Radio links between rovers, landers and orbiters on Mars can double as weather instruments. Dust in the path weakens the signal, so the measured loss on each link reveals how much dust lies between the two ends. The existing design of this idea (Martian Dust Storm Detection With THz Opportunistic ISAC) hits two walls. The first is a physics problem: the conversion from signal loss to dust amount depends on how large the dust grains are, and the authors identify this particle-size uncertainty as "the primary contributor to estimation errors." The second is a networking problem that the paper does not model at all. It assumes attenuation readings arrive continuously. Real links fade, time out and drop packets, and they drop out most often exactly when the dust is thickest. A system that quietly ignores lost readings will underestimate storms in a systematic way, and collecting more data will not fix that.
This note proposes to treat both walls with the same tool that precision physics uses when counting statistics stop helping and residual systematic uncertainties dominate. Each unknown bias is written as an explicit nuisance parameter. Where possible, it is cancelled with ratios of measurements that share it. Whatever remains is carried into honest confidence intervals. In plain terms, the plan has three parts. Several carrier frequencies on the same link are compared, so the per-link hardware offsets and the dust amount drop out and the grain size can be inferred in place. Link outages are counted as information: an outage says "the loss exceeded this link's budget," not "no data." The system then reports frequentist error bars whose stated coverage can be checked. The central claim is that, once outages are informative, getting the missing-data model right matters as much for accuracy as getting the dust physics right. A pipeline that handles only one of the two will be biased and overconfident during storms.
Core Claim
Consider link $\ell$ at carrier frequency $f$. Its excess attenuation relative to clear sky can be written as
$$A_{\ell f} = g_{\ell f} + L_\ell\,\kappa_f(r)\,M_\ell + \epsilon_{\ell f}.$$
The terms are defined as follows. $g_{\ell f}$ is an unknown hardware or pointing offset. $L_\ell$ is the known path length. $\kappa_f(r)$ is the dust extinction per unit mass at frequency $f$, which depends on the effective grain radius $r$. $M_\ell$ is the path-averaged dust mass loading. $\epsilon_{\ell f}$ is measurement noise. The link is lost whenever $A_{\ell f}$ exceeds its link budget $B_{\ell f}$, or when a fade or collision erases the packet for reasons unrelated to dust.
The estimator we propose maximizes a censored likelihood. Received readings contribute their density. Budget-exceeding outages contribute the probability that $A_{\ell f} > B_{\ell f}$. Independent packet erasures contribute nothing, and they are distinguished from budget outages using the link-layer outage indicators already available to the radio. Grain size $r$ enters as a shared nuisance parameter. It is constrained by differences of attenuation across frequency bands on the same link and time window. Taking these differences removes offsets that are common across bands. The ratio of the remaining dust terms, $\kappa_{f_1}(r)/\kappa_{f_2}(r)$, does not depend on the dust amount and so carries information about size. Uncertainties come from the M-estimator sandwich covariance, the approach Frequentist Uncertainties on Neural Density Ratios with wifi Ensembles uses to obtain valid coverage without bootstrap pseudo-experiments.
The ratio trick is borrowed on purpose. Reactor neutrino experiments measure the same quantity in near and far detectors so that shared detector and flux uncertainties cancel (New Measurement of Antineutrino Oscillation with the Full Detector Configuration at Daya Bay). We apply the same logic across frequency bands on one radio path.
One honest caveat shapes the program. Martian dust grains are micrometres across, while THz wavelengths are tens to hundreds of micrometres. Over much of the band, extinction may therefore scale almost purely with dust mass and carry little size information. If that holds, the band ratio cannot identify $r$. It then also means that mass loading, rather than number concentration, is the observable that resists size uncertainty. Either outcome removes the size systematic, but by different routes. The first task is to find out which regime applies.
Research Program
The first step is an identifiability study. We use the HITRAN and Mars Climate Database attenuation setup from the MDSD paper and compute, across candidate band pairs in the CO₂ absorption windows, how much information the band ratio carries about $r$ compared with $M$. The output is a map of where size can be inferred and where the pipeline should switch to estimating mass loading directly. CO₂ absorption depends on temperature and pressure, and those enter as further nuisance parameters constrained by climatology.
The second step is the loss model. We add a link budget and a fading process to the MDSD simulator, so outages become correlated with dust, and we also inject dust-independent Bernoulli and bursty packet erasures. This contrast is the conceptual core. Distributed Training under Packet Loss shows that renormalizing the surviving shards keeps estimates unbiased when loss is random and independent of the data. That result supplies the correct treatment for the dust-independent erasures. It also sets a limit: the same renormalization is biased when loss depends on the quantity being measured, and storm-driven outages have exactly that property.
