How to pick a reference region for validating a satellite fleet count

Open wheat field stretching to a tree line and distant power lines under a pale sky.

Before you roll a satellite-derived fleet count into a market-share denominator you present to the board, you need to check it against something you already trust. That's the reference region: a smaller geography where you have a real count to compare against, so you can see whether the scaling method is reading the fleet correctly before you lean on it for a whole state or province.

Picking that region badly is how teams end up defending a number they can't back up. Here's what the choice comes down to.

What makes a region usable as ground truth

The core requirement is a count you'd stake your job on, independent of the imagery. That usually means one of a few sources: dealer registration records for a defined territory, state or provincial equipment registration data where it exists, or an association count that's audited rather than self-reported. If your "ground truth" is itself a dealer's estimate pulled from memory, you're not validating anything, you're comparing two guesses.

Size matters more than people think. Too small, and you're validating against a handful of machines where one missed unit swings the error rate by ten points. Too large, and you can't audit it by hand when the numbers don't match, which defeats the purpose of using it as a check at all. A single dealer territory, a county, or a cluster of three or four adjacent counties tends to be the right scale: big enough that occlusion and coverage effects show up, small enough that someone on your team can go unit by unit if the first pass looks off.

The region also needs to look like the places you want to scale to. If your end goal is a regional installed-base figure across mixed row-crop and pasture land, don't validate against a dense urban-fringe market where equipment sits in tight yards and under cover. A reference region with the wrong terrain mix, canopy cover, or storage density will tell you the method works when it's really just working for that specific landscape.

Avoiding the biases that ruin a validation

One mistake shows up constantly: using a brand's own strongest dealer territory as the reference region. Dealer records there tend to be meticulous because the brand dominates and has every reason to track penetration closely, which stacks the deck in that brand's favor before you've even run the comparison. Pick a region where no single OEM has an obvious home-field advantage in record quality, so the comparison isn't flattered by one dealer's unusually clean books.

Timing is the other place validations fail. If your ground-truth count is from last season and your imagery is from this spring, machines bought or sold in between will read as method error when they're really just fleet turnover. Match the imagery date to the registration snapshot as closely as you can, or at least know the size of that gap before you judge the result.

Finally, don't pick a region solely because the imagery is good there. A validation that only works under ideal conditions, clear skies, flat terrain, no canopy, won't tell you anything about how the method holds up across a real multi-state region where some counties have tree cover and some don't. If anything, lean toward a reference region with a bit of the messiness you'll hit at scale: some wooded parcels, some dealer lots packed tight, a mix of urban-adjacent and rural geography. That's a harder test, and a harder test is the one that actually earns your trust.

Run the comparison, look at where the satellite count overshoots or undershoots the ground truth, and figure out why before you ever extend the method past that one region. That's the whole point of picking a reference region in the first place: it's cheaper to find the method's blind spots in a county you can audit by hand than in the regional total you're about to put in front of the board.

If you're building that denominator and want a second opinion on a region-by-region installed-base count, that's the specific problem we work on.

Request early access

← Back to the blog