fog between the sensors and the satellite

Problem statement

Dense fog on a highway is a highly localized phenomenon — a valley, a river crossing, a few miles of low ground — and the state DOTs responsible for warning drivers observe it through widely spaced roadside weather stations (RWIS/ESS) because, as the FHWA problem statement says, it is "difficult and cost prohibitive to install surface (ground)-based visibility sensors continuously along the highway system." Geostationary satellites (GOES) see fog and low cloud everywhere every few minutes, but their operational fog products were built for aviation and synoptic meteorology: infrared resolution is about 2 km at nadir, the products struggle to tell ground-touching fog from elevated stratus, cannot compute cloud thickness under multiple cloud layers, lose skill at sunrise and sunset, and do not detect smoke. The unsolved problem is turning the combination — sparse but exact ground sensors, plentiful roadside cameras, and coarse but continuous satellite products — into a validated, road-segment-level visibility hazard notification that a traffic management center can act on.

Why this matters

FHWA's Road Weather Management program reports that over 38,700 vehicle crashes occur in fog each year, killing over 600 people and injuring more than 16,300 annually (long-run annual averages, no data year stated; a 2025 NOAA NESDIS article restates the same figures as 2024 numbers), and fog pile-ups on rural interstates are among the deadliest multi-vehicle crashes. Because the failure is spatial — the fog forms between the sensors — a warning system built only on RWIS will miss the events that matter most, while a warning built only on satellite will over-warn where low cloud is aloft, and drivers learn to ignore both. The FHWA author frames a validated satellite-plus-ground product as a step toward the national Vision Zero goal specifically for weather-visibility crashes.

What’s been tried and why it hasn’t worked

State DOTs have deployed RWIS visibility sensors and, more recently, machine-learning classification of roadside camera images and in-vehicle video (Khan & Ahmed's RoadweatherNet and webcam CNN work, cited in the statement); these are accurate where they exist but are point observations, and cameras need landmarks at known distance or image-sharpness heuristics to infer range. On the satellite side, UW-SSEC/CIMSS produce the GOES-R Fog/Low Stratus products, which fuse GOES imagery with Rapid Refresh model fields to give an "IFR probability" that stays consistent from night into day and even under high cloud — but the products themselves document the limits: cases where the nighttime microphysics signal indicates low cloud that observations show to be elevated stratus rather than fog, no cloud thickness in multi-layer situations, 2-km infrared resolution that misses fog in narrow valleys, and no smoke detection. Fog products are also designed and validated against airport IFR conditions, not highway visibility distance. The problem statement's proposed remedy — an AI/ML fusion of satellite, RWIS, and camera data validated during known fog events and crashes, at $548,000 over 30 months — was reviewed favorably by NCHRP staff but no funded project was found, and the reviewer noted the proposal bundled three separate objectives (fusion model, nationwide satellite fog product, NWP validation).

What would unlock progress

The unlock is a highway-specific validation dataset: time-aligned satellite fog-product pixels, RWIS visibility readings, and camera-derived visibility for the same road segments during many fog events, so that a fusion model can learn when the satellite signal means ground fog on the road and when it means stratus aloft. With that in hand, a probabilistic segment-level "visibility hazard" product becomes feasible, and its skill can be reported honestly (hit rate, false-alarm rate) rather than assumed. The adjacent solved problem is aviation, where satellite IFR-probability products were validated against airport ceilometer/visibility observations for exactly this fog-vs-stratus discrimination; highways need the same calibration against roadside truth.

Entry points for student teams

A team could pick one fog-prone interstate corridor with public RWIS data and DOT camera feeds, pull archived GOES-R Fog/Low Stratus product output for a season, and build the aligned dataset plus a first fusion classifier, reporting skill against RWIS visibility. A computer-vision team could develop a camera-based visibility estimator that works without surveyed landmarks (e.g., using lane markings as known-geometry references) to densify ground truth. A human-factors/operations team could design how a probabilistic segment alert would be displayed to TMC operators and pushed to navigation apps without over-warning. Relevant skills: remote sensing, machine learning, meteorology, computer vision, and traffic operations.