The third step builds the censored-likelihood estimator with the size nuisance parameter, then maps the per-link estimates of $M_\ell$ into a spatial field. It propagates the sandwich covariance through the interpolation stage so that every map pixel carries an interval.
The fourth step tests robustness to mismodeling. We deliberately use the wrong grain-size distribution shape and the wrong budget $B_{\ell f}$, and we measure how coverage degrades.
A secondary thread asks whether the per-link offset $g_{\ell f}$ can be removed without channel-state calibration. The motivation comes from over-the-air computation, where CHARLES adapts to imperfect channel estimates and NCAirFL avoids them entirely with power-only detection. Our band-difference construction is a sensing-side counterpart of the second approach, since it needs only received power and never phase.
Baselines
The naive reference is the MDSD pipeline as published, fed only the readings that survive: dropped samples are discarded and grain size is fixed at the climatological value. The strongest existing method is MDSD's best configuration, linear interpolation. It reported mean absolute error "approximately 50% lower than other methods" across node densities and a correlation "up to 0.93 during dust storms" (MDSD). We will rerun it on the identical simulated scenes, node densities and loss processes, with the same interpolation stage. The only differences will be the per-link estimator and its uncertainty.
The file gives no absolute MAE values, and no coverage or calibration result for MDSD was found on Lacuna. Our comparison therefore has two targets. We must match or beat its storm-time correlation and MAE. We must also deliver 68% and 95% interval coverage within a few points of nominal, at outage rates where the baseline's intervals, however they are constructed, under-cover.
Two ablations isolate the key ingredients. The first drops censoring by treating every outage as missing at random, with survivor renormalization in the spirit of Distributed Training under Packet Loss. The second drops the band-ratio size constraint and marginalizes $r$ over its climatological prior. A third reference uses naive ensemble spread in place of the sandwich covariance, which the wifi-ensemble paper found under-covers (wifi ensembles). All variants share the same scenes, random seeds and hyperparameter search budget.
Why not just…?
Why not just retransmit lost packets? Retransmission can recover dust-independent erasures, but it cannot turn a budget outage into a reading. While the path is saturated, no packet gets through, and the eventual successes are biased toward clearer moments. Reliable-transport feedback also scales poorly in multi-node settings, as shown in the unicast/broadcast comparison and Distributed Training under Packet Loss. The first ablation measures the remaining bias directly by adding an automatic-retransmission arm to the simulator.
Why not just add more links? MDSD shows that higher node density cuts MAE by 40–60%. However, denser coverage shrinks statistical scatter, not a size bias shared across links, and correlated storm outages hit neighbouring links together. We will plot error against node density for each variant. If the naive pipeline's error flattens at a floor that ours avoids, the systematic is confirmed. If denser deployment alone closes the gap, that is a real threat and we will report it.
Why not just marginalize grain size over climatology? This is the most serious alternative. If the identifiability study finds the band ratio uninformative about $r$, marginalizing and reporting mass loading may do as well as our approach. The second ablation settles the question: our method must produce narrower intervals at equal coverage. If it does not, the contribution reduces to the censoring model plus a recommended choice of observable, and we will present it that way.
Why not just impute missing readings by spatial interpolation? Interpolation fills gaps from neighbours, and those neighbours are the other clear links. That is exactly the selection bias we are trying to remove. The comparison against the linear-interpolation baseline under dust-correlated outages tests this.
Decisive Checks
Three results decide whether the idea holds. The first is the identifiability map, which determines whether size is inferred or cancelled. The second is a bias curve in storm peak mass loading against outage fraction. Naive handling should degrade monotonically, while the censored estimator should stay flat until outages are nearly total. The third is interval coverage under deliberate mismodeling.
All evaluation here is in simulation, as in MDSD. Field validation would require a THz testbed with controlled dust loading, and no such dataset appears in the file. If coverage fails under mild mismodeling of the link budget, the censoring premise is too fragile for deployment, and that negative result is itself worth publishing.
References
- Martian Dust Storm Detection With THz Opportunistic Integrated Sensing and Communication in the Internet of Space (IoS)
- Frequentist Uncertainties on Neural Density Ratios with wifi Ensembles
- New Measurement of Antineutrino Oscillation with the Full Detector Configuration at Daya Bay
- Distributed Training under Packet Loss
- CHARLES: Channel-Quality-Adaptive Over-the-Air Federated Learning over Wireless Networks
- NCAirFL: CSI-Free Over-the-Air Federated Learning Based on Non-Coherent Detection
- The Performance Comparison of the Unicast Routing Protocol and the Broadcast Routing Protocol in the Small-sized Ad hoc Network