What is a revision policy, and why your land-cover figure needs one

Every national statistical office that publishes a CPI or a GDP series has a revision policy sitting somewhere in its methodology notes: a schedule for when preliminary figures get superseded, a rule for how far back a correction propagates, a public log of what changed and why. Land-cover change statistics rarely get the same treatment. A figure goes into the SDG 15.1.1 forest area ratio or the national land degradation baseline, and then sits there, unrevised, until someone notices the underlying classification was wrong for three reporting cycles running.
A revision policy is the set of rules that governs what happens when new evidence changes an old number: which years get reopened, how far back the correction reaches, who approves it, and how the change gets disclosed to whoever cited the original figure. For land-cover change, that's not a nice-to-have. It's the difference between a time series your reporting partners trust and one they stop citing after the second unexplained revision.
Why land-cover figures need revising at all
A land-cover change statistic is a comparison between two classified images, usually taken years apart, run through algorithms that have gotten measurably better since the baseline year was set. Cloud masking has improved. Sensor calibration has shifted. The training data used to tell shrubland from early-regrowth forest has grown. None of that is a scandal. It's just what happens when a multi-decade dataset keeps absorbing better inputs.
The problem is that most land-cover programs treat the first published number as permanent, because reclassifying a decade of baseline imagery by hand is its own multi-month study, and nobody budgets for redoing work that's "already done." So a ministry ends up reporting one classification lineage for the 2015 baseline and a different one for 2023, stitched together at the seam, and calling the resulting trend line a time series. An auditor, or a UN agency doing cross-country SDG review, will find that seam.
Backcasting instead of patching
The fix statisticians reach for in every other domain is backcasting: reprocessing the entire historical series through the current method every time the method changes, rather than patching only the newest year. For a land-cover layer, that means re-running the classification on 2010, 2015, and 2020 imagery with the same model and the same spectral bands used for this year's figure, not bolting a corrected 2023 number onto an uncorrected 2015 baseline. The output is a single internally consistent lineage: every year in the series was produced the same way, so a change between any two years is change on the ground, not change in the measuring instrument.
Backcasting only works if the underlying pipeline is annual and automated. A one-off classification study commissioned for a single reporting cycle has no mechanism for going back and reprocessing five prior years when the method improves; the contractor delivered the deliverable and closed the project. A pipeline that refreshes every year, on the same wide-swath multispectral base, can rerun the archive because the archive and the method live in the same place.
What a revision policy actually says
Write one down, even a short one, and three things need to be in it. A trigger: what counts as a reason to reopen a published year, such as a sensor recalibration, a training-data update, or an error report from a provincial office. A depth: how many years back a given trigger reaches, since not every fix justifies redoing the whole archive. And a disclosure line: a dated note in the published table saying this year's 2018 figure differs from the 2018 figure cited in last year's report, and why.
That third point is the one environment ministries skip most often, and it's the one that actually protects credibility. A revised number with no note looks like an error. A revised number with a one-line footnote looks like a statistical office doing its job.
If you're assembling the land-cover change layer behind an SDG 15.1.1 or 15.3.1 figure and want a lineage that survives a revision without a seam, Habitat Loss Mapping builds that layer as a single annually-refreshed series rather than a stack of one-off studies. Get in touch if a defensible, revisable national time series is what the next reporting cycle needs.