Cities plan to route delivery drones and air taxis over quiet residential courtyards on the assumption that buildings block their sound. Current noise tools estimate the drone's emission, the shielding provided by buildings, and the background noise that hides it as three separate quantities, each taken at a typical value. The weather drives all three at once. A gusty downwind hour makes rotors noisier, bends sound over rooftops into the shielded "quiet side", and scatters it there. A calm night with a temperature inversion (warm air above cool air near the ground) also bends sound downward, and it arrives just as traffic thins and the background that would mask a drone disappears.
This note proposes to measure that coupling where it matters, in the rare hours when all three terms are bad together. We will describe weather-driven changes in rotor emission with acoustic analogy theory. We will describe weather-driven loss of building shielding with wave-based propagation models. Then we will model their joint extremes with multivariate peaks-over-thresholds statistics. The central question is whether treating the three terms as independent, as current practice does, badly underestimates how often a drone becomes noticeable on the quiet side or pushes total exposure past the roughly 65 dB(A) level where masking stops helping.
Core Claim
A drone's noise at a shielded receiver, measured relative to the background, adds up in decibels from parts that share weather as a common driver. We write the noticeability margin as
$$N = L_0 + S + G - L_a,$$
where $L_0$ is the drone's reference level at the receiver under still air (the Noise-Power-Distance lookup used in routing studies), $S$ is the extra source level caused by turbulent or gusty inflow to the rotor, $G$ is the extra propagation gain relative to a still atmosphere, and $L_a$ is the ambient level. Under the noticing model, an event is noticed when $N > 0$.
Our claim is that $S$, $G$, and $-L_a$ are dependent in their upper tails. They are not independent, and they are not linked through a smooth Gaussian correlation. We expect the dependence to come from at least two distinct weather regimes:
- In the gust regime, $S$ and $G$ peak together.
- In the nocturnal inversion regime, $G$ and $-L_a$ peak together.
The useful structural fact comes from Multivariate peaks over thresholds models, which defines extreme episodes as hours when at least one component exceeds its threshold. That paper proves that positively weighted sums of multivariate generalized Pareto components, conditioned on being large, are again generalized Pareto. Because $N$ is a unit-weighted sum in decibels, a fitted joint tail model for $(S, G, -L_a)$ gives the tail of $N$ in closed form. A tail-calibrated estimate of rare quiet-side exceedance probabilities is the quantity a routing constraint actually needs.
Research Program
Source term. The source term $S$ is where acoustic analogy enters. The Ffowcs Williams–Hawkings (FW-H) equation computes far-field rotor sound from surface pressure loading and blade thickness on an integration surface around the rotor. It has been validated on propellers in Experimental and Numerical Study of a UAV Propeller Printed in Clear Resin and on tandem configurations in Acoustic and aerodynamic evaluation of POLIMI tandem propellers in hover and cruise flight conditions within GARTEUR AG26. Those studies use clean inflow. We will run the same FW-H pipeline with synthetic turbulent inflow at several turbulence intensities and mean gust speeds. From these runs we will build a low-dimensional response surface for $S$ as a function of inflow turbulence and thrust demand, in A-weighted octave bands. We need only a response surface here, not a new solver. Its role is to replace the weather-independent lookup table that Noise-Aware and Equitable Urban Air Traffic Management uses for the NASA quadrotor.
Propagation term. The term $G$ comes from two models of different fidelity. The cheap model is the engineering turbulence-scattering model of A Model of Sound Scattering by Atmospheric Turbulence for Use in Noise Mapping Calculations. It is fast enough to evaluate for every hour of a multi-year weather record. However, its canyon error has a standard deviation of about 6 dB, which is too large to trust in the tail. The expensive model is the coupled FDTD–PE solver of Parameter study of sound propagation between city canyons with coupled FDTD-PE model:
- FDTD (finite-difference time-domain) computes the wave field inside the source canyon.
- PE (parabolic equation) carries the sound over rooftops into the receiver canyon.
That study reports about 10 dB of lost shielding at friction velocity $u_* = 1$ m/s, but only for a few fixed weather states. We will run FDTD–PE only for weather states selected in the tail and use those runs to fit a correction to the cheap model where the correction matters.
Ambient term. $L_a$ comes from measured sound levels. The routing study draws on national exposure maps. For the statistical work we will use sensor streams from the SONYC network in New York, paired with hourly meteorology from nearby stations.
