Earthquakes, tsunamis and volcanic eruptions push air upward. The resulting waves ripple the ionosphere, the electrically charged layer of the upper atmosphere, and those ripples show up in the total electron content (TEC) that GPS-type receivers measure along their line of sight to each satellite. A natural way to catch such ripples without many labeled examples is reconstruction-based anomaly detection. A model trained only on normal ionospheric data rewrites each new signal as its nearest "normal" version, and whatever it cannot rewrite, the residual, is flagged. This proposal argues that these detectors fail unevenly across the globe. The ionosphere near the magnetic equator and in the auroral zones is naturally turbulent and full of wave-like structure. A model trained there learns that wave-like structure is normal and reproduces a tsunami ripple instead of removing it. The residual then shrinks toward zero exactly where many coastlines and sparse sensor networks are. This failure is called the identical shortcut. We hypothesize that its strength depends on the latitude group, and that a per-station threshold cannot fix it because the problem is lost separability, not an offset.
The opportunity is to borrow a distinction from medical anomaly-detection fairness and make it measurable for geophysical sensors. That literature separates a uniform upward shift in reconstruction error, which recalibration fixes, from a loss of the model's ability to tell normal from abnormal within a subgroup, which it does not. We will measure how much of an injected, physically shaped wave each detector "heals" in each latitude band. We then tune the detector's reconstruction bottleneck, meaning how much of the input it is allowed to see, separately per band until the healing is equalized. Success means that after per-station calibration, the worst latitude band's detection performance moves up toward the best.
Core Claim
The claim is specific and falsifiable. For reconstruction detectors trained on multi-latitude TEC, the fraction of a precursor-like perturbation that survives into the residual is systematically lower in equatorial and high-latitude bands than at mid-latitudes. This gap persists after per-station threshold calibration. A detector whose masking strength is chosen per band, using physically realistic synthetic wave injections, closes most of it. We define the retained-residual fraction for latitude group $g$ as
$$\rho_g = \mathbb{E}_{\delta}\left[\frac{\lVert (x - \hat{x}) \odot m_\delta \rVert}{\lVert \delta \rVert}\right],$$
where $x = x_{\text{norm}} + \delta$ is a real normal TEC segment from group $g$ with an injected disturbance $\delta$, $\hat{x}$ is the model's reconstruction, and $m_\delta$ masks the time-and-space support of the injection. A value of $\rho_g$ near 1 means the model removed the disturbance and left it in the residual, so it is detectable. A value near 0 means the model copied the disturbance, which is the identical shortcut.
Where This Sits in Prior Work
Three ionospheric papers supply the application and its gaps. GNSS TEC-Based Earthquake Ionospheric Perturbation Detection Using a Novel Deep Learning Framework reconstructs expected TEC with an LSTM encoder-decoder conditioned on space-weather indices (Kp, F10.7, ap) and flags residuals beyond ±2σ. It found negative-then-positive anomalies 4–8 days before the 2016 Xinjiang earthquake at a handful of stations in one region, and it never asks whether its residual behaves the same elsewhere. SegED-ConvLSTM forecasts global TEC maps from CODE Global Ionospheric Maps. Its role here is motivating evidence that error differs between equatorial, mid- and high-latitude bands, but it studies forecasting, not anomaly separability. Deep Learning Driven Detection of Tsunami Related Internal Gravity Waves (also Constantinou et al.) takes the supervised route: it trains a ResNet50 on image encodings of slant TEC and needs multi-station voting to push precision from 34.7% to 100%. It is our strongest competitor and a reminder that single-station detectors raise many false alarms.
