henriksen Observatory, space weather