Latitude-Conditioned Spectral Gating for Satellite Earthquake-Precursor Detection

Satellites in low Earth orbit, such as DEMETER and the China Seismo-Electromagnetic Satellite (CSES), measure plasma density and electric fields in the ionosphere, the electrically charged upper atmosphere. Several studies report that these measurements change in the days before large shallow earthquakes. The difficulty is that the ionosphere is also shaken by the Sun, by geomagnetic storms, and, near the magnetic equator, by its own strong daily dynamics. Detectors trained to spot precursors therefore produce many false alarms, and they get noticeably worse at low latitudes. We propose a detector that first splits each satellite track segment into frequency bands. A small network then decides how much to trust each band, and that decision depends on where the satellite is relative to the magnetic equator and on local time and solar activity. The model is trained end to end as a binary seismic/non-seismic classifier. The bet is that background disturbances live in latitude-specific frequency bands, while the seismic signature is more stable across latitudes. If so, a latitude-dependent frequency filter should transfer better across latitude belts than either a fixed filter or a time-domain model that is simply told its latitude.

The second goal matters as much as the first. The field has few labelled events (the five-year CSES retrospective covers 47 earthquakes), so a flexible classifier can easily learn "disturbed ionosphere" rather than "earthquake coming." We therefore build in tests with fake earthquake dates, and we hold out entire latitude bands, so that a positive result would actually be believable.

Core Claim

The low-latitude collapse documented across ionospheric deep learning is a frequency-domain problem. Accuracy falls near the Equatorial Ionization Anomaly in TEC forecasting (Ionospheric TEC Prediction Base on Attentional BiGRU, Forecasting Ionospheric foF2 Based on Deep Learning Method), and Xiong et al. note that GNSS precursor detection likewise degrades at low magnetic latitudes (GNSS TEC-Based Earthquake Ionospheric Perturbation Detection Using a Novel Deep Learning Framework). The Equatorial Ionization Anomaly is a pair of dense plasma crests near ±15–20° magnetic latitude, produced when daytime electric fields lift plasma that then slides down field lines (Anomalies of the Ionospheric F2-layer above Guangzhou Region). The crests, and the post-sunset plasma bubbles that form near them, inject structure at characteristic spatial scales along the satellite track. These scales differ from those of the mid-latitude background.

A per-band gate whose weights are a smooth function of magnetic latitude, local time, and solar flux can learn to suppress these regime-specific bands. We claim it will improve detection skill in latitude bands held out of training, relative to the same backbone with a fixed band-pass filter, with an unconditioned learned filter, or with latitude given as an ordinary input. It should do so without gaining skill on fake-date controls.

What Exists and How This Differs

Three supervised precursor detectors frame the problem as classification, and each leaves the latitude question open. SeqNetQuake stacks a convolutional feature extractor and a bidirectional LSTM over 20-observation CSES windows inside the Dobrovolsky radius, with transfer from DEMETER and a three-way label that includes solar/magnetic disturbance (Pre-Earthquake Ionospheric Perturbation Identification Using CSES Data via Transfer Learning). It works purely in the time domain and has no latitude parameterization. It is our main competitor and our source of data conventions, including nighttime-only windows, which that paper found to be less noisy.

Asaly and Gottlieb use a polynomial support vector machine (SVM) on moving-average-detrended GIM TEC. They throw out disturbed days, which leaves only 106 usable earthquakes from 23 years (Using Support Vector Machine (SVM) with GPS Ionospheric TEC Estimations to Potentially Predict Earthquake Events). We keep disturbed periods as hard negatives instead.

Constantinou et al. convert slant TEC into Gramian-angle images for a ResNet50. The raw network reached only 34.7% precision, and multi-station voting was needed to recover precision (Deep Learning Driven Detection of Tsunami Related Internal Gravity Waves). That evidence motivates attacking false positives at the representation level. Their events were also mid-latitude only.

The ED-LSTME approach forecasts background TEC and flags 2σ residuals (GNSS TEC-Based Earthquake Ionospheric Perturbation Detection Using a Novel Deep Learning Framework). It is a two-stage detector that never optimizes the detection decision. It also defers the low-latitude problem to future work.

