Written 15 Oct 2026 for the fog-forecast tool. Sources are the operational references collected in research/A-fog-science-and-data.md (NWS Houston fog guide, Haby's forecasting notes, NWS Elko and Miami fog studies, NAV CANADA, Met Office/Boutle 2018, KNMI LES work), Marc Adamus's weather notes (research/B-marc-adamus-fog-notes.md), and the tool's own backtest (research/backtest.md). Where a number is a rule of thumb rather than a measured threshold it says so.
Almost every photogenic fog morning here in October and November is radiation fog: the ground cools by radiating heat to a clear night sky, the air touching it cools to its dewpoint, and water condenses in the lowest few metres to tens of metres. It is a flat-land, calm-night, clear-sky phenomenon. The Kankakee and Iroquois valleys add a second ingredient, steam fog over the open river when the water is much warmer than the air; the Wilmington gauge still reads about 19 °C in mid-October while dawn air is 5–10 °C, so river steam will show on many mornings even when the fields stay clear. Marc's own one-line rule is the right starting point: "if the air temperature falls below the dew point there must be fog" (research/B, Windy discussion at 36:46).
What this region does not get: valley fog from cold-air drainage down mountain slopes (there are no slopes), and marine layer. The closest local analogue to a "valley" is the river bottom itself, a few metres lower than the fields, where cold air pools and the river feeds moisture.
1. Dewpoint depression at dawn (weight 0.30). The air has to saturate. Operational guidance says fog forms when the spread between temperature and dewpoint closes to under about 5 °F / 2.8 °C with the right wind and sky [NWS Houston guide]. The tool reads the minimum spread in the three hours before sunrise from each model. The backtest found that on real fog mornings the day-ahead models typically still show a spread of 1.5–2.3 °C, because a 3 km grid does not resolve the coldest low ground, so a model spread under 2 °C is already a strong signal here.
2. 10 m wind (0.20). Calm air gives dew or a film of fog on the grass; 2–7 kt stirs the cooled air upward and thickens the layer [Houston]; above about 8 kt the mixing warms the surface air and fog cannot form (the tool applies a hard penalty above 10 kt). The deep fog bank you want for the drone frame needs that light breeze, not dead calm.
3. Sky overnight (0.15). Cloud above the ground reflects longwave radiation back and stops the cooling. Unexpected cirrus can stop fog; a cloud sheet arriving over an existing fog dissipates it early [NAV CANADA, Boutle 2018]. The tool uses mid plus high cloud averaged over the night; low cloud that the model pairs with good visibility is treated as a stratus deck, which is a different and worse morning.
4. Cooling from evening to dawn (0.10). A big temperature fall under clear skies is the signature of radiational cooling. The tool discounts this when the dewpoint also falls a lot, because that is a dry air mass moving in behind a front, not radiation.
5. Moisture: crossover and wet ground (0.10). The crossover method: take the afternoon dewpoint; if the forecast overnight minimum reaches it, expect fog; if the minimum goes about 1.5 °C below it, visibility under half a mile [NWS Elko, citing Baker et al. 2002]. Rain in the last 48 h, wet soil, and the river all keep the dewpoint from falling with the temperature. The classic set-up is afternoon rain, clearing at night, light wind [Haby].
6. The model's own fog signal (0.15). HRRR and NBM carry a visibility field. When it drops under 1 km in the pre-dawn hours that is the model explicitly producing fog. ECMWF and ICON have no visibility field, so the tool uses their relative humidity instead.
Two things act as gates rather than weights: precipitation at dawn (drizzle, not fog) and a stratus deck with good visibility.
Nobody publishes a metre-based rule for this; the split below is the tool's synthesis of the sources, and it says "estimate" on every depth number.
Shallow ground fog (the lone-cottonwood frame, 0.5–5 m deep): the air saturates only in the last hour or two before sunrise, the wind is near calm, and only the layer touching the wet ground or the river fogs. The fog sits below the tree crowns, the sun rises over a clear horizon or over a deeper bank behind, and shadows of trunks and branches project toward you through the fog. Shoot into the sun. Marc: "the stronger the light the more I like to look directly into the source of it"; the sun behind the subject "picks up atmosphere" (research/B §3).
Deep fog bank (the spruce-tops frame, 20–80 m): the air saturated four or more hours before sunrise and a 2–6 kt breeze kept stirring the cooled air upward, so the saturated layer grew until it buried the tree lines. Half of radiation fog events at one English site stayed thin and stable; the other half became well-mixed deep fog [Boutle 2018]. The HRRR vertical profile for Kankakee (Penn State BUFKIT) gives the only direct read on this: its lowest levels are about 10, 35, 80 and 160 m above ground, and the tool reports how high the saturated layer reaches. Saturation at 35 m or more supports a bank deep enough for the aerial frame; saturation only at 10 m means ground fog.
