PaperlayerView exact revision on arXiv

Source-linked AI summary

Meteorology-driven Causal Nowcasting of Fugitive Landfill Emissions Enables Proactive Public Health Response

Timothy C. Pearce, David J. T. Smith, Alec Dobney, Alessia Freddo

arXiv:2608.14254v1cs.CYcs.AIcs.LGphysics.ao-phphysics.geo-ph

TL;DR

Communities near fugitive-emission sources remain exposed while monitoring and response are largely retrospective. The study uses meteorology-linked causal analysis and nowcasting to predict receptor-scale gas exposure, finding that measured temporal dynamics improve high-exposure detection and support tiered alerts validated against independent odour complaints.

  • Problem

    Public-health responses to fugitive landfill emissions are largely retrospective, despite disproportionate health burdens near odorous and toxic sources.

  • Method

    CAIRN links its temporal receptive field to causally measured transport timescales and predicts gas exposure from routine weather variables and calendar information.

  • Results

    CAIRN leads the engineered benchmark on every high-exposure metric while producing 612 versus 626 false alarms across 8,040 timesteps, with paired High-class decisions significant at McNemar p = 0.043.

  • Takeaways & Limitations

    Meteorology-driven temporal structure can reduce reliance on site-specific feature engineering and provide a validated basis for operational community-protection alerts.

  • Takeaways & Limitations

    The headline κw = 0.709 has a wide block-bootstrap confidence interval of [0.557, 0.852], so conservative operational planning should use the interval’s lower end.

Abstract

from arXiv · show

Fugitive emissions from waste sites increasingly expose communities to toxic and odorous gases, yet public-health responses remain largely retrospective, with episodes investigated only after residents have been exposed. Here we show that the meteorological drivers of elevated hydrogen sulphide (HS) at a long-monitored European landfill, and the timescales over which they act, can be identified directly from routine monitoring data. We introduce CAIRN (Causal-Anchored Inference for Receptor Nowcasting), a machine-learning framework whose internal memory is matched to these measured timescales: a fast component tracking hour-scale wind-borne transport and a slow component tracking multi-hour weather changes. Trained to predict gas measurements, CAIRN operates using only routine weather variables and the calendar, without hand-engineered features. Its behaviour is consistent with the identified transport mechanisms, and the framework transfers unchanged to a second monitoring station and to co-emitted methane. Combining four such nowcasters produces a site-level, tiered alert aligned with WHO odour guidance that closely reproduces the alert generated by a direct sensor network and tracks an independent record of community odour complaints. Weather-driven nowcasting can therefore estimate community impact as an emission episode unfolds, providing public-health authorities with a validated, graded trigger for intervention and enabling exposure to be reduced during events rather than after them.

Introduction

The paper identifies meteorological controls and timescales governing landfill H2S exposure, then uses them to anchor CAIRN, a transferable nowcasting and alert framework externally checked against community odour complaints.

  • Introduction: Retrospective responses leave communities near fugitive-emission sources exposed before impacts are documented, while waste growth and under-reported methane increase the challenge.H2S is an odorous landfill gas associated with both acute toxicity and chronic wellbeing impacts from persistent malodour.
  • Introduction: Three evidential gaps concern weather-to-exposure links, interacting timescales, and external validation against independently recorded community impact.Sensor-only validation cannot establish whether predictions track human exposure and distress; daily resident odour complaints provide such an external record.
  • Causal hierarchy of meteorological drivers: Wind direction, wind speed and atmospheric pressure form the causal meteorological core across ten scales spanning 15 minutes to 96 hours.Directional analyses identify a westerly source sector, while pressure carries directed information across scales through 12 hours.
  • Physics-anchored learnt nowcasting: CAIRN uses a dual-pathway state-space memory anchored to measured timescales, predicts gas spikes from raw meteorology and calendar variables, and transfers across receptor geometries and methane.Its leading model features include calendar hour, a 2-hour stagnation index and wind speed, with interpretable inverse-wind-speed behaviour.
  • Multi-site network tier alerting and public-health framing: Four per-channel nowcasters are fused into WHO-aligned tiers that reproduce deterministic sensor alerts and track independent daily community odour complaints at lag zero.This connects meteorology-only exposure inference to a graded operational signal tied to community impact.
  • Causal hierarchy of meteorological drivers: Boundary-layer observations show exposure varies with thermal history, diurnal mixing, sunrise, wind, temperature and temperature tendency.The mean H2S profile peaks at 02:00 LST, and sunrise reduces the pre-sunrise baseline from 9.49 µg m−3 to 1.30 µg m−3 within four hours.
  • Physics-anchored learnt nowcasting: A 6 h slow-lane anchor achieves 0.57 ± 0.01 F1-High and outperforms every tested alternative, while removing the slow-lane bound lowers performance to 0.48 ± 0.01.The 6 h anchor lies within the causally significant pressure band, although intermediate anchors were not tested and 12 h was not included.