The mechanism comes from frequency-domain attention outside geophysics. SAAM adds global and local learned weights over Fourier components of a forecaster's hidden states (Deep autoregressive models with spectral attention). FAMNet shows that weighting mid-frequency bands differently from low and high bands helps under CT-to-MRI shift, and that uniform attention across bands costs over 7% (FAMNet). TF-FACE gates frequency bands but with a fixed low/high split (TF-FACE). None of these conditions the band weights on a physical covariate that defines the domain.

The conditioning idea follows the evidence from SegED-ConvLSTM that space-weather inputs must be stratified by latitude band to avoid equatorial collapse (A New and Tested Ionospheric TEC Prediction Method Based on SegED-ConvLSTM). We make that stratification continuous and move it into the frequency domain. Time-domain attention TEC models such as Att-BiGRU and MA-LSTM (Ionospheric TEC Prediction in China Based on Multiple-Attention LSTM) are regression forecasters and serve only as motivation.

Research Program

The input is a nighttime CSES or DEMETER half-orbit segment. It is resampled to uniform along-track spacing and covers electron density, ion density, and electric-field spectral power. We work on segments long enough, a few hundred samples, that the along-track spectrum resolves scales from tens to thousands of kilometres. A 20-point window is too short for meaningful frequency resolution.

Each channel is tapered and transformed with a short-time Fourier transform. The taper and short windows limit the spectral leakage that sudden storm onsets cause in a single global transform. We then pool the frequency bins into $B$ bands whose edges are learned, initialized on a log scale.

A gating network produces one weight per band, $g_b = \sigma(h_\theta(\lambda_m, t_{\mathrm{LT}}, F_{10.7}, K_p))_b$. Here $\lambda_m$ is magnetic latitude, $t_{\mathrm{LT}}$ is local time, $F_{10.7}$ is solar radio flux, $K_p$ is the geomagnetic activity index, $h_\theta$ is a small network, and $\sigma$ is the logistic function. The gated spectrogram goes into a light temporal convolutional backbone that outputs a single logit $z$.

We train with class-weighted binary cross-entropy computed directly on logits:

$$\mathcal{L} = -\frac{1}{N}\sum_{i} \left[ w_+ \, y_i \log \sigma(z_i) + (1-y_i)\log(1-\sigma(z_i)) \right]$$

Here $y_i \in \{0,1\}$ marks a pre-seismic window and $w_+$ is the positive-class weight. Following Trovato et al., computing the loss from raw logits is necessary for numerical stability when operating at very low false-alarm rates (Neural network time-series classifiers for gravitational-wave searches in single-detector periods). Storm-time and solar-disturbed windows stay in the data as negatives. We fit post-hoc temperature scaling on validation data, because binary cross-entropy is known to produce overconfident outputs on unfamiliar noise.

Labels follow SeqNetQuake's convention: $M \ge$ 5-class events, shallow depth, and a pre-event window inside the Dobrovolsky radius. The CSES retrospective found most anomalies within one week before the event, with secondary peaks at 11–15 and 20–25 days (Retrospective Study on Seismic Ionospheric Anomalies Based on Five-Year Observations from CSES). We use the one-week window as primary and test the longer windows as variants. The same study reports that anomalies drift equatorward along field lines. As a secondary ablation, we compare epicentre-centred positive sampling with sampling centred on the field-line-mapped footprint. Because the drift claim comes from the same kind of retrospective analysis, we treat that ablation as exploratory and not as a reason to tune labels.

The interpretable output is the learned map $g(\lambda_m, t_{\mathrm{LT}})$. If the hypothesis is right, it should down-weight the bands where crests and bubbles concentrate near ±15–20° and after sunset. The weights near mid-latitudes should look different. That map is a falsifiable physical prediction, separate from accuracy.

Baselines

The naive reference is a $K_p$/Dst rule that labels a window seismic whenever it is geomagnetically quiet and its detrended density exceeds a 2σ envelope. This tells us how much skill comes from simply recognizing quiet ionosphere. We pair it with the class prior, which gives MCC 0.

The classical baseline is the degree-2 polynomial SVM on moving-average-detrended inputs from Asaly and Gottlieb. It reported 82.85% accuracy, 0.85 precision, and 0.80 recall on its own GIM test set of filtered quiet days (Using Support Vector Machine (SVM) with GPS Ionospheric TEC Estimations to Potentially Predict Earthquake Events). That test set is not comparable to unfiltered CSES, so we re-run the SVM on our splits.

