Fog Banks: Imaging, Meteorology & Computing

• Fog banks are spatially coherent fog volumes that attenuate light through exponential decay, scattering, and multiple scattering effects.
• They are modeled in imaging with depth-dependent transmission functions and play a crucial role in meteorological nowcasting and visibility studies.
• The term also extends to distributed computing, where 'Fogbanks' refer to organized clusters of vehicular or edge resources that optimize power and latency.

Fog banks are spatially coherent fog volumes that suppress visibility through attenuation of scene radiance, addition of scattered ambient light, and multiple scattering, often with layered or patchy structure rather than uniform opacity. In imaging-oriented formulations, their effect is commonly written as

$$\mathbf{I}(\mathbf{x}) = \mathbf{J}(\mathbf{x})\,T(\mathbf{x}) + \mathbf{A}\bigl[1 - T(\mathbf{x})\bigr], \qquad T(\mathbf{x}) = e^{-\beta\, d(\mathbf{x})},$$

where $\mathbf{J}$ is clear radiance, $\mathbf{A}$ is global airlight, $T$ is transmission, $\beta$ is the extinction coefficient, and $d$ is path length (Ingold et al., 27 Jun 2026). In meteorology, the term covers organized marine, coastal, and freezing-fog structures; in computer-systems literature, the cognate term “Fogbanks” denotes pooled vehicular or edge resources rather than atmospheric condensate (Cardoso-Bihlo et al., 4 Jun 2026, Alahmadi et al., 2020).

1. Optical structure and visibility degradation

Fog banks are difficult observational targets because they simultaneously weaken image-bearing light and add a veiling background. Outdoor fog experiments emphasize three coupled effects: attenuation or extinction of target light, backscatter or veiling background, and multiple scattering. The latter is often described in terms of ballistic photons, snake photons, and diffusive photons, with conventional cameras integrating all three contributions (Kumar et al., 2021). This is why dense fog banks can make distant beacons, objects, and horizon structure disappear even when simple frame averaging is applied.

Computer-vision work treats the same phenomenon as a depth-dependent volumetric process. In monocular fog synthesis for multiple-object tracking, fog formation is modeled with the classical Koschmieder form

$$\mathbf{I(x)}=\mathbf{I_0(x)}\cdot \mathbf{T(x)} + L_\infty(1 - \mathbf{T(x)}),$$

with transmission decreasing exponentially with depth and atmospheric light estimated either from visible far-field regions or from the dark channel prior when sky or horizon is absent (Kirillova et al., 2024). This directly captures the defining property of fog-bank visibility loss: farther structures are degraded first, while nearby structures remain comparatively legible.

Neural radiance field work makes the volumetric interpretation explicit. Standard NeRFs trained on foggy imagery can represent visible fog as low-density volume distributed through space, while denser occupancies encode solid scene content. Post-training thresholding of low-density samples can therefore reduce visible fog at render time, but it cannot recover geometry that was never observed through a dense bank during training (Teigen et al., 2023). This distinction is important: fog-bank removal in imagery is not equivalent to reconstructing physically hidden structure.

2. Meteorological forms, spatial organization, and unusual occurrences

Marine fog banks are prominent in the northwest Atlantic, including the Grand Banks area and the vicinity of Sable Island, where fog is common, highly variable, and operationally consequential for shipping and offshore activity. Minute-resolution ship observations from the FATIMA campaign were collected in July 2022 aboard the Research Vessel Atlantic Condor using a Vaisala Forward Scatter Sensor model FD70, a Vaisala Weather Transmitter model WXT50, and a Gill R3A ultrasonic anemometer (Gultepe et al., 2024). The same regional setting also motivates a stochastic-deterministic fog-cover model for St. John’s, Newfoundland, where warm, moist air is advected over colder coastal waters influenced by the Labrador Current and terrain-induced upslope effects near the Avalon Peninsula. In that model, fog cover is treated as a two-scale phenomenon: mesoscale moisture tendencies drive saturation, while an Ising-type subgrid model produces clustered horizontal structure, including “bands, rolls, and closed or open cells” (Cardoso-Bihlo et al., 4 Jun 2026).