Statistical model. Weather drivers (friction velocity, gust intensity, near-ground temperature gradient, wind direction relative to the street axis, and traffic volume) map through the physics into hourly vectors $(S, G, -L_a)$. We fit these vectors with the observed-scale multivariate generalized Pareto construction and the censored likelihood of Peaks Over Thresholds Modeling With Multivariate Generalized Pareto Distributions. Censoring keeps ordinary hours from biasing the tail fit, and the observed-scale form keeps the decibel sum physically meaningful.
Regimes. A single dependence structure may be wrong, so we first run the angular spectral clustering of Spectral learning of multivariate extremes on the exceedances. That paper groups extreme events by the direction of their joint excess and found distinct winter clusters in Leeds air pollution. Here it tests whether separate gust and inversion regimes really exist. If they do, we fit a mixture with one component per regime. The inversion regime is motivated directly by Influence of correlation between diurnal traffic pattern and meteorological conditions on long-term average Lden, which found that traffic peaks line up with inversions and raise long-term exposure.
Constraint from asymptotic independence. Multivariate generalized Pareto models assume asymptotic dependence. Some pairs, for example rotor gust noise and nocturnal low ambient, may instead be asymptotically independent. For each pair we will report the tail dependence measure $\chi(q)$ as a function of the quantile $q$. Where $\chi(q)$ decays toward zero, we will say so rather than force a dependent model.
Outputs. The deliverables are:
- the fitted per-regime tail models;
- closed-form tail probabilities for $N > 0$ and for total level above 65 dB(A) at quiet-side receivers;
- a chance-constraint version of the corridor masking rule used in the routing study, evaluated on representative canyon geometries.
Baselines
Naive reference. The do-nothing reference is current practice. It uses the still-air Noise-Power-Distance level and quiescent geometric shielding, then adds a fixed safety margin of 3 or 6 dB. This is how the routing study and regulatory mapping tools treat weather. It is the baseline any operator would actually use. We will apply it to the same canyon geometries and the same held-out weather years. It is beaten if its implied exceedance frequency of the noticeability and 65 dB(A) thresholds differs from the held-out truth by a larger factor than our model's does. We expect it to be either far too permissive or uniformly over-restrictive.
Independence product. The second baseline uses the same physics but treats $S$, $G$, and $L_a$ as independent. It fits each marginal tail carefully, with univariate generalized Pareto distributions, and convolves them. This isolates the key ingredient of the proposal, joint tail dependence, because marginals, physics, and data are identical. Only the dependence assumption changes.
Gaussian copula. The third baseline keeps the same marginals and joins them with a Gaussian copula fitted to all hours. This is the standard "account for correlation" fix. Because it fixes tail dependence at zero, it tests whether bulk correlation is enough.
Empirical Monte Carlo. The strongest competitor is empirical Monte Carlo. The cheap-plus-corrected physics is run for every hour of a long weather record, and exceedances are simply counted. On a long enough record this is close to the truth, and we use a decade-scale record as the reference for the held-out evaluation. The fair comparison gives every method the same short training record, one to two years, which is typical for a new deployment site. Every method also gets the same budget of FDTD–PE runs: random hours for Monte Carlo, tail-selected hours for ours. Each method is then scored against counts from the held-out long record.
What counts as beating it. We use the occurrence ratio and severity ratio of Tail calibration of probabilistic forecasts. The occurrence ratio compares how often the threshold is exceeded with how often it was forecast to be. The severity ratio compares the size of excesses above it. Winning means ratios closer to one at return periods longer than the training record, together with narrower uncertainty intervals at equal FDTD–PE cost. At return periods inside the training record, we expect empirical Monte Carlo to tie, and we will report that tie as a result.
No result on quiet-side exceedance tails exists to compare against: no reported result found on Lacuna for any method on this quantity. The nearest reported figures are physics anchors rather than tail scores: about 10 dB of shielding loss at $u_* = 1$ m/s from the FDTD–PE study, and about 6 dB canyon error standard deviation for the engineering scattering model. Each fitted correction must reproduce these first.
Source-term ablation. A further ablation replaces the FW-H weather-dependent $S$ with zero, the static lookup assumption. This measures how much of the compound tail is due to the source.
Why not just…?
Why not just add a fixed safety margin to standard noise maps? Weather-driven shielding loss in a single canyon can reach about 10 dB according to the FDTD–PE study. Canyon geometry alone changes scattered levels by 7 to 14 dB according to the turbulence scattering study. A margin large enough to cover those cases would block corridors on most ordinary hours, and a smaller one fails exactly on the hours that matter. The naive baseline above measures this trade-off directly. We will report how often each fixed margin is violated and how much corridor capacity it costs.
