astronomy/ai produced the result/Earth Planets and Space 2025 · v2
Neural network scans eleven years of night-sky images for mesospheric wave fronts
Researchers trained a YOLOv3 object detector to spot frontal waves in faint airglow pictures from a weather satellite, then used its detections across eleven years to chart where and when the waves appear.
spectrum · one line per step, placed by what the step does · bright lines used AI
Machine-learning detection and variability of mesospheric frontal waves observed by VIIRS day/night band
Earth Planets and Space, 2025
doi:10.1186/s40623-025-02308-4 · record aix-00167 v2 · checked 2026-10-09
- AI was for
- Detection
- Model family
- Convolutional neural network
- Checked by
- Held-out928 tested, 575 worked
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
About 90 kilometres up, in a layer of the atmosphere called the mesosphere, the air glows faintly at night. This airglow acts like a sheet of dye in a fluid: when waves ripple through the air, they press the glowing layer into bright and dark bands that a sensitive camera in orbit can see. Some of these patterns take the form of a sharp front with ripples trailing behind it, known as a frontal wave. Finding them is tedious. The glow is dim, cloud and other weather features in the lower atmosphere can mimic the same shapes, and a satellite builds up enormous numbers of night-time images over years of operation.
The researchers set out to automate the search in images from the Day/Night Band of the VIIRS instrument on the Suomi NPP satellite, a channel sensitive enough to record faint night-time light. Using moonless nights, when moonlight does not swamp the glow, they wanted a list of frontal wave events long enough to show how often the waves occur, how large they are, and how their numbers change with latitude, season and year.
Where AI came in
The AI was a convolutional neural network for object detection called YOLOv3, trained from scratch on images in which people had drawn boxes around frontal waves. Infrared imagery was used while labelling to reject features belonging to the lower atmosphere, and images containing no waves were added so the model would produce fewer false alarms. The labelled set was enlarged by rotating images, giving 3,584 training images, and the network was trained for 4,000 epochs. On withheld test images containing 928 labelled wave objects it found 575.
The trained detector then read every available moonless Day/Night Band image from January 2012 to June 2023 — 515,187 of them — and flagged 3,283 as containing waves. People looked through those flagged images, discarding mistakes and duplicates, which left 1,150 confirmed events. The network therefore stood in for the eye-by-eye survey of half a million pictures; human judgement remained, but only on the small set the model proposed. That event list was the basis for the occurrence statistics, which the authors compared with a model of atmospheric tides.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
A YOLOv3 object detector was trained on manually labelled bounding boxes of frontal waves in Suomi NPP VIIRS Day/Night Band airglow images, with negative background images added to reduce false positives. On the held-out labelled test images the model reached 83.19% average precision at an IoU threshold of 0.2, falling to 30.74% at an IoU threshold of 0.6. Applied to 515,187 moonless images from January 2012 to June 2023, the detector returned 3,283 positive images, and a visual survey confirmed 1,150 frontal wave events. The resulting climatology shows monthly occurrence falling from around 15 in 2012 to around 5 in 2022, a high occurrence peak at equatorial latitudes and weaker peaks at winter mid-latitudes, which the authors relate to regions of large migrating diurnal and semidiurnal tidal temperature amplitude.
How AI was used
VIIRS DNB level-1B granules from moon-free nights were converted to 8-bit images by clipping, minimum subtraction, median scaling and a fitted cumulative-distribution transform, then resized. Frontal waves were identified by human inspection assisted by a YOLOv3 detector from a preliminary study, with VIIRS M15 infrared imagery used to reject tropospheric look-alikes, and each object was enclosed in an axis-aligned bounding box. The labelled images were split into training and testing sets and augmented by 90, 180 and 270 degree rotations, and 800 background images without frontal waves were added, giving 3,584 training images. A YOLOv3 model with a Darknet-53 backbone, run at an input resolution of 1,028 by 1,028 instead of 608 by 608, was trained on this set for 4,000 epochs. Performance was measured on the withheld labelled images across IoU thresholds at a confidence threshold of 0.2, and separately on all qualifying 2012 images with manual inspection of every positive. The trained detector was then run over all available moonless DNB images from January 2012 to June 2023, and its positive detections were visually surveyed to drop false positives and duplicates before the event list was used for the occurrence analysis.
The shape of the work
Structural · the record, drawn
no AI
Collect moonless DNB observations
Obtaining raw data, whether by measurement, download or retrieval.
The Suomi NPP VIIRS DNB was used in this study.where the paper describes this · verbatim
no AI
Preprocess granules into 8-bit images
Cleaning, filtering, normalising or labelling data already obtained.
Finally, we convert the data to 8-bit integers and bilinearly resize itwhere the paper describes this · verbatim
AI
Label frontal waves and build augmented training set
Cleaning, filtering, normalising or labelling data already obtained. The AI stood in for manual curation.
The selection process included both visual inspection by human eyes and detection by a YOLOv3 model we trained in a preliminary studywhere the paper describes this · verbatim
AI
Train YOLOv3 frontal wave detector
Fitting model parameters, including fine-tuning an existing model. The AI stood in for manual curation.
We trained the model with the aforementioned 3,584 images for 4,000 epochswhere the paper describes this · verbatim
AI
Evaluate detector on held-out and full-year imagery
Testing outputs against ground truth. The AI stood in for manual curation.
The performance of the trained model is tested with the 880 images containing 928 frontal wave objectswhere the paper describes this · verbatim
AI
Apply detector to eleven years of imagery
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
the ML model was applied to the whole available images of Suomi NPP VIIRS/DNB from January 2012 to June 2023where the paper describes this · verbatim
no AI
Visually confirm detections and remove duplicates
Testing outputs against ground truth.
We conducted a visual survey of the positive images and excluded FP detections and duplicate detections.where the paper describes this · verbatim
no AI
Derive occurrence variability and compare with tidal model
Extracting understanding from model behaviour.
With the detected 1,150 frontal wave events, the 11-year variability of frontal waves was studied.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The event list underpinning the eleven-year climatology was produced by the trained detector; manual inspection only removed false positives from what the model proposed
The YOLOv3 machine learning model, short for “You Only Look Once version 3,” was trained to detect frontal wave eventswhere the paper describes this · verbatim
The performance of the trained model is tested with the 880 images containing 928 frontal wave objectswhere the paper describes this · verbatim
The trained weight of the network of YOLOv3 is provided in https://doi.org/10.5281/zenodo.14812061.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of YOLOv3 preliminary frontal-wave detectorWhich version of the model was used is not stated.
About this article
Record aix-00167, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error