How many days of imagery before a site baseline is reliable

Every analyst who's built a baseline from scratch has hit this question on day three: is this enough yet, or am I still looking at noise? There's no single number that applies to every site, but there's a way to reason about it that holds up across target types.

It's an observation count, not a calendar count

The instinct is to ask "how many days" because that's how tasking gets planned. What makes a baseline trustworthy is the number of independent looks you've accumulated at the site in a settled state. A site collected daily for two weeks with clean VHR passes gives you fourteen real data points. A site collected every three or four days because of weather or tasking priority gives you far fewer over that same two weeks, even though the same amount of time has passed.

Daily cadence matters for a specific reason: it compresses the calendar time needed to reach a given number of observations. That's what lets you stand up a baseline on a newly tasked site without waiting a month to say anything useful.

What sets the minimum observation window

Three things drive how many observations you need before a baseline stops moving around on you:

Site tempo. A parking lot at a logistics depot that turns over daily needs fewer looks to characterize than a site with a weekly delivery cycle or a facility that only shows activity around a monthly maintenance window. If the site runs on a seven-day cycle, you need at least one full cycle of good passes before you can say you've seen "normal," and ideally two or three cycles before you trust the edges of that normal.

Confounders in the collection record. Cloud cover, off-nadir angles that distort shadow and vehicle counts, and gaps from tasking conflicts all add noise that a short run of imagery can't average out. A ten-day window with three unusable passes gives you seven real observations, and seven might not be enough if the site has any weekly pattern at all.

How much the site varies day to day on a good day. A static facility with little vehicle or equipment movement stabilizes fast. A site with constant construction, rotating equipment, or seasonal storage swings needs a longer run before the baseline reflects the real range of normal rather than whatever happened to be parked there during week one.

A working rule of thumb

Most analysts converge on a practical floor: enough consecutive good observations to cover at least two full cycles of whatever rhythm the site runs on, plus a handful of extra days as a buffer against bad passes. For a site with no obvious weekly or monthly pattern, two to three weeks of usable daily imagery is a reasonable starting point to call a baseline provisional, meaning good enough to start flagging deviations while you keep refining it in the background. For a site tied to a longer cycle, a weekly resupply run or a monthly inspection, you want to see that cycle repeat at least twice before you trust the baseline's lower and upper bounds.

The key word there is provisional. A baseline isn't a one-time calculation you finish and file away. It should keep absorbing every new pass, tightening its sense of normal as more observations come in, and getting more precise about what counts as a real deviation versus a once-off. A snapshot comparison stops there; a working pattern-of-life product keeps going, so the baseline is never really "done." It's just continuously current.

If you're trying to figure out whether your own site has enough collection history yet, the fastest way to find out is to start the accumulation and watch when the flagged anomalies start looking like real signal instead of routine variation. Activity-Based Intel is built around exactly that loop, daily VHR passes feeding a baseline that checks every new look against the pattern it's already learned.

Start logging daily passes on your site now, and you'll have a working answer to this question sooner than you'd expect.