Freezing fog can also behave as a source region for precipitation. Observations from the UK winter of 2008/09 document snowfall during anticyclonic, non-frontal, foggy conditions, with freezing fog reported as the sole weather at most synoptic stations across the UK and central Europe on 31 December 2008 and a typical fog depth of 300–400 m from radiosonde ascents (0902.1326). Cloud radar indicated no cloud over the UK until after 12:00, when only thin cirrus above about 7 km appeared. The reported precipitation consisted of very slight, small snow grains, leading to the proposed term “swizzle,” a contraction of “snow” and “drizzle.” The report links such episodes to earlier literature on Anthropogenic Snowfall Events and treats the anthropogenic-particle explanation as plausible but not definitively proven.

The notion of a fog bank has even been extended beyond Earth. Post-landing analysis of 82 Side Looking Imager images from the Huygens probe on Titan found six images containing an extended, horizontal feature whose inferred optical-depth change was between 0.005 and 0.014. The preferred interpretation is a fog bank close to the horizon that rises and falls during the observing interval, although the authors retain detector artifacts, terrain effects, and mirage as alternatives (Smith et al., 2016). This suggests that the fog-bank concept is fundamentally geometric and radiative: a coherent near-surface opacity boundary may be recognizable even in very different atmospheres.

3. Observation, discrimination, and active clearing

Field imaging through fog banks has motivated both passive and active optical methods. A notable passive-active hybrid is quadrature lock-in discrimination (QLD), which modulates a source sinusoidally and then demodulates camera frames at the same frequency to recover only the tagged optical component. In outdoor field tests, a 10-LED red panel at 640 nm, modulated at 13 Hz, was viewed over 150 m in visibility of 40 m. A raw frame yielded contrast-to-noise ratio (CNR) 2.3, averaging all 10,140 frames gave CNR 2.5, and QLD increased CNR to 11.0 (Kumar et al., 2021). The same approach revealed a cardboard target at 75 m in about 30 m visibility, improving CNR from 0.3 to 1.9, and suppressed strong daylight clutter at 150 m, increasing CNR from $-0.3$ to 2.4. The method is explicitly not true ballistic filtering at the employed 13–17 Hz modulation frequencies; rather, it selects light that retains the imposed modulation after propagation through fog.

QLD also clarifies a common misconception about fog imaging. Simple temporal averaging reduces random noise but not the strong static haze and backscatter background, so it does not recover obscured targets when the background dominates. Frequency-selective demodulation can cancel the unmodulated component, but its performance depends on fog microphysics, ambient illumination, and fog dynamics; in practice, around 100 modulation periods, or about 7.7 s at 13 Hz, were often needed for strong CNR gains (Kumar et al., 2021).

More aggressive active control seeks not merely to see through fog banks but to displace droplets. A plasma-free approach based on molecular quantum wakes uses resonant trains of eight 56 fs pulses separated by the nitrogen rotational revival time $T_R = 8.36\,\text{ps}$. The pulse train creates a rotational wave packet in $\mathrm{N_2}$, whose collisional thermalization launches a single-cycle acoustic wave and a transient low-density channel that ejects droplets radially (Schroeder et al., 2020). In chamber fog with initial attenuation of 9 dB, a resonant 8-pulse train increased $\mathbf{J}$0 transmission by a factor of 3, or 4.8 dB; full transmission recovery was achieved up to $\mathbf{J}$1 initial extinction. The cleared channel remained sub-millimetric, with a measured density-depleted diameter of 500–700 $\mathbf{J}$2, so the result is a proof of principle rather than a field-scale solution for wide fog banks.