Discussion

The study shows that meteorological causal structure can improve landfill-emission nowcasting and connect model outputs to community impact. The resulting tiered indicator supports more anticipatory, proportionate intervention, while deployment remains constrained by training-data and forecasting requirements.

  • Causal structure: Wind direction, wind speed and atmospheric pressure formed the dominant meteorological information pathways to receptor H2S across multiple temporal scales.The hierarchy was reproduced across estimator configurations and disjoint half-years; pressure's directed-information share peaked at 12 hours.
  • Model design: CAIRN's causally anchored temporal design outperformed tested alternatives and the engineered benchmark on every high-exposure metric while producing 612 versus 626 false alarms across 8,040 timesteps.Releasing the physically plausible slow-pathway bound reduced F1-High by 0.04 on 10 of 13 folds.
  • External validation: The meteorology-only model tracked independent daily community odour complaints at zero lag, with H2S exposure probability correlating at r = 0.585 across 88 days.The complaint record was withheld from model development, providing an external benchmark for a human-reported endpoint.
  • Operational alerting: The fused tier correlated with daily complaints at r = 0.721 under leave-one-week-out refitting, compared with r = 0.793 for the deterministic raw-sensor reference.The reported paired difference was Δr = +0.072, with a 95% CI of −0.303 to +0.127.
  • Operational alerting: Bayesian fusion of four channels produced a tiered exposure indicator that agreed substantially with sensor-derived classifications, with weighted κ = 0.709 and 0.703 under leave-one-week-out refitting.The ordinal scale was aligned with the FIDOL/WHO odour-annoyance framework.
  • Implications: A real-time, tiered exposure signal could shift response from retrospective complaint-driven action toward anticipatory and proportionate public-health and regulatory intervention.The identified meteorological conditions also provide a basis for targeting oversight and scheduling relevant operational activities away from higher-risk periods.

Methods

The study combines multiscale causal analysis with a causally constrained CAIRN architecture and expanding-window validation to nowcast H2S from routine meteorology and calendar variables.

  • Data and causal analysis: 34,609, 23,104, 28,331 and 11,374 valid H2S observations were collected at MMF9, MMF1, MMF2 and MMF1A, respectively, during 2024.Sampling was performed at 15-minute intervals against a complete grid of 35,136 intervals.
  • Data and causal analysis: Multiscale transfer entropy tested directed nonlinear information flow from candidate meteorological drivers to H2S across ten scales from 15 minutes to 96 hours.The analysis used coarse-graining, conditional estimation, surrogate testing, effective transfer entropy, and false-discovery-rate correction across 70 tests.
  • Nowcaster design: Four quarterly XGBoost classifiers predicted three H2S classes defined by WHO guidance: Low < 2, Medium 2–7, and High ≥7 µg m−3.The models used expanding-window training rooted in September 2023.
  • Nowcaster design: CAIRN uses a minimal sixteen-dimensional input and learns meteorological memory without engineered features, while preserving strict one-sided causality throughout the model.Inputs include raw weather channels, angular and calendar encodings, and seasonal harmonics.
  • Nowcaster design: The architecture splits channels into fast and slow lanes, with |A_fast| = 10.0 and |A_slow| = 0.5 to represent sub-hour-to-few-hour and half-day-to-multi-day timescales.The lanes use a 50/50 channel split and physics-anchored timescale initialization.
  • Validation: CAIRN evaluation used 13 weekly expanding-window blocks with a 96-hour causal context reserve, while the engineered benchmark trained to the week boundary without that reserve.The production configuration reproduced F1-High within ±0.006 across three independent seeds, with pooled within-condition standard deviation 0.013.