Why not just run Monte Carlo over the observed weather record and count? This is the real threat. Within the span of a long local record, plain counting with good physics may match our model, and the experiments are designed to show where it does. Our model has an advantage only when one of two things is true: the required probability is rarer than the record supports, or the expensive solver can be afforded only for a few hundred states. These are the conditions of a new vertiport site with one year of data and a chance constraint at roughly the once-per-several-years level. The short-record versus long-record evaluation settles this question empirically. If counting wins there too, the defensible conclusion is that the statistics are unnecessary and only the physics coupling matters. That would still be a finding.
Why not just train a tail-aware deep surrogate on the urban sound benchmark? Urban Sound Propagation: a Benchmark for 1-Step Generative Modeling shows that fast image-to-image models degrade in shielded courtyards. Its ground truth contains no wind, inversion, or turbulence, and the same holds for the Beijing mapping model in A rapid approach to urban traffic noise mapping with a generative adversarial network. Adding heavy-tailed outputs such as Pareto latents or tail-transform flow layers to such a model would sharpen the spatial peaks of a weather-free field. It cannot represent the weather coupling that drives our claim. A surrogate might later replace the cheap propagation model. That swap is orthogonal to the question asked here.
Why not just use a Gaussian copula on sensor data? A Gaussian copula fixes tail dependence at zero. The forecasting literature also shows that models calibrated on typical data can miss tail exceedance frequency at high thresholds: in the precipitation study of Allen et al., a logistic model's occurrence ratio drifted away from one at the 99.5th percentile while a GEV model held. The copula baseline tests this directly on acoustic data.
Proposed Experiments
Physics anchors. First we reproduce the physical anchors:
- the roughly 10 dB shielding loss at $u_* = 1$ m/s in the FDTD–PE canyon configuration, extended to a grid of temperature gradients and street orientations;
- FW-H rotor levels under clean inflow against the clear-resin and GARTEUR propeller measurements, before any turbulent inflow is added.
These runs also give the cheap model's tail correction. We check it on held-out FDTD–PE runs drawn from the most extreme 2% of weather states.
Empirical test of propagation tail dependence. We test the driver–propagation dependence on real data without drones. At SONYC nodes on shielded sides, versus street-facing nodes, we estimate $\chi(q)$ between excess received traffic level and downwind or inversion indicators from colocated weather. We then cluster the exceedances spectrally. If shielded nodes show no more tail dependence on weather than exposed nodes, the propagation leg of the hypothesis fails.
Decisive held-out comparison. All baselines are compared on quiet-side receivers in representative canyon geometries. Each is trained on short records and scored on long held-out records with the occurrence and severity ratios, at thresholds $N > 0$ at night and total level above 65 dB(A).
The proposal succeeds if two things hold:
- The independence and copula baselines underestimate the occurrence of compound exceedances by a clear factor, while the regime-mixture model stays near calibration.
- The tail model matches or beats short-record Monte Carlo beyond the record length at equal solver budget.
The source-term ablation then reports what fraction of the excess risk comes from weather-dependent rotor emission. If that fraction is small, the acoustic analogy component should be dropped from future routing tools, and we will say so.
Downstream consequence. Finally we re-evaluate the corridor masking rule from the Austin routing study as a chance constraint. We count how many highway-masked corridors fail it on inversion nights.
References
- Multivariate peaks over thresholds models
- Experimental and Numerical Study of a UAV Propeller Printed in Clear Resin
- Acoustic and aerodynamic evaluation of POLIMI tandem propellers in hover and cruise flight conditions within GARTEUR AG26
- Noise-aware and equitable urban air traffic management: An optimization approach
- A Model of Sound Scattering by Atmospheric Turbulence for Use in Noise Mapping Calculations
- Parameter study of sound propagation between city canyons with coupled FDTD-PE model
- SONYC: A System for the Monitoring, Analysis and Mitigation of Urban Noise Pollution
- Peaks Over Thresholds Modeling With Multivariate Generalized Pareto Distributions
- Spectral learning of multivariate extremes
- Influence of correlation between diurnal traffic pattern and meteorological conditions on long-term average Lden (invited paper)
- Tail Calibration of Probabilistic Forecasts
- Urban Sound Propagation: a Benchmark for 1-Step Generative Modeling of Complex Physical Systems
- A rapid approach to urban traffic noise mapping with a generative adversarial network