4. Computational restoration, rendering, and robustness evaluation

Recent computer-vision work treats fog banks as a transfer problem across increasingly out-of-distribution domains. A paired defogging pipeline first learns from a single-camera fog imager that photographs a flat-panel display through an artificial-fog enclosure with a fixed 114 mm scattering path, producing 5,495 pixel-aligned foggy/clear pairs, split deterministically into 4,943 training pairs and 552 held-out test pairs (Ingold et al., 27 Jun 2026). Exact registration enables pixel-exact $\mathbf{J}$3 reconstruction and a paired Laplacian ratio that predicts per-image NAFNet PSNR with Spearman $\mathbf{J}$4, versus 0.399 for the best single-image proxy. Under identical training conditions across 30 restoration backbones, NAFNet achieved 24.33 dB PSNR and 0.7912 SSIM on the held-out fog-chamber split, while SpecAT S2 remained within 1.29 dB at about 3% of the parameter count. After synthetic fine-tuning on randomized Mapillary Vistas fog, the same model generalized without target-domain training from chamber fog to chamber-free free-flowing fog and to iPhone 13 aircraft-window video; on the unpaired aircraft domain, mean NIQE improved from 6.22 to 4.97. The paper is careful to frame this as visual restoration rather than physical inversion of the fog bank itself.

The same caution appears in neural rendering. A NeRF trained on foggy images can reproduce the fog-filled scene and then remove visible fog by thresholding away densities below a scene-specific global threshold chosen by contrast flattening over $\mathbf{J}$5 (Teigen et al., 2023). This works because visible fog occupies a lower-density regime than true scene content in the learned volume. However, if a fog bank fully hides background structure during training, thresholding cannot reveal accurate geometry behind it. The method therefore removes adverse volumetric effects from the representation; it does not estimate the physical structure of the bank or hallucinate unseen content.

Fog-bank degradation also exposes weaknesses in downstream perception. A physics-based volumetric fog simulation pipeline augmented MOT17 with four fog-intensity levels, using MiDaS 3.0 with a DPT BEiT$\mathbf{J}$6-Large backbone for framewise monocular depth estimation, depth-dependent transmittance, atmospheric-light estimation via dark channel prior, and heterogeneous fog from five-level Perlin-noise turbulence with brightness reduced to 80% (Kirillova et al., 2024). All four evaluated trackers degraded sharply. Under homogeneous Fog 4, ByteTrack fell from HOTA 80.34 in clear weather to 26.23, while FairMOT fell from 75.40 to 13.05; heterogeneous Fog 4 was slightly less damaging, with ByteTrack at 37.70 and FairMOT at 24.99. This establishes that fog banks are not merely an aesthetic nuisance: they alter the depth structure on which tracking, association, and re-identification depend.

5. Forecasting and nowcasting

Short-lead prediction of fog-bank visibility remains difficult because fog can form and dissipate rapidly, even within minutes. In the Grand Banks and Sable Island region, generative nowcasting used 120 minutes of lagged visibility, wind speed, dew point depression, and relative humidity with respect to water, filtered to $\mathbf{J}$7, at one-minute resolution (Gultepe et al., 2024). The study considered $\mathbf{J}$8 km and $\mathbf{J}$9 km subsets and compared cGAN regression with XGBoost, persistence, and persistence-window baselines. At 30-minute lead time and $\mathbf{A}$0 km, cGAN achieved RMSE 0.151 km, compared with 0.170 km for XGBoost and 0.164 km for persistence; at 60 minutes, XGBoost was best at 0.167 km and cGAN reached 0.209 km. For $\mathbf{A}$1 km, XGBoost slightly outperformed cGAN at both 30 and 60 minutes. The practical implication is narrow but important: generative models showed the most promise when the task was the short-range evolution of already foggy conditions, not broad reduced-visibility prediction over longer horizons.