Cross-species CH4 nowcaster and complaint validation

The study extends CAIRN to co-emitted methane and combines four nowcasters into a meteorology-only alert pipeline evaluated against sensor-derived tiers and independent community complaints.

  • Cross-species transfer: A second architecturally identical CAIRN model was trained on co-emitted CH4 under the weekly walk-forward protocol.Its reported values retain the original checkpoint-selection rule and are optimistic to the extent quantified in walk-forward validation.
  • Cross-species transfer: CH4 alert classes were derived from H2S thresholds using the regression H2S = 5.55 CH4 − 8.29 fitted on the monitoring record.
  • Alert-tier classifier: Four independently trained H2S and CH4 nowcasters at MMF9 and MMF2 were fused into a continuous network-event probability and four ordinal alert tiers.The predicted pipeline uses meteorology at inference time and does not read raw target-species concentrations.
  • Validation against sensors and complaints: The alert pipeline was compared with a deterministic raw-concentration ground-truth tier using quadratic-weighted Cohen’s kappa on 8,536 synchronous 15-minute timesteps.Confidence intervals used 1,000 24-hour block-bootstrap resamples, and an independent daily complaint record served as a held-out endpoint.
  • Reproducibility: The reference implementation releases the CAIRN model, checkpoints, standardisation parameters, feature construction, labelling code, and acceptance tests.The complaint record is withheld as personal data, so complaint-based analyses cannot be regenerated from the public release.

Data availability

The public release includes the principal-receptor meteorological and gas record, while community complaint data remain unavailable because they are personal regulatory-process data.

  • Released data: The released Parquet file covers 1 September 2023 to 31 December 2024 at the principal receptor and includes weather, H2S and CH4 measurements.The variables are wind direction, wind speed, air temperature, barometric pressure, H2S and CH4.
  • Unavailable data: Community odour-complaint records are not released because they are personal data collected under a regulatory complaints process.Derived daily counts are available from the corresponding author on reasonable request, subject to information governance.
  • Site identification: The monitoring site is identified only by station code.

Funding acknowledgement

The study was part-funded by the NIHR Health Protection Research Unit in Chemical Threats and Hazards, with views attributed to the authors.

  • Funding: The study is part-funded by the NIHR Health Protection Research Unit in Chemical Threats and Hazards under grant award NIHR207293.The authors state that the views expressed are theirs and not necessarily those of NIHR or the Department of Health and Social Care.

SUPPLEMENTARY INFORMATION

The Supplementary Information contains extended quantitative tables, procedural notes, methods, and supporting numerical and graphical detail for the manuscript.

  • Supplementary Information: Extended Data tables contain peer-reviewed quantitative results referenced by the main figures and Results text.
  • Supplementary Information: Supplementary Notes describe multiscale transfer entropy, accumulated local effects, and the S4D state-space model.
  • Supplementary Information: Supplementary Methods specify the optimiser schedule and walk-forward validation procedure, including leakage safeguards and attribution analyses.

Extended Data

The Extended Data materials document the physics-anchored ablation, walk-forward model comparison, and meteorology-only multi-site alert-tier evaluation, with calibration-related metrics reported separately.