Three fairness papers constrain the design. Bias in Unsupervised Anomaly Detection in Brain MRI showed on a VAE that shifts in scanner, sex and race separate into additive "domain shift" and within-group "methodological bias." We reuse that split directly, evaluating within each latitude group, but on time series and with modern detectors rather than one VAE. (Predictable) performance bias in unsupervised anomaly detection found that subgroup performance tracks training representation and that 50% balancing still left significant gaps (p < 0.01). This negative result rules out rebalancing as the fix, and it poses the question we try to answer: what inductive constraint decouples reconstruction difficulty from scoring? FairDiffusion reweights group losses through Bayesian optimization. We treat that as an alternative to test. Loss reweighting targets how well each group is reconstructed, and our hypothesis is that equatorial groups are already reconstructed too well, perturbations included.
The method comes from anomaly detection. DRÆM trains a discriminator on the pair (input, reconstruction) using synthetic Perlin-noise anomalies, reaching 98.0 image AUROC on MVTec AD. Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals blocks copy-through by masking large fractions of the input at several ratios. We borrow both ideas and change two things. The synthetic anomalies become physically shaped traveling disturbances instead of textures, and the masking ratio becomes a per-group quantity calibrated against $\rho_g$ instead of a global hyperparameter.
Research Program
We build a masked reconstruction model for slant-TEC sequences. It is a transformer that sees a station's recent TEC, local solar time, magnetic latitude and space-weather indices, and reconstructs masked spans of the series. A DRÆM-style discriminator then reads the observed and reconstructed series together and outputs a per-time-step anomaly score. Training uses only data with no catalogued hazard. Synthetic injections are drawn from a physics-shaped family: wave packets with periods and apparent horizontal speeds in the ranges typical of acoustic and gravity-wave disturbances, amplitudes scaled to the local background, and a propagation delay across neighboring pierce points (the points where each line of sight crosses the ionosphere) that is consistent with a moving source. This last feature separates them from the stationary textures DRÆM uses.
The central analysis measures $\rho_g$ for latitude bands (equatorial anomaly crests, geomagnetic equator, mid-latitude, auroral), split by local time and by quiet versus storm days. We do this for four detector families: the LSTM-style reconstruction of Xiong et al., a ConvLSTM, our masked transformer at a single global mask ratio, and a diffusion denoiser at a single global noise level. If $\rho_g$ does not differ across bands after controlling for injection amplitude relative to local variance, the core hypothesis fails and we report that.
If it does differ, the intervention is group-matched healing. For each band, we choose the mask ratio, or the diffusion noise level, that brings $\rho_g$ on held-out quiet data to a common target. We impose a cap on the false-residual rate on uninjected data, so that the model cannot game the target by raising residuals everywhere. Scores are then converted to per-station quantiles. This keeps calibration as a baseline component rather than the thing under test.
Baselines
The naive reference point is the Xiong et al. recipe as published: LSTM encoder-decoder reconstruction with space-weather inputs and a global ±2σ residual rule. It is retrained on the same stations, time span and splits as every other method. It tells us how much any change is worth. Beating it within the worst latitude band is necessary but not sufficient.
The baseline we consider most dangerous is the same reconstruction model with per-station, per-local-time quantile thresholds. It removes all additive domain shift at essentially zero cost. It gets the same training data and an equal hyperparameter search budget, with thresholds set on the same quiet-period calibration windows as ours. Our method must beat it on worst-group within-group AUROC. This baseline defines the headline test.
The strongest existing detector is the supervised ResNet50 on Gramian Angular Difference Field images with spatial voting from Deep Learning Driven Detection of Tsunami Related Internal Gravity Waves. On the out-of-sample 2015 Illapel event it reports F1 of 91.7%, precision of 100% and recall of 84.6%, against 96.2% recall and 34.7% precision without voting. We will reproduce it on the same Maule/Tohoku-Oki/Haida-Gwaii training split and Illapel test. We compare at matched voting: our per-station scores go through the same voting rule. Beating it means matching its Illapel F1 while holding precision at a higher rate on a held-out low-latitude event it was never trained near.
Ablations isolate the key ingredient. The first uses one global mask ratio and changes nothing else. The second replaces physics-shaped injections with DRÆM's generic Perlin-texture anomalies. The third replaces per-group mask calibration with FairDiffusion-style group loss reweighting under the same tuning budget. If the global-ratio ablation matches the full method, the per-group calibration contributes nothing and the claim fails.