The strongest existing method is SeqNetQuake with DEMETER-to-CSES transfer. Direct CSES training gave F1 0.6792 and MCC 0.427, and transfer learning gave a reported 12% increase in F1 and 29% increase in MCC over that (Pre-Earthquake Ionospheric Perturbation Identification Using CSES Data via Transfer Learning). We re-implement it and reproduce its in-distribution numbers first. We then run it on our segments, labels, splits, and pretraining schedule, with the same hyperparameter search budget as our model, measured in trials and GPU-hours. The SVM, the GADF ResNet, and the time-domain model with latitude input get the same budget.

Three ablations isolate the key ingredient. The first replaces the learned gate with a fixed physics-motivated band-pass, the acoustic-gravity and medium-scale band chosen once on training data, feeding the identical backbone. The second keeps learned band weights but removes conditioning, giving a single global mask in the style of SAAM and FAMNet. The third removes the frequency stage entirely and appends $\lambda_m$, $t_{\mathrm{LT}}$, $F_{10.7}$, and $K_p$ as input features to the time-domain backbone. We add a regional variant that trains separate models per latitude belt.

Beating the field means three things at once. Mean MCC on held-out latitude bands must exceed transferred SeqNetQuake and all three ablations, with bootstrap intervals over earthquakes, not windows, excluding zero. The low-latitude gap must shrink. And the fake-date control must show no gain.

Why not just…?

Why not just band-pass filter the signal and use a simple classifier? This is the most serious threat. If the seismic signal sits in a known wave band, a fixed filter could match a learned gate. The water-quality study's finding that frequency attention lowered residual entropy beyond fixed decomposition comes from a different domain and does not settle this (FECA–CEEMDAN–VMD LSTM). We run the fixed band-pass ablation with the same backbone and budget. If it matches the conditioned gate on held-out latitudes, we will report that the benefit comes from frequency selection and not from latitude conditioning, and recommend the simpler filter.

Why not just give a time-domain model the magnetic latitude as an input? This is cheaper and exploits the same information. Lei et al. show that time-domain attention alone still degrades at 15°N (Ionospheric TEC Prediction Base on Attentional BiGRU), but they did not supply latitude as an input feature. The third ablation tests exactly this. Our hypothesis predicts it interpolates within trained latitudes but extrapolates worse to a held-out belt, because it must relearn the noise suppression from scratch in each regime instead of moving a band weight.

Why not just train separate models per latitude belt? Xiong et al. effectively recommend regional retuning. With only tens of low-latitude events, however, a separate equatorial model has very few positives to learn from. We include the regional variant. If it wins, conditioning offers no sample-efficiency benefit, and the result is negative but useful.

Why not just discard disturbed days? Asaly and Gottlieb's filtering left 106 earthquakes from 23 years. An operational system cannot ignore storm periods, and SegED-ConvLSTM shows storms affect latitude bands differently, so a single global cutoff is miscalibrated. We report results both on all data and on a quiet-only subset. That shows whether our gains come from handling disturbed windows.

Proposed Experiments

We pretrain on DEMETER (2005–2010) and fine-tune and test on CSES, matching SeqNetQuake. Splits are by earthquake and by time, so no event's windows appear on both sides. The primary test is leave-one-latitude-belt-out over three belts: |λ_m| < 10°, 10–25° (the crest zone), and > 25°. Metrics are MCC and F1 at a threshold chosen on validation data, the true-positive rate at a fixed false-alarm rate per orbit-day, and Brier score with reliability curves after temperature scaling.

The decisive check is the fake-date control. We reassign every earthquake to a random date with matched season and solar-cycle phase, retrain all models on the shuffled labels, and evaluate. Any model that keeps above-chance MCC on fake dates is detecting ionospheric state, not precursors. That model's real-date skill must be discounted by its fake-date skill. We expect several baselines to fail this check.

The second decisive check is physical. Learned band weights at crest latitudes after sunset should differ significantly from mid-latitude weights in the directions the plasma-bubble and crest physics predict. A gate that changes with latitude in ways unrelated to known ionospheric physics would suggest it is fitting noise.

Finally, we run the secondary field-line-shift labelling ablation and a payload ablation. The CSES retrospective reports anomaly recognition above 70% for the plasma probe, electric field, and search-coil magnetometer payloads (Retrospective Study on Seismic Ionospheric Anomalies Based on Five-Year Observations from CSES), so we compare these payloads alone against all channels. Because the event count is small, every comparison is reported with confidence intervals computed by resampling earthquakes.