  • Extended Data: Table ED1 isolates the architectural contribution of physics-anchored timescale priors while holding the model, training, input, validation, and hyperparameter settings constant.
  • Extended Data: Table ED2 reports period-wide walk-forward operational performance for the dual-pathway S4 model versus an engineered XGBoost benchmark under matched stopping-point selection.
  • Extended Data: The S4 model leads on every High-class detection metric, while probabilistic metrics differ between models after evaluation-block checkpoint selection is removed.The ranked probability score and log-loss favour the benchmark, whereas Brier-High marginally favours the S4 model.
  • Extended Data: Table ED3 evaluates a meteorology-only multi-site Bayesian alert-tier classifier on a synchronous Jan–Mar 2025 walk-forward window.The predicted tier excludes raw target-species concentration at inference time.

S1 Regulatory and public-health context

The regulatory and public-health context frames landfill H2S as a community exposure and odour problem managed largely through retrospective assessment, despite available guidance and dispersion models.

  • S1 Regulatory and public-health context: 3.40 billion tonnes is the projected global municipal solid-waste generation by 2050, while engineered landfills continue releasing methane, carbon dioxide, water vapour, and co-emitted gases.
  • S1 Regulatory and public-health context: H2S is produced anaerobically by sulphate-reducing bacteria and has an odour threshold below concentrations of toxicological concern.The relevant sulphate source in mixed waste streams is identified as gypsum from construction and demolition debris.
  • S1 Regulatory and public-health context: 7 µg m−3 is the WHO 30-minute odour-annoyance guideline, whereas acute high-concentration H2S exposure can cause severe toxicity and death.The passage distinguishes the community-wellbeing benchmark from acute-exposure thresholds by several orders of magnitude.
  • S1 Regulatory and public-health context: Persistent odorous emissions are associated with annoyance, insomnia, headaches, sensory irritation, nausea, depression, and respiratory symptoms.
  • S1 Regulatory and public-health context: UK statutory-nuisance assessment relies on local-authority judgement and witnessed evidence, which can constrain responses to intermittent odours and sensitive receptors.
  • S1 Regulatory and public-health context: The source–pathway–receptor framework places atmospheric dispersion between landfill emissions and residential populations, but communication and decisions occur after exposure.Forward-looking dispersion modelling exists but relies on simplifying assumptions about atmospheric conditions and emissions.

S2 Extended environmental characterisation

Environmental characterisation links elevated H2S at the principal receptor to a westerly source direction and identifies a common emitting region across monitoring stations.

  • S2 Extended environmental characterisation: 34,609 valid 15-minute H2S observations from 2024 underpin the principal-receptor analysis at MMF9.The complete grid contained 35,136 observations.
  • S2 Extended environmental characterisation: CPF ≈0.37 peaks at a 265° westerly bearing for the MMF9 receptor using a 5 µg m−3 exceedance threshold.The threshold lies between the odour-detection threshold and the WHO 30-minute guideline value.
  • S2 Extended environmental characterisation: 11.2 µg m−3 in the W sector and 10.6 µg m−3 in the NW sector exceed 1.1 µg m−3 in the opposing E sector.The peak/opposing-sector relative-strength ratio is 4.09×.
  • S2 Extended environmental characterisation: r = 0.149 links wind direction with H2S concentration, with p < 10−3 across the 34,609-observation record.
  • S2 Extended environmental characterisation: CPF cones from MMF9, MMF1, and MMF2 converge on a common source area, providing spatial triangulation of the emitting region.

Chemical fingerprint: H2S–CH4 co-emission (Fig. 1b)

H2S and CH4 show strong co-emission and synchronized spike behavior, while particulate controls and phase offsets argue against shared non-gas explanations. H2S varies sharply with meteorological conditions, especially wind, time of day, sunrise, and season.

