FogCam Waits for Dark
FogCam has watched the San Francisco State campus since 30 September 1994, and this week it turned 32. Since July, Justin Earl's Fog-Bank has been catching every frame it publishes. I downloaded the timestamps of 102,907 of them and found that the camera keeps time by the sun.
Slower after dusk
Fog-Bank says a new picture arrives every 20 to 60 seconds. The gap is not random. Below is the median gap between frames on summer evenings (14 July to 9 August) and on September evenings (10 September to 2 October). Switch the clock for the sky:
By the clock the two steps sit far apart, because September's evenings end about 1¼ hours earlier. Line them up on the end of civil twilight, when the sun is 6° below the horizon, and they fall on top of each other. Uploads stay quick, a frame every 21 to 26 seconds, until twilight ends. Five to ten minutes later the median jumps to 52–66 seconds and stays there all night. In the morning it runs backwards: the gap falls from about 60 seconds to 26–32 within 10 to 15 minutes of the start of civil twilight.
Why, I can only guess. A webcam in the dark usually exposes for longer, which slows each capture. That is a guess, not something these timestamps can show. Midday is slow too (median 57 s at noon), and the light curve does not explain that.
Around the clock
A sunrise that keeps clock time
Fog-Bank's open API compares "measured first light versus astronomical sunrise" for every day. I checked its sunrise and sunset against the US Naval Observatory for the campus (37.7233°N, 122.4797°W):
Two things are off. The midpoint of Fog-Bank's sunrise and sunset sits at 13:06 PDT on every date, while real solar noon drifts from 13:16 in July to 12:59 in October; that drift is the equation of time. And Fog-Bank's days come out 13 to 18 minutes too short. So its "first light minus sunrise" figures are off by up to a quarter of an hour, in a direction that changes with the season.
How many pictures since 1994?
The home page says "roughly 50 million". The API's own estimate is 45,414,642. Its method: seconds since 1994 divided by a cadence of 20 s up to the 2019 rescue, then by the current median of 37 s. I reproduce it to within 0.003%. But a count is time divided by the mean gap, and the gaps are lopsided: one in eight is longer than two minutes. Those long gaps fill 37% of the time.
| Assumption | Pictures since 30 Sep 1994 |
|---|---|
| A frame every 20 s throughout | 50.5 million |
| Fog-Bank: 20 s to 2019, then the median gap | 45.4 million |
| Frames per working day: 2019 rescue days (median 4,067), 2026 days with no outage (median 1,330) | 40.5 million |
| Mean gaps: 22.1 s (2019 rescue), 64.0 s (2026 days with no outage) | 39.1 million |
The bigger uncertainty is upstream. Every row assumes the camera ran in 1994–2019 as it did during five weeks of 2019, and nobody kept the frames that would show it. The camera now makes about a third as many pictures a day as it did then.
What stands on what
| Claim | Source |
|---|---|
| Frame timestamps, day list, sunrise/sunset fields, published figures | Fog-Bank open API (/api/frames, /api/days, /api/forensics, /api/figures), fetched 3 Oct 2026, 07:13 UTC. Data CC BY-SA 4.0. FogCam, San Francisco State University, archived by Fog-Bank (fog-bank.org). |
| Official sunrise, sunset, civil twilight, solar transit | US Naval Observatory API, one-day queries for the campus coordinates |
| Campus coordinates | Wikipedia, San Francisco State University |
| Twilight for every day | My implementation of the Almanac for Computers formula, checked against USNO to within a minute on two dates |
| Medians, means, the twilight alignment, the estimates | My analysis of 79 days (14 July to 2 October 2026) and 34 rescue days in 2019. Gaps over an hour (outages) are left out of the charts. |
| File size and gap | A sample of 274 frames on 30 Sep and 1 Oct: larger frames go with longer gaps (r = 0.56), but sizes vary by about 25% while gaps vary 2.5-fold, so size is at most part of it |
| Origin story, the 2019 rescue (126,835 frames) | Fog-Bank history page and figures; Hackaday, 1 Oct 2026 |
This is a friendly check of a fine project. Fog-Bank publishes its own methods, which made all of this possible.