Field notes: the first annual refresh after the satellite programme changed its sensor

The email came in on a Tuesday, the way these things always do: "Loss figures for Region 4 look off, can you check before we send to the working group." Region 4 hadn't had a fire season, hadn't had a concession auction, hadn't had anything that would explain a jump in detected forest loss that large. What it had was a new sensor on the satellite programme we'd been drawing from for six years.
Nobody tells you this when you start compiling land-cover change figures, but a sensor swap upstream is its own kind of disturbance event. The spacecraft operator announces it months in advance, usually with a calibration report and a reassuring paragraph about continuity. Then the first annual refresh lands and a chunk of your national total has shifted in a way no field visit will ever explain, because nothing on the ground moved. The ground cover is reading differently because the instrument looking at it is different.
The specific culprit varies. Sometimes it's a shift in the near-infrared band's spectral response, so vegetation that used to sit comfortably above your loss threshold now sits just below it, or vice versa. Sometimes the new sensor has a slightly finer native resolution and your classifier starts resolving field boundaries and treelines it used to lump together, which shows up as "loss" at every edge pixel in the country simultaneously. Either way the artefact has a signature once you know to look for it: it's diffuse, it tracks sensor swath boundaries rather than administrative or ecological ones, and it's suspiciously uniform across land-cover types that have no business behaving the same way.
The first thing I do in a year like this is stop treating the refresh as a single comparison and start treating it as two. One comparison covers the years before the handover, built entirely on the old sensor's record. The other starts fresh from the overlap period, where both satellites were producing usable imagery over the same ground for a few months. If the classification agrees with itself across that overlap window, you've got a baseline you can trust going into the new sensor's data. If it doesn't agree, you've found your correction factor before it gets anywhere near a published figure.
This is tedious in a way that doesn't show up in any methodology note. It means pulling tile-by-tile comparisons for the overlap months, not just national totals, because a shift that nets out to near zero nationally can still be hiding a 3% swing in one biome and the opposite swing in another. It means going back to the classifier settings and checking whether the loss threshold was tuned against the old sensor's radiometry, because if it was, you're not comparing land cover, you're comparing two different instruments' opinions about what counts as bare ground.
What goes in the report, in the end, is a footnote most readers will skip: a break in series, a note that figures before and after the transition year use harmonised inputs, a line explaining why one year's change figure carries a wider margin than the others. That footnote is the only honest way to handle a sensor swap. Pretending it didn't happen, or worse, letting the artefact ride into the headline number because the deadline was close, is how a country ends up explaining to a UN review panel why its land-cover trend has a kink in it that no satellite record anywhere else shows.
This is also, mechanically, the argument for building your change layer on a wide-swath programme that handles continuity as part of its own operations, and for treating the yearly update as a managed process rather than a fresh commission each time. We built Habitat Loss Mapping around exactly that annual, same-basis refresh, so a sensor transition is something we work through once on your behalf rather than something your team discovers in a Tuesday email.
If your next reporting cycle is due for a refresh and you want to avoid finding out about a sensor change the hard way, have a look at what a consistent annual layer looks like in practice.