  • Co-emission: r = 0.830 Pearson correlation links simultaneously measured H2S and CH4.Spearman correlation was ρ = 0.683, with both p < 10−15.
  • Co-emission: J = 0.654 spike co-occurrence and 1,386 observed events versus 87.7 expected support synchronized H2S–CH4 excursions.The excess was significant at χ2 = 21,298, p < 10−15.
  • Negative controls: PM10 and PM2.5 show no positive association with H2S, providing negative controls against shared particulate resuspension.PM10 had r = −0.028 and PM2.5 had r = −0.009, with PM2.5 not significant.
  • Negative controls: NOx correlates moderately with H2S, but a six-hour peak offset supports joint nocturnal trapping rather than a common emission pathway.NOx peaks near 08:00 LST, whereas H2S peaks near 02:00 LST.
  • Meteorological modulation: H2S concentration is strongly diurnal and seasonal, peaking at 02:00 LST and in February.The mean peak/trough ratio was 7.4×, while the February-to-August monthly ratio was 15.1×.
  • Meteorological modulation: Low wind produces substantially higher H2S than high wind, with 11.8 µg m−3 below 1 m s−1 versus 1.8 µg m−3 above 5 m s−1.Spike samples also had lower wind speeds than non-spike samples: 1.75 versus 3.67 m s−1.

S3 Per-season interpretation of the XGBoost ALE feature hierarchy

Seasonal XGBoost attributions vary with atmospheric regime: diurnal timing leads in winter, spring, and autumn, whereas wind speed leads in summer. Agreement with the causal hierarchy is therefore partial, not a full ranking correspondence.

  • Winter: Winter has the flattest feature hierarchy, with diurnal timing leading at Ī = 0.037 and no feature exceeding Ī = 0.04.Wind direction follows, while stagnation and temperature tendency contribute little.
  • Spring: Spring assigns nearly equal importance to diurnal timing and two-hour stagnation, at Ī = 0.146 and 0.130 respectively.This pattern is associated with stagnation as a departure from generally well-ventilated spring conditions.
  • Summer: Summer is the only season led by a measured meteorological variable: wind speed at Ī = 0.058, ahead of diurnal timing and stagnation.The attribution is consistent with convective mixing making mechanical dilution more discriminating.
  • Autumn: Autumn has the largest single attribution, with diurnal timing at Ī = 0.333, roughly an order of magnitude above two-hour stagnation.Shortening daylight and longer nocturnal stable-boundary-layer periods make time of day a stability proxy.
  • Cross-seasonal pattern: Wind speed and wind direction appear in the leading group every season, while raw pressure has zero ALE importance because its synoptic content is represented by stagnation.Diurnal timing leads in winter, spring, and autumn; wind speed leads in summer.
  • Methodological interpretation: The ALE and transfer-entropy rankings do not fully mirror one another because calendar terms are absent from the causal driver set and the estimators credit different information.Among shared features, ALE ranks wind speed first and WD sin second; raw pressure is the genuine discrepancy.

S4 Extended comparison with prior literature

The paper positions its contribution as a bridge between causal meteorological analysis, structured state-space modeling, and external community-impact validation. Its evidence supports a site-specific, physics-anchored nowcasting framework rather than a universally transferable hierarchy.

  • Architectural contribution: The dual-pathway MSTE-anchored S4 initializes memory on causally significant horizons rather than using data-agnostic timescales.A six-condition ablation found the optimum among tested anchors at 6, 48, and 147 h, with q ≤ 0.048.
  • Architectural contribution: The tested slow-anchor optimum outperformed alternative initializations, but anchors between 6 and 48 h were not evaluated.The optimum is therefore identified within the tested set, not precisely located throughout the causal band.
  • Representation: The model reproduces marginal dP/dt and dT/dt response curves without using those derivatives as input features.The weak lane-resolved association, ρs = +0.14, supports consistency with a joint synoptic signature rather than pathway-specific encoding.
  • Community validation: A meteorology-only nowcaster tracks an independent community-wellbeing endpoint at 15-minute resolution across a full quarter, with r = 0.585 at lag zero.The record included n = 88 days.
  • Operational implication: The operational target is the lower 30-minute WHO odour-annoyance guideline, rather than the regulatory acute-exposure threshold.The paper links this target to substantial community impact at the case-study site.
  • Scope boundary: Operational deployment requires independently auditable monitoring and re-estimation of the causal hierarchy at each new site.The study uses only post-calibration-adjusted records and treats the hierarchy as regime-specific.

