Satellites estimate how much haze, dust, or smoke is in the air by measuring sunlight reflected back to space. Over bright ground such as deserts, cities, and bare soil, most of that light comes from the ground rather than the air. A thin aerosol layer then becomes a small excess on top of a large, poorly known surface background, and current retrievals either get it wrong or discard the pixel. Rare-event physics faces the same situation in a more extreme form. Dark matter and neutrino experiments look for a few genuine events buried under backgrounds that are orders of magnitude larger. Their working method has three parts. They measure the background directly on data known to contain no signal. They report how often background "leaks" past a cut at a given signal acceptance. They combine independent detectors only after checking that the detectors' backgrounds are actually independent.
This note proposes carrying that method into aerosol retrieval. A polarimeter on a polar-orbiting satellite and an imager on a geostationary satellite each give a weak, noisy hint of aerosol over the same bright pixel. We would combine these hints as a calibrated likelihood ratio, the way gravitational-wave and neutrino triggers are combined. The surface-induced background of each instrument would be measured on clean, ground-verified scenes, not assumed. The central claim is that the gain from combining sensors is set by how correlated their surface errors are. Two instruments that both lean on the same surface database will not help each other much. Two instruments that reject the surface by different physics can turn individually ambiguous detections into reliable ones.
Core Claim
Over bright land at low aerosol optical depth (AOD, the column-integrated haze extinction), each existing retrieval's error is dominated by surface misestimation. The rare-event literature shows that low signal-to-background problems are solved by discriminants whose background response is measured, not modelled.
Two examples show the pattern. In α-event characterization and rejection in point-contact HPGe detectors, a slow trailing-edge feature rejects surface alpha events with only 0.2% loss of bulk signal. The rejection was established by scanning a source across known positions, not by trusting a field model. In First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB, electronic-recoil leakage below $1.2\times10^{-7}$ is quoted as a measured number from calibration data.
Enhancing the significance of astrophysical events with multimessenger coincidences supplies the combination step. It updates the probability that a weak gravitational-wave trigger is real using a coincident neutrino, raising one candidate from 7.7% to 86–97%. That work relies on the two detectors having physically unrelated backgrounds.
Our claim is that the same structure works for aerosol detection over bright land, provided the joint surface background of the two sensors is estimated empirically. For each pixel we form residual statistics $x = (x_1, x_2)$, one per sensor: the departure of each sensor's retrieval from what its clean-scene surface model predicts. We then compute
$$\lambda(x) = \log \frac{p(x \mid \text{AOD} \ge \tau)}{p_B(x \mid s)},$$
where $\tau$ is a target low-AOD threshold (for example 0.1). The denominator $p_B(x \mid s)$ is the joint background density of the two residuals for surface class $s$. It is measured on scenes where ground photometers confirm near-zero aerosol.
If $p_B$ factorizes across sensors, background leakage multiplies down and coincidence is powerful. If the two residuals share a surface error, the off-diagonal structure of $p_B$ erodes that gain. Measuring this correlation, and selecting sensor pairs that minimize it, is what the proposal adds.
Prior Work and Where This Sits
The remote-sensing papers supply the per-sensor statistics and also motivate the gap.
Aerosol Retrieval Study from a Particulate Observing Scanning Polarimeter Onboard Gao-Fen 5B without Prior Surface Knowledge, Based on the Optimal Estimation Method separates the surface using polarized 2250 nm light, which it assumes is barely touched by aerosol, and then solves for aerosol by optimal estimation. Its surface rejection is therefore physically independent of external surface databases. It still struggles at the bright Baotou site (R = 0.718).
A Simple Band Ratio Library (BRL) Algorithm for Retrieval of Hourly Aerosol Optical Depth Using FY-4A AGRI Geostationary Satellite Data uses MODIS-derived red/shortwave-infrared surface ratios at 15-minute cadence. It reaches R = 0.84 and RMSE = 0.16, with about 60% of retrievals inside the expected-error envelope.