Steam fog over the river can be deep and dramatic over the water (visibility under 300 m under a strong inversion [Houston]) while the fields are clear. It forms when the water is roughly 5 °C or more warmer than the air and the wind is light; the tool reports the Wilmington water temperature against the forecast dawn air.
Beams need three things at once: fog that is still there after the sun is up, sun strong enough to be a point source (little mid or high cloud, no stratus deck above the fog), and something to cast shadows through the fog (tree lines, lone trees, corn rows). The tool's "light above the fog" line is mid plus high cloud at sunrise, with a low deck added when the model says that deck is cloud rather than fog.
Sun geometry for Bourbonnais: sunrise moves from 07:04 at 101° (ESE) on 16 October to 06:22 at 109° on 1 November (after the clock change) and 06:39 at 114° by mid-November. The sun climbs about 10° in the first hour. Long tree shadows across fog need the sun under about 12°; the tool prints the window from 2° to 12° as the beam window. Be set up by civil dawn (about 30 minutes before sunrise): Marc's "two peak times" are the sunrise itself and the 30–60 minutes either side (research/B §3), and for fog the second peak is the one that matters because the fog is backlit only once the sun is up.
Burn-off. Thin, stable fog goes quickly; well-mixed fog lingers [Boutle]. Visibility can dip briefly right after sunrise as the first heating stirs the layer [NAV CANADA]. The rules of thumb the tool uses: shallow fog clears 30–75 minutes after sunrise, patchy fog 60–120, a deep bank 2–3.5 hours; add half an hour if cloud above weakens the sun; subtract if the wind picks up. The satellite-based depth rules (about 3 h for a 300 m deck) do not transfer to fog this shallow. Next to the estimate the report prints the hour at which HRRR's own visibility and dewpoint spread recover, labelled as model output.
329 airport mornings (Kankakee, Joliet, Romeoville, Gary, Rensselaer, Pontiac; 21 Aug–15 Oct 2026) scored from archived model runs older than the ones the evening run sees (HRRR and NBM from 13:00 the afternoon before, ECMWF and ICON from 07:00); 34 had fog at or under one mile, on 13 distinct mornings. Ranking skill (AUC) 0.76. Fitted on the other five stations and scored on the sixth, the probability beats climatology (Brier 0.085 vs 0.093), and forecasts in the 30–40 % band verified on 5 of 10 cases. The calibration is airport-based: airports sit on open, drained ground and fog less than river bottoms and low fields. Read the probability as "chance the airport reports fog" and the site table as a ranking. At any threshold the tool misses more fog mornings than it catches, because the models themselves rarely saturate the 2 m air at this lead; that is why the pre-dawn 04:45 refresh exists, and why the report shows the NBM's own visibility probabilities and the live Kankakee observation as independent checks.
1. Grade and probability at home, with confidence from model agreement (Marc: agreement means confidence, disagreement means low confidence).
2. Type and depth, and whether the HRRR profile shows a saturated layer.
3. Light above the fog at sunrise.
4. Sunrise time and azimuth, beam window, burn-off window.
5. The NWS Chicago forecast discussion paragraphs that mention fog or stratus (Marc calls the AFD "the best forecast that you can possibly get for any region in the United States").
6. The live Kankakee METAR against the HRRR's value for the same hour: if the model is already running drier than the air outside, trust the air.
7. River water temperature versus dawn air.
The outlook page scores each of the next seven mornings for a region's home point with the same factor engine, using ECMWF, GFS, ICON and NBM (HRRR joins where its 48-hour reach allows). It reports a set-up score, the raw factor blend before calibration, and ranks the mornings. It is not a probability: in the day-1 backtest a "strong" set-up (score ≥ 0.75) fogged the airport on a quarter of mornings, "decent" on about one in six, and "poor" on fewer than one in ten. Skill falls with lead; the lead backtest in research/backtest.md gives the ranking skill at 2, 3 and 5 days out, and the outlook page prints those numbers so the wording is backed. Use days 5–7 to spot a calm, clear high-pressure stretch, re-check nightly, and let the day-1 report make the decision.
Grayslake sits 12 miles inland from Lake Michigan in Lake County. The same radiation-fog mechanics apply on calm clear nights, with the Des Plaines River, Fox River and the chain of lakes and marshes as the moisture sources. Lake influence is the part this tool does not model. What could be verified (research/D-grayslake-locations.md): CIMSS case studies show lake-breeze fog cutting visibility to half a mile at Waukegan when an easterly wind brings a shallow, stable layer of cold air ashore, both in the warm season; the Midwestern Regional Climate Center notes dense fog is most common in the Midwest from October to March between midnight and mid-morning and is denser in low ground; and a NOAA Lake Erie study notes a still-warm autumn lake can short-circuit fog near the shore. Read that as: an easterly flow off the lake is a different animal from a radiation night, and the tool's score does not know about it.