S5 MSTE estimator ablation study

The MSTE ablation finds that the core wind-direction, wind-speed, and pressure relationships are robust to estimator settings, while embedding depth and confounder conditioning materially affect fine-scale significance.

  • Robustness: Wind direction, wind speed, and atmospheric pressure remain significant across all six configurations at scales ≤3 h.Temperature is robustly significant only at 6 h, except under the baseline configuration.
  • Estimator sensitivity: Bivariate transfer entropy inflates effect sizes and false positives by failing to condition on confounders.For dP/dt, ETE reaches 0.6–19.5 × 10−3 nats at all scales bivariately but appears only at ≥2 h under conditional configurations.
  • Estimator sensitivity: History embedding depth is the dominant sensitivity axis, with h ≤2 at fine scales eliminating spurious temperature significance while preserving genuine links.Larger embeddings increase nearest-neighbour sparsity and estimation difficulty.
  • Production configuration: Configuration 4 was selected for production because it balances target-history context, confounder isolation, and a low conditioning dimension.It uses h = 2 at fine scales and maintains d ≤3 for reliable KSG estimation.
  • Sample sensitivity: The split-half analysis preserves the ordering WDsin > Pressure > WS in both halves and the full-year analysis.Significance is retained by 20 of 21 cells in the first half and 18 of 21 in the second.
  • Sample sensitivity: Wind-speed estimates disagree in sign at 6 h and 12 h, making its coarse-scale signal the least robust part of the hierarchy.Both second-half estimates are consistent with zero, although wind speed still ranks third in both halves.
  • Limitations: Split halves differ seasonally and cannot detect bias applied uniformly by the estimator.The analysis can confirm the hierarchy under this confounding but cannot fully refute it.

S6 Joint site analysis

The joint alert system converts meteorology-only channel predictions into tiered site alerts and validates them against raw-sensor tiers and independent community complaints. Its agreement is substantial overall, but intermediate-tier discrimination and parameter derivation remain important limitations.

  • Calibration caveats: The deployed cut-points (θ1, θ2, θ3) = (0.15, 0.50, 0.92) are operational choices not reproduced by the documented grid search.A leave-one-week-out refit selected θ1 = 0.20 in all folds and changed κw by 0.013.
  • Calibration caveats: The fusion assumes conditional independence despite shared sources and meteorology across channels, while higher-order joint-state treatment is left for future work.The present likelihood-ratio values are deliberately conservative.
  • Validation: Δκw = +0.070 versus the vote-count baseline, with a paired block-bootstrap 95% CI of 0.025–0.109 and one-sided p = 0.003.The Bayesian fusion also provides a calibrated continuous posterior and handles missing data and CH4-quiet neutrality.
  • External validation: Daily mean predicted tiers correlate with complaint counts at Pearson r = 0.729, R2 = 0.53, and lag 0.This captures 84% of the ground truth’s explained variance and approaches the raw-sensor reference of r = 0.793.

Supplementary Methods

The supplementary methods specify environmental diagnostics, causal analysis, model architecture, validation, and reproducibility procedures. They connect meteorological patterns to exposure mechanisms while documenting data gaps, estimator settings, and non-separable lane contributions.