Why not just…?
Why not just calibrate thresholds per station? This is the most serious alternative. On pure offset shifts it will match us, because Bias in Unsupervised Anomaly Detection in Brain MRI shows domain shift acts as a uniform elevation that recalibration removes. What recalibration cannot restore is separation that the reconstruction has already destroyed. If the model copies the wave, no threshold recovers it. That same paper shows within-group AUPRC collapses of 69–84%, which is the part recalibration does not touch. The decisive experiment compares within-band AUROC under per-station thresholds against our method. If the gap is negligible, per-station calibration is the answer and we will say so.
Why not just use the supervised CNN with spatial voting? It works well on Illapel. However, it was trained on three events and tested on one, all at mid-latitude subduction zones, and it cannot flag disturbance types it never saw, such as volcanic waves like Tonga 2022 (Feng et al.). Its 34.7% raw precision also shows that voting carries much of the load. We test it on a low-latitude event and on synthetic injections across bands. If it keeps high precision there, the case for reconstruction weakens to open-set coverage alone.
Why not just rebalance or reweight latitude groups during training? (Predictable) performance bias in unsupervised anomaly detection found significant gaps persist at 50% balance. Our mechanism also predicts that reweighting makes things worse: it improves reconstruction in turbulent bands, which lowers $\rho_g$ further. The reweighting ablation tests this prediction directly.
Why not just add more space-weather inputs or model capacity? Xiong et al. already condition on Kp, F10.7 and ap. More capacity is precisely what makes copying easier. We track $\rho_g$ as model size grows to check this.
Proposed Experiments
Normal-period training and injection experiments use CODE Global Ionospheric Maps from 2012–2024, the SegED-ConvLSTM data span, sampled across all latitude bands, together with slant-TEC from the tsunami benchmark stations. Real-event tests use the Maule, Tohoku-Oki and Haida-Gwaii training events with Illapel held out, following Constantinou et al.. Tonga 2022 serves as a low-latitude held-out event, using the CMONOC Xinjiang case of Xiong et al. as a pre-seismic sanity check. Labeled real events are few and cluster at mid-latitudes. For that reason the controlled injection benchmark carries the statistical weight, and real events serve as external checks rather than proof.
We report within-band AUROC and AUPRC, subgroup AUROC in the sense of Schulz et al., worst-band performance, the $\rho_g$ curves, false alarms per station-day on storm days with Kp ≥ 5, and detection delay on real events. There is no reported result on Lacuna for any reconstruction detector's per-latitude anomaly-detection performance, so the per-band baselines here are new measurements.
The decisive checks are three. First, $\rho_g$ must differ across bands at matched relative amplitude. Second, that difference must survive per-station calibration as a within-band AUROC gap. Third, per-band mask calibration must narrow the gap more than both the global-ratio and reweighting ablations, without raising storm-day false alarms. If the first check fails, the project reduces to a useful negative result: reconstruction detectors transfer across latitudes once thresholds are local.
References
- GNSS TEC-Based Earthquake Ionospheric Perturbation Detection Using a Novel Deep Learning Framework
- A New and Tested Ionospheric TEC Prediction Method Based on SegED-ConvLSTM
- Deep Learning Driven Detection of Tsunami Related Internal Gravity Waves: a path towards open-ocean natural hazards detection
- Bias in Unsupervised Anomaly Detection in Brain MRI
- (Predictable) performance bias in unsupervised anomaly detection
- FairDiffusion: Enhancing Equity in Latent Diffusion Models via Fair Bayesian Perturbation
- DRÆM - A discriminatively trained reconstruction embedding for surface anomaly detection
- Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals for Cross-Domain Audio Deepfake Detection
- Analysis of Ionospheric Anomalies before the Tonga Volcanic Eruption on 15 January 2022