How does a satellite picture become a number a researcher can line up with a calendar? This note walks through that conversion, using public imagery of Walmart parking lots as the example, and lists the checks that happen before the count is a series.
A count is a decision about pixels
A picture is not a time series. It is an array of measurements taken at one moment, over one patch of ground, with one instrument. To become a series, someone has to define a place, a quantity, and a rule that turns each later picture of that place into the same quantity. The quantity might be a count of cars, a fraction of a lot that looks occupied, or a change in brightness. The rule has to be stable enough that a jump in the series means the ground changed, not that the rule changed.
The instrument changes too. Clouds hide the lot. A different satellite passes at a different angle. Shadows move with the season. A raw count that ignores those conditions will show a pattern of orbits and weather. The file needs a flag for pictures that failed the visibility test, and the series needs a policy for gaps. Filling a cloudy day with the previous clear day invents a measurement. Leaving the gap empty forces the later user to notice it. Those are different datasets.
A picture becomes a time series only when a place, a quantity, and a stable rule turn each later image into the same measurement, with failures and gaps left visible.
Parking lots as a public illustration
Retail parking lots are a common public illustration because the object is easy to describe. Consider stores operated under the Walmart name. A picture taken on a Saturday and a picture taken on a Tuesday do not mean the same thing even if both are clear. The series has to know the day of week, the hour, and whether the store was open. A count of cars is not store revenue. It is a count of cars under a definition of what counts as a car and what counts as the lot.
Drawing the lot is part of the definition. A box that includes the road will count traffic. A box that includes a neighboring store will mix two businesses. A box that was right last year may be wrong after a construction project. The polygon and the date it was last checked belong with the series. Without them, a drop in the count might be a smaller crowd or a box that no longer covers the asphalt. A reviewer cannot tell which from the number alone.
Vision models can propose the count. A model may mark regions it believes are vehicles and return a tally. That tally is a proposal. A researcher accepts it as a feature only after the failures are sampled: a snowy lot, a night pass if the instrument cannot see at night, a store with a covered garage the model cannot see into. Acceptance is a written yes for a stated version of the model and a stated set of places. A new model version is a new proposal, not an automatic continuation of the old series.
Lining the series up with other facts
Once the count exists, it still has to meet other clocks. The picture has a capture time. A sales figure, if it is used at all, has a reporting lag. Joining them on a calendar date without those clocks puts a later report next to an earlier picture and calls the pair simultaneous. The point-in-time habit applies to images as much as to filings. The row should say when the picture was taken and when the derived count became available to the study.
The Walmart lot is an example of the conversion, not a claim about that retailer. A series of car counts can be flat because traffic was flat, because clouds removed the interesting days, or because the box drifted. The note beside the feature should say which pictures were dropped and which model version proposed the tally. Until that note exists, the column is still a picture, stored as a number.
This piece does not claim that parking lots measure sales, that a vision model is correct, or that imagery should be used in a strategy. It describes the conversion from a picture to a column, and the acceptances that conversion requires.