  • Environmental characterisation: CPF assigns each 15-minute observation to one of 36 ten-degree wind sectors and computes exceedance probability from sector-specific H2S counts.The exceedance threshold is Cthr = 5 µg m−3.
  • Environmental characterisation: Diurnal temperature–H2S trajectories quantify hysteresis using enclosed area and separate cooling- and warming-limb slopes.Normalised areas above 0.30 are classified as strong hysteresis.
  • Environmental characterisation: Hourly means and 95th percentiles describe diurnal concentration variation, while sunrise-aligned profiles test pre-sunrise versus post-sunrise concentrations.The sunrise analysis uses a one-sided Mann–Whitney U test and predefined evidence categories.
  • Environmental characterisation: Meteorological associations with H2S are evaluated using Pearson correlation and OLS regression for wind speed, temperature, and temperature tendency.All three associations are predicted to be negative because of dilution and boundary-layer destabilisation mechanisms.
  • Multiscale causal analysis: Multiscale transfer entropy uses non-overlapping block averages with right-closed, right-labelled blocks to prevent leakage across aggregation boundaries.Coarse blocks at τ ≥ 3 h are aligned to 18:00 UTC to centre nocturnal accumulation within a block.
  • Data and reproducibility: Missing-data handling affects coarse-scale estimates: the 96 h scale is not estimable, and 31.0% of 6 h blocks are affected by reconstruction.Scales τ ≥24 h are sample-limited and should be interpreted descriptively.
  • S4D state-space model: The S4D architecture uses pre-normalisation, gated linear units, and zero-initialised skips to stabilise deep residual processing and preserve slow-lane information.The prediction representation is the final-timestep slice, and the kernel is causal rather than bidirectional or future-padded.
  • Walk-forward validation: Across 13 weeks, the lane ablation does not separately identify fast and slow contributions because their ordering reverses with the masking formulation.The interaction exceeds either marginal contribution, so the decomposition does not establish mechanism-specific skill attribution.

Supplementary Tables

The supplementary tables document datasets, evaluation protocols, model comparisons, transfer-entropy analyses, and limitations underlying the paper’s results. They report seasonal and weekly performance, memory ablations, statistical testing, and estimator sensitivity.

  • Environmental characterisation: 34,609 valid 15-minute H2S records characterize receptor-level exposure at MMF9 across 2024.Table S9 defines the principal receptor and record count for the environmental characterization.
  • Seasonal performance: The seasonal XGBoost evaluation reports per-class precision, recall, and F1, with High-class prevalence ranging from 5.2% in Summer to 12.8% in Winter.Accuracy summarizes overall quarterly performance, while class counts are reported separately.
  • Limitations and scope: Six samples after a seven-hour meteorological outage are missing under the causal derivative formulation, but all are Low-class and leave Medium and High counts unchanged.Other reported boundaries include sensitivity to fold composition and the absence of a state-space model on engineered features.
  • Transfer-entropy analysis: The transfer-entropy supplement compares six estimator configurations across 42 driver–scale combinations, with production significance based on 2,000 circular-shift surrogates.The B = 10 screen is explicitly coarse and not false-discovery-corrected.
  • Memory decomposition: Across 13 weekly folds, the matched memory ablation gives ∆engineering = +0.040 and ∆architecture = +0.073 in pooled F1-High over XGB-16.The endpoint difference is CAIRN − XGB-24 = +0.032, and both increments are reported as primary comparisons despite not clearing the two-sided sign test.

Supplementary Figures

Figure S1 shows co-located 15-minute H2S and CH4 measurements at the principal receptor during autumn 2024 validation. H2S thresholds and linearly transformed CH4 class boundaries establish the basis for the cross-species transfer analysis.

  • Figure S1: The top series plots H2S concentration against WHO Medium and High thresholds of 2 and 7 µg m−3.These thresholds define the displayed H2S alert classes.
  • Figure S1: The bottom series plots simultaneous CH4 concentration using the corrected relation H2S = 5.55 CH4 −8.29.The corresponding CH4 classes are Low < 1.854, Medium 1.854–2.756, and High ≥2.756 ppm.
  • Figure S1: The concurrent H2S–CH4 coupling motivates the cross-species transfer experiment reported in main Fig. 6.The figure caption links this experiment to reliability, complaint cross-correlation, and cross-model agreement analyses.
Loading 2608.14254v1…