Methods
Gaps, closed in pairs.
Existing air-quality systems share known gaps: single-method estimates with hidden uncertainty, no independent validation, unmonitored data quality, one-overpass blindness, reactive alerts. Each gap here is closed by a pair of peer-reviewed techniques that cover each other’s failure modes.
The core claim
Gap-cover matrix
| Failure mode | Covered by |
|---|---|
| Sparse stations | Satellite AOD-to-PM regression plus kriging spatial structure |
| Satellite cloud gaps (monsoon) | Ground kriging plus meteorology features and compositing; GEMS hourly when keyed |
| One overpass per day (low-earth-orbit) | GEMS geostationary diurnal (keyed) plus diurnal ML priors |
| Model bias drift | Independent embassy-network validation plus calibration receipts |
| Bad station data | Station trust scores down-weight fusion inputs, weights on record |
| Point estimates hide risk | Quantiles plus per-ward exceedance probabilities |
Method stack
How each layer works
- 01
Ensemble gap-fill
Four independent estimators (IDW, ordinary kriging, satellite AOD-to-PM, LightGBM) combined by stacked generalization (Wolpert) trained only on leave-one-station-out folds. Published per cell: p50 and p90, per-method weights, and a method-disagreement index as honest epistemic uncertainty.
- 02
Weather normalization
Grange and Carslaw meteorological normalization: resample the meteorology and predict a weather-neutral PM2.5 series, so a drop is the policy and not the rain.
- 03
Causal effect estimation
Event-study around each GRAP transition on the normalized series, with placebo tests on matched high-pollution non-intervention days and block-bootstrap confidence intervals.
- 04
Trajectory attribution
Kinematic 48h back-trajectories and concentration-weighted trajectory fields (Hsu 2003), overlapped with FIRMS fires and Sentinel-5P NO2 columns. CPF gives direction; CWT gives geography.
- 05
Health translation
GEMM exposure-response (Burnett 2018) with WorldPop 100m population, aggregated per ward to avoided exposure and avoided premature deaths, always labeled a modeled estimate.
Formulas
Key equations
CPCB sub-index (linear interpolation)
Ip = (IHi - ILo) / (BPHi - BPLo) * (Cp - BPLo) + ILo
Inverse-distance weighting baseline
c(x) = sum(wi * ci) / sum(wi), wi = 1 / d(x, xi)^p
Weather normalization
c_norm(t) = mean over resampled meteorology of f(meteo*, calendar(t))
GEMM attributable mortality
deaths = population * baseline_rate * (1 - 1 / RR(exposure))
Reuse
Downloadable data products
Researchers can reuse the outputs. Published products, with lineage on the receipts page:
- Hourly nowcast grid + ward rollupsJSON now; Parquet and GeoTIFF from the publish pipeline
- Ward forecast (24 to 72h, p50/p90)JSON
- Intervention Ledger (effects + mortality)JSON
- Validation receiptsJSON
On the record
Peer-reviewed methods
- van Donkelaar et al., satellite AOD to surface PM2.5 (two-stage)
- Wolpert 1992, stacked generalization
- Hsu et al. 2003, concentration-weighted trajectory
- Wackernagel, ordinary kriging (multivariate geostatistics)