Aerosol Retrieval Method Using Multi-Angle Data from GF-5 02 DPC over the Jing–Jin–Ji Region refines AOD by minimizing variance across viewing angles. It explicitly underperforms at low AOD because errors in its MODIS MOD09 surface input dominate. This is the clearest evidence that a shared surface database is the bottleneck.
Evaluation and comparison of MODIS Collection 6.1 aerosol optical depth against AERONET over regions in China with multifarious underlying surfaces adds two further observations: only 54% of Dark Target retrievals fall inside the expected-error envelope, and stricter quality flags remove coastal and wetland coverage entirely.
None of these papers measures how one retrieval's surface error co-varies with another's. None of them states its performance as background leakage at fixed signal acceptance.
The physics papers supply the method. The multimessenger framework provides the likelihood-ratio combination. The germanium and argon papers provide the discipline of measuring background on signal-free calibration data and quoting leakage.
We do not reuse the multimessenger paper's astrophysical emission priors, which that paper itself flags as a source of model dependence. Our signal model is instead built from ground-verified polluted scenes and from simulated aerosol injected onto clean-scene radiances.
Research Program
The study region is northern and central China, where GF-5B POSP (November 2021–July 2022), FY-4A AGRI, and Himawari-8 overlap. Ground truth comes from AERONET sun photometers, the standard ground network for aerosol validation. The sites are Beijing, Beijing-RADI, Hefei, and Baotou, plus any further AERONET Level 2.0 sites with collocations. Pixels are grouped into surface classes (urban, arid/bare, vegetated, mixed coastal) using MODIS land products.
The first stage builds the background. We collect "clean" scenes where AERONET AOD falls below 0.05. On these scenes we compute each sensor's residual and estimate the joint density $p_B(x \mid s)$ for each surface class, using a kernel or Gaussian-mixture fit. We report the cross-sensor correlation of residuals as a primary scientific result. This is the analogue of the source-scan calibration in the germanium work.
The second stage builds the signal response. We take polluted scenes with AERONET AOD in the 0.1–0.4 range, which is the regime the multi-angle paper flags as weak. We supplement them with synthetic injections: aerosol layers simulated with UNL-VRTM, the vector radiative transfer code already used by the POSP retrieval, added onto clean-scene radiances. This maps signal acceptance as a function of AOD and aerosol type. We treat the injections only as a cross-check, because surface–atmosphere coupling makes additive injection approximate.
The third stage adds coincidence structure. The geostationary sequence brackets each polar overpass by about ±1 hour. Aerosol tends to persist and advect smoothly across this window, while surface-error residuals are fixed in place, so temporal persistence enters the likelihood as an additional independent factor. Its own background is again measured on clean scenes.
The final output for each pixel is a posterior probability that AOD is at least $\tau$, together with a posterior-mean AOD. Pixels that current quality flags discard are kept and scored rather than dropped.
Baselines
The naive reference point is each sensor's existing retrieval used alone. For POSP this is the optimal-estimation product, for AGRI the BRL product, and for DPC the multi-angle variance product. Each is converted to a detection by thresholding its AOD at $\tau$. The paper-reported figures we must at least match on estimation are these:
- POSP reaches R = 0.944 and RMSE = 0.025 for AOD < 0.7, but that comparison is against MODIS, not AERONET.
- At individual AERONET sites, POSP reaches R = 0.982 (Beijing) and R = 0.718 (Baotou).
- BRL reaches R = 0.84 and RMSE = 0.16.
- DPC reaches R² > 0.9 and RMSE < 0.1 at Beijing-RADI.
For the detection metric, background leakage at fixed signal acceptance over bright land, we found no reported result on Lacuna for any of these sensors. We will compute it ourselves for every baseline on identical collocations.
The strongest learned competitor is the deep network of Himawari-8 Aerosol Optical Depth (AOD) Retrieval Using a Deep Neural Network Trained Using AERONET Observations. It maps top-of-atmosphere reflectances, physically motivated band ratios, geometry, and auxiliary data to AOD. It reports R² = 0.730 and RMSE = 0.172 under leave-one-station-out validation, and R² = 0.915 under random k-fold.
