Automated Analysis of Ripple-Scale Gravity Wave Structures in the Mesosphere Using Convolutional Neural Networks

TL;DR

A deep learning framework using convolutional neural networks to automatically detect ripple-scale wave structures in all-sky airglow images and performs a statistical characterization of the detected ripples to infer underlying physical mechanisms, such as localized instability conditions and background wind filtering.

Abstract

The mesosphere and lower thermosphere (MLT), spanning approximately 80--100~km in altitude, is a region of intense dynamical activity where atmospheric gravity waves amplify due to decreasing air density. These waves often undergo breaking, inducing energy and momentum dissipation as well as turbulent mixing. These processes can be described by atmospheric instabilities -- including convective and shear instabilities -- that regulate the vertical coupling between atmospheric layers. Ripple-like structures observed in mesospheric airglow imagery represent small-scale, short-lived signatures of such instabilities. Their occurrence often reflects localized instabilities, critical wave interactions, or wave ducting phenomena. Therefore, detecting and analyzing these features across broad spatial and temporal domains remains a challenge. In this study, we develop a deep learning framework using convolutional neural networks (CNNs) to automatically detect ripple-scale wave structures in all-sky airglow images. Trained on labeled datasets with recognized ripples, the model learns the spatial morphology associated with instability-driven gravity waves and achieves high accuracy in identifying ripple events. Beyond detection, we perform a statistical characterization of the detected ripples to examine their frequency of occurrence, orientation distributions, scales, and geographic and seasonal variability. These statistics are used to infer underlying physical mechanisms, such as localized instability conditions and background wind filtering. This work advances our understanding of instability-driven dynamics in the upper atmosphere through AI-powered detection, while also highlighting the potential of deep learning in scientific research.

Figures (6)

• Figure 1: Schematic illustration of the squeeze-and-excitation (SE) block used in the convolutional neural network. Input feature maps of size $[C, H, W]$ are first passed through a convolutional layer followed by a ReLU activation. The squeeze operation $F_{sq}(\cdot)$ applies global average pooling over the spatial dimensions to generate channel-wise descriptors $Z_c$. These descriptors are then processed by the excitation operation $F_{ex}(\cdot)$, which consists of two fully connected layers with a nonlinear activation and a sigmoid function to produce channel-wise weights. The resulting weights are applied through the scaling operation $F_{scale}(\cdot)$ to recalibrate the original feature maps, which are subsequently passed to the next convolutional layer.
• Figure 2: Example of an all-sky airglow image with several detected ripple events. The ripple structure was automatically identified by the CNN, marked by Cyan starts. The red dots are manual ripple recognitions.
• Figure 3: Distribution of ripple propagation directions for the detected events. The histogram indicates the relative frequency of ripples propagating in each azimuthal sector.
• Figure 4: Seasonal variation of ripple occurrence frequency. The plot shows the mean number of ripple events detected in each season of the year.
• Figure 5: Seasonal distribution of ripple occurrence as a function of local time (LT) hour for spring, summer, autumn, and winter. Bars show the number of detected ripple events in each LT bin, with green representing model detections and red representing human-identified events. The comparison highlights seasonal differences in the diurnal timing of ripple occurrence and the level of agreement between the model and human observations.