Using Convolutional Neural Networks to Detect Atmospheric Gravity Waves in AIM/CIPS Data

A new data set is being provided from the AIM (Aeronomy of Ice in the Mesosphere) satellite's CIPS instrument; it provides high resolution raleigh albedo data at 50-55km - fluctuations in these measurements can be used to observe gravity waves. Since February 2016, the instrument has been in "continuous imaging/CI" mode, giving global coverage. Due to the large size of both the original polar and the new CI data sets, hand-classifying GWs is likely to be a significant source of error and bias.

We design and train a convolutional neural network using a relatively small set of hand-labeled images. We then use the network to classify the entire CIPS/RAA data set and provide a global map of GW activity. Our initial network, when trained with 3000 images, reached 92% accuracy on a 2167 image evaluation set with an intentional bias toward false-positives. We bias this way since false-positives can be detected and corrected easily (positives account for <4% of the entire data set) - false negatives would be missed entirely outside of an evaluation run.

We look to include data sets from other sources in order to corroborate the AIM data and to provide GW occurrence data at other altitudes.