Spatiotemporal Reconstruction of Auroral Precipitations with Neural Fields

Nazar Misyats, Mizuki Fukizawa, Yoshimasa Tanaka, Satoshi IkehataarXiv 2026 · arXiv

TL;DR

Researchers have developed a new method to reconstruct the hidden "engine" of the aurora—the field of precipitating electrons—from ground-based video. By combining coordinate-based neural networks (Neural Fields) with a recurrent temporal model, they can turn sparse, multi-view auroral footage into high-fidelity, physically consistent 3D reconstructions significantly faster than traditional tomographic methods.

What Problem Does This Paper Address?

Auroras are more than just a light show; they are a visual map of the complex electromagnetic interactions between the Earth's magnetosphere and the upper atmosphere. To understand these processes, scientists need to know the electron precipitation field: the spatial distribution and energy spectra of electrons as they slam into our atmosphere.

However, recovering this field is a "severely ill-posed" inverse problem. Ground-based cameras capture 2D projections (images) of 3D light emissions. To make matters worse, these emissions are an indirect, energy-dependent response to the incoming electrons. Traditional Aurora Computed Tomography (ACT) tries to solve this by carving the atmosphere into a discrete 3D grid, but it is computationally expensive (taking hours per frame), struggles with temporal consistency in videos, and often requires manual "tuning" of camera calibrations.

Methods

The authors propose a physics-constrained neural field approach that treats the precipitation field as a continuous function rather than a discrete grid.

  1. Neural Field Representation: Instead of a voxel grid, they use a Multi-Layer Perceptron (MLP) that takes a 2D horizontal position $(x, y)$ and a temporal latent code as input, outputting the local electron energy spectrum.
  2. Temporal Dynamics via GRU: To handle video, they don't just process frames independently. They use a Gated Recurrent Unit (GRU) to evolve a latent code over time. This ensures that the reconstructed aurora moves smoothly and logically from one frame to the next.
  3. Differentiable Forward Model: The system "renders" its predicted electron flux into synthetic camera images using a physical emission model (the GLOW model). This includes a clever accounting for delayed emissions—specifically the 557.7 nm green line, which comes from "metastable" states that stay lit for about 0.7 seconds after the electrons have passed.
  4. Joint Calibration: Since real-world cameras are rarely perfectly calibrated, the pipeline jointly optimizes for camera sensitivities alongside the neural field parameters.

Datasets

The method was evaluated on both simulated and real-world data:

  • Synthetic Datasets: Two scenarios were created—"Patches" (simulating diffuse, pulsating auroras) and "Arc" (a bright, moving sinusoidal arc). These provided a ground truth to measure the accuracy of the recovered energy spectra.
  • Real Observations: Multi-view video captured on March 26, 2023, from three stations in Northern Scandinavia (Tromsø, Skibotn, and Kilpisjärvi).

Key Insights and Ideas

  • Source vs. Effect: Unlike standard Neural Radiance Fields (NeRF) which model the light you see, this model learns the unobserved physical source (electrons) and uses physics to derive the light. This allows the model to generalize better from very few viewpoints (as few as three).
  • The Power of Recurrence: The "aha moment" in the temporal ablation study was that the GRU-based latent codes significantly outperformed independent frame processing or simple continuous-time inputs. It captures the "innovation" (new changes) in the aurora while maintaining the history of the structure.
  • Continuous is Better: By using a continuous representation, the authors avoid the "staircase" artifacts common in grid-based tomography and can represent high-frequency spatial features more efficiently.

Key Results

  • Speedup: The neural field approach takes 3 to 5 minutes to reconstruct a sequence, compared to the hour-per-frame requirement of traditional ACT.
  • Accuracy: On synthetic data, the method achieved significantly lower Mean Absolute Error (MAE) in both total energy flux and mean energy compared to the state-of-the-art Generalized ACT (G-ACT).
  • Visual Fidelity: Re-rendered images from the neural field showed much lower Normalized Root Mean Square Error (NRMSE) against actual observations, meaning the model "explained" the camera data much better than previous methods.

Limitations and Future Work

While a major step forward, the authors note a few hurdles:

  • Sparse Viewpoints: With only three cameras in the real-world test, the problem remains highly under-constrained. Adding more viewpoints or integrating other sensors, like scattering radars, would further stabilize the reconstruction.
  • Hyperparameter Sensitivity: Like many neural field methods, the quality of the result depends on choosing the right regularization strengths (e.g., how "smooth" you expect the energy spectra to be).
  • Spectral Smoothing: The model tends to produce smoother energy spectra than might actually exist in nature (e.g., it might blur a sharp "peak" in electron energy). Future work could involve more sophisticated priors on the shape of the energy distribution.

This research demonstrates that the tools of modern computer vision—specifically neural fields and differentiable rendering—can be powerful allies for space physicists trying to decode the mysteries of the upper atmosphere.

Summary updated

Abstract

Auroras provide remote optical observations of energetic electrons precipitating into the upper atmosphere, and whose energy distribution encodes key information about large scale magnetospheric processes. Recovering the underlying electron precipitation field from ground-based camera images of auroras remains a challenging inverse problem. Existing auroral tomography techniques are often limited to simplified static settings, discrete spatial and energy grids, or strong assumptions on the electron energy spectrum. We propose to model the dynamic auroral electron precipitation field via a temporally conditioned neural network. The state of the precipitation field at each frame is represented by a per-frame latent code driven by learned discrete dynamics. A shared decoder maps horizontal positions and latent states to electron energy spectra. The predicted flux is converted to multi-wavelength volume emission rates and rendered into camera images accounting for delayed emissions. We evaluate the method on synthetic dynamic auroral sequences and real multi-view auroral videos, showing that the proposed representation produces temporally coherent precipitation reconstructions and image projections consistent with the observations.
BibTeX
@misc{misyats2026spatiotemporal,
  author = {Nazar Misyats and Mizuki Fukizawa and Yoshimasa Tanaka and Satoshi Ikehata},
  title = {{Spatiotemporal Reconstruction of Auroral Precipitations with Neural Fields}},
  year = {2026},
  eprint = {2611.19757},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19757},
}
Figure