A more explicitly spatial approach uses a stochastic lattice model. The St. John’s prototype solves a moisture-evolution equation on a $\mathbf{A}$2 km domain at 30 km resolution, giving $\mathbf{A}$3 mesoscale cells, each with a $\mathbf{A}$4 sub-lattice of binary fog/no-fog sites (Cardoso-Bihlo et al., 4 Jun 2026). The Ising Hamiltonian

$\mathbf{A}$5

introduces local positive interactions $\mathbf{A}$6, while the external field $\mathbf{A}$7 with $\mathbf{A}$8 couples fog occupancy to supersaturation; $\mathbf{A}$9 controls stochasticity. Over 95 independent 10-hour samples from May 2014–2017, using fog forecast if $T$0 and no fog if $T$1, the contingency table produced TP 354, FP 98, FN 35, and TN 463, corresponding to Accuracy 86%, Precision 78%, Recall 91%, False Positive Rate 21.6%, and $T$2 Score 84%. These values indicate a model better at detecting fog occurrence than at avoiding false alarms, which is consistent with a system designed to capture coherent fog-bank structures rather than pointwise station conditions alone.

6. Terminological extension in distributed computing

Outside atmospheric science, “Fogbanks” has become a distinct term in edge and vehicular computing. In this literature, Fogbanks are vehicular fog clusters formed from underused processing, storage, and sensing resources in connected and autonomous vehicles, particularly in parking lots, charging stations, roads, and intersections (Alahmadi et al., 2020). Vehicles are described as being used for only 2–4 hours per day, while office car parks may hold them for 7–8 hours per day and airport parking for one to two weeks. A cloud-supported Fogbanks architecture with 4 vehicular Fogbanks, each containing 5 to 15 vehicles with 3200 MIPS per vehicle, was optimized by MILP for total power consumption. Relative to central-cloud-only processing, the reported savings reached up to 80%; increasing vehicular-node density saved 38% to 64% compared with cases using only fixed fog or low vehicular density; and distributed allocation reduced power by up to 28% compared with single allocation.

A closely related parked-vehicle study treats a vehicular fog as a pool of clustered vehicles coordinated by a roadside unit within a cloud–fog–vehicular-fog hierarchy (Ma et al., 2020). There, 50 user requests each required one of 10 software packages, and the comparison between same-type and random-type software deployment showed that random-type deployment yielded up to 9% additional power savings over same-type deployment. In the largest random-type case, power savings reached 48% versus cloud-only and 34% versus cloud-plus-fixed-fog. The core architectural lesson is that a fog bank, in this computational sense, derives value from cluster-level diversity and coordination rather than from uniform replication alone.

The broader fog-computing literature uses the term “fog” even when “Fogbanks” is not explicit. Agentic Fog models fog nodes as policy-driven autonomous agents communicating through shared memory and localized coordination, formalized as an exact potential game with convergence under asynchronous bounded-rational best-response dynamics and stability under node failures when shared memory persists and the communication graph remains connected (Akbar et al., 28 Jan 2026). In simulation, this framework reduced average latency by 15–30% relative to greedy heuristics and 10–18% relative to ILP under dynamic workloads. Another line of work asks how many fog nodes should be upgraded in a cloud–fog–thing network; in finite-area stochastic geometry, the optimum number is sublinear in total node count, and optimizing it can increase average data rate nearly an order of magnitude in shadowing and fading (Balevi et al., 2018). At a smaller scale, a smart fog gateway for wearable IoT was implemented on Intel Edison and Raspberry Pi, with the Raspberry Pi completing the process almost two times faster and showing around 12.39 seconds average wait time versus 64.65 seconds for Intel Edison (Constant et al., 2017). In all of these cases, “fog bank” denotes an organized pool of proximate compute resources, not an atmospheric opacity field.

The dual usage is conceptually revealing. In atmospheric science, a fog bank is a structured, visibility-reducing volume whose morphology, radiative effects, and evolution matter operationally. In computing, Fogbanks are structured, proximate resource pools whose locality, heterogeneity, and coordination matter operationally. The shared term reflects clustering and near-field interaction, but the referents are otherwise distinct.