[Paper Review] Correlation of Auroral Dynamics and GNSS Scintillation with an Autoencoder

Kara D. Lamb, Garima Malhotra+8 more|arXiv (Cornell University)|Oct 4, 2019

Ionosphere and magnetosphere dynamics14 references4 citations

TL;DR

This paper proposes an unsupervised approach using a residual autoencoder (Res-AE) to learn low-dimensional representations of auroral images from the THEMIS network, followed by clustering via t-SNE and UMAP. It finds that specific clusters—particularly those corresponding to discrete and arc-like auroral structures—show strong correlation with elevated GNSS phase scintillation indices (σφ), indicating these dynamic auroral features are key drivers of ionospheric irregularities.

ABSTRACT

High energy particles originating from solar activity travel along the the Earth's magnetic field and interact with the atmosphere around the higher latitudes. These interactions often manifest as aurora in the form of visible light in the Earth's ionosphere. These interactions also result in irregularities in the electron density, which cause disruptions in the amplitude and phase of the radio signals from the Global Navigation Satellite Systems (GNSS), known as 'scintillation'. In this paper we use a multi-scale residual autoencoder (Res-AE) to show the correlation between specific dynamic structures of the aurora and the magnitude of the GNSS phase scintillations ($σ_ϕ$). Auroral images are encoded in a lower dimensional feature space using the Res-AE, which in turn are clustered with t-SNE and UMAP. Both methods produce similar clusters, and specific clusters demonstrate greater correlations with observed phase scintillations. Our results suggest that specific dynamic structures of auroras are highly correlated with GNSS phase scintillations.

Motivation & Objective

• To identify physically meaningful patterns in auroral dynamics without relying on human-labeled categories.
• To investigate whether unsupervised representation learning can reveal correlations between auroral morphology and GNSS scintillation.
• To determine if specific dynamic auroral structures are more strongly associated with high phase scintillation (σφ) than others.
• To evaluate the effectiveness of t-SNE and UMAP in preserving physical correlations in the latent space of autoencoded auroral images.
• To explore the generalizability of the method across different geographic sites using multi-site auroral data.

Proposed method

• A residual autoencoder (Res-AE) is trained on 35,277 THEMIS all-sky imager (ASI) images from Fort Simpson, Canada, to learn hierarchical, multi-scale features in the auroral image data.
• The encoder of the Res-AE maps input images into a lower-dimensional latent space, preserving spatial and structural information through residual blocks and strided convolutions.
• Latent representations from the Res-AE are projected into 2D using t-SNE and UMAP for visualization and clustering, with UMAP using 15 nearest neighbors to define local manifold structure.
• Spectral clustering is applied to the UMAP-embedded latent space to identify distinct image clusters based on structural similarity.
• Clusters are then correlated with co-located GNSS phase scintillation indices (σφ) measured at the CHAIN receiver in Fort Simpson.
• The method is evaluated using a separate set of 7,700 manually annotated auroral images (classified as arc, diffuse, discrete, moon, clear, clouds), excluding cloudy images to focus on ionospheric physics.

Experimental results

Research questions

• RQ1 Which dynamic auroral structures, as captured in ASI images, are most strongly correlated with elevated GNSS phase scintillation (σφ) values?
• RQ2 Can unsupervised representation learning via a Res-AE and dimensionality reduction (t-SNE/UMAP) reveal physically meaningful clusters in auroral images without human-annotated labels?
• RQ3 Do clusters identified in the latent space of the autoencoder correspond to known auroral morphologies (e.g., arcs, discrete features) and show measurable differences in σφ?
• RQ4 How do the latent representations and resulting clusters vary across different geographic sites, and what constraints affect transferability of the method?
• RQ5 Can the latent space embeddings preserve topological relationships that reflect ionospheric electron density irregularities linked to scintillation?

Key findings

• Specific clusters identified in the UMAP projection of the Res-AE latent space—particularly clusters 0 and 5—show significantly higher median values of GNSS phase scintillation (σφ) compared to other clusters.
• These high-σφ clusters are predominantly composed of images labeled as 'discrete' and 'arc' in the manual annotations, suggesting that these dynamic auroral structures are strongly linked to ionospheric irregularities.
• The cluster containing 'moon'-labeled images also shows elevated σφ, indicating that brightness and intensity of auroral emissions correlate with scintillation magnitude, likely due to enhanced particle precipitation.
• t-SNE and UMAP produce highly similar clusterings in the latent space, validating the robustness of the learned representations across different non-linear dimensionality reduction techniques.
• The method reveals physically meaningful correlations between auroral morphology and ionospheric disturbances without relying on pre-defined human-labeled categories, demonstrating the value of unsupervised feature learning.
• The approach shows site-specific behavior, with greater separation between images from different sites (e.g., RANK vs. FSIM) than between image classes at a single site, suggesting potential for future adaptation via edge masking or domain adaptation.