However, because airglow emits only a very small amount of light, this gravity wave signature is only visible on moonless nights. This limited window of opportunity to observe gravity waves along with the challenges inherent in finding a relatively rare event in the large amount of data flowing from the DNB sensors on the Suomi-NPP, NOAA-20, and NOAA-21 satellites means that there are very few labeled examples of gravity wave occurrences available to train an automated detection algorithm. In addition, at the sensitivity level required to detect gravity waves on moonless nights, there is a high degree of noise which is also a significant challenge in building an automated detector.
To attempt to address these concerns, particularly the very limited size of available training dataset, we utilize a combination of mathematical methods to isolate and enhance key features of gravity waves which can then be utilized by a simple classification algorithm that can be trained on the small amount of data available. We utilize methods from a number of areas of mathematics including harmonic analysis and topological data analysis, which all have the advantage of being highly interpretable and explainable, so that we can at each step understand what our detection algorithm is basing its decisions on.
References:
[1] Miller, S. D., Straka, W. C., Yue, J., Smith, S. M., Alexander, M. J., Hoffmann, L., Setvák, M., & Partain, P. T. (2015). Upper Atmospheric Gravity Wave details revealed in nightglow satellite imagery. Proceedings of the National Academy of Sciences, 112(49).