We will retrain this architecture on exactly our fused inputs, which are both sensors' residuals and the temporal features. It gets the same collocations and the same hyperparameter search budget as our density fits. It is evaluated only by leave-one-station-out, since that is where surface transfer is tested.
Beating it means two things. First, lower leakage at matched acceptance on held-out bright stations, especially Baotou. Second, better-calibrated probabilities, measured by expected calibration error, with no loss in low-AOD RMSE.
The key ablation replaces the measured joint background with a factorized one. In that version each sensor's background is measured separately and the two are treated as independent, which is the direct multimessenger-style product. A second ablation pairs DPC with AGRI. Both depend on MODIS surface information, so we expect their surface errors to be correlated. We compare this pair against POSP with AGRI, whose surface rejection mechanisms differ.
Why not just…?
Why not just average or inverse-variance-weight the two sensors' AOD? Weighted averaging assumes errors are independent and roughly Gaussian. Our hypothesis is that over bright land the errors share a surface component, and that this component is heavy-tailed at the low-AOD end, where the DPC paper shows MOD09 errors dominate. If our measured cross-sensor correlation turns out near zero and residuals look Gaussian, then inverse-variance weighting is essentially optimal and the proposal's advantage collapses. We will report that outcome as a finding, not hide it.
Why not just train a neural network on both sensors? This is the real threat. A flexible network could learn the joint background implicitly. Against it, the Himawari network's drop from R² = 0.915 to 0.730 when whole stations are held out suggests that learned models absorb station-specific surface behaviour. Our density is instead conditioned on surface class and measured on clean scenes. The matched retraining described above under leave-one-station-out settles the question. If the network wins on both leakage and calibration, we would conclude that its value lies in feature design and adopt it.
Why not just relax the quality flags? Relaxing thresholds recovers coverage, but the extra pixels it admits have unknown leakage. The MODIS evaluation documents the coverage loss but never quantifies what relaxing the flags would cost in false detections. We will sweep each baseline's flag threshold to trace its leakage–acceptance curve and compare whole curves, not single operating points.
Why not just wait for hyperspectral oxygen A-band sounders? The A-band information-content study is simulation-only. The broad-swath imagers flying now lack 0.01 nm channels, so this route does not help existing records.
Proposed Experiments
The first decisive check is the measured residual correlation between sensors per surface class. The proposal predicts it is substantially higher for the DPC–AGRI pair than for the POSP–AGRI pair. If the two are indistinguishable, the physical-independence argument fails.
The second check compares the full method with the factorized ablation on held-out stations. Background leakage at 80% acceptance for AOD ≥ 0.1 should be lower when the joint background is measured. If the two match, correlation does not matter in practice.
The third check compares the full method with the retrained network and with inverse-variance fusion, using leakage–acceptance curves and calibration error under leave-one-station-out. Baotou, the site where POSP alone reaches only R = 0.718, is the primary test.
The fourth check counts pixels that current quality flags discard but that receive high posterior scores. We verify these against AERONET where it is available and against geostationary temporal persistence elsewhere. This tests whether coincidence recovers real low-signal aerosol rather than surface artifacts.
References
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB
- Enhancing the significance of astrophysical events with multimessenger coincidences
- Aerosol Retrieval Study from a Particulate Observing Scanning Polarimeter Onboard Gao-Fen 5B without Prior Surface Knowledge, Based on the Optimal Estimation Method
- A Simple Band Ratio Library (BRL) Algorithm for Retrieval of Hourly Aerosol Optical Depth Using FY-4A AGRI Geostationary Satellite Data
- Aerosol Retrieval Method Using Multi-Angle Data from GF-5 02 DPC over the Jing–Jin–Ji Region
- Evaluation and comparison of MODIS Collection 6.1 aerosol optical depth against AERONET over regions in China with multifarious underlying surfaces
- Himawari-8 Aerosol Optical Depth (AOD) Retrieval Using a Deep Neural Network Trained Using AERONET Observations