Canopy Impact

Is it working?

Every protected area is compared with the 10 km ring of land just outside it. If cropland and buildings grew more slowly inside the boundary than next door, protection is holding. If they grew faster inside, something is wrong.

Loading...

First leg: land use

What the land-use comparison finds

Cropland and built-up ground from Dynamic World, across the cohort, with every park that cannot be compared honestly held out and counted here rather than quietly dropped.

Park by park

Where the boundary held, and where it did not

Sorted by the size of the difference. Bars run left where pressure grew more slowly inside the park than outside it, and right where it grew faster.

grew more slowly insideno differencegrew faster inside

Showing the 20 largest differences in each direction. The full cohort is in the table below.

Full results

All compared parks, on both legs

Every park with a published result on either leg. Click any column heading to sort. A park read by one leg and not the other says why in the other leg's cells rather than leaving them blank, and the two differences are never combined into one number.

Land use (cropland and built-up ground) Forest loss (canopy)
Park Country Inside Ring outside Difference Reads as Confidence Ring match Difference Reads as

The two differences are measured in percentage points of different things: land use in points of park area under cropland or building, forest loss in points of that unit's own baseline forest. They are not subtractable and are never subtracted.

Held out

Parks with no honest comparison

These parks are not scored as zero and not quietly dropped. Each one is listed with the reason the comparison would have been misleading.

Land use

Method

How this is measured

Take how much cropland and built-up ground grew inside the park. Take how much it grew in the ring outside over the same years. Subtract the second from the first. That difference is what this page publishes, in percentage points of area, and a negative number means the inside grew more slowly than the outside.

The arithmetic, and what it is called

The design is a difference-in-differences, the standard way to ask whether an intervention changed an outcome. Two places, two dates:

difference = (pressure inside, later − pressure inside, earlier) − (pressure in ring, later − pressure in ring, earlier)

The control

The ring is a 10 km band of land running outward from the boundary. Every neighbouring protected area is cut out of it first, so the park is compared with unprotected ground rather than with more park. Where that leaves too little ring to read, the park is held out instead.

Where the ring is still part of the park

Sometimes the boundary we hold is one unit of a much larger protected area, and the rest of that area reaches into the ring, where it is left in place. The ring is then partly the park, which pushes the difference toward zero and makes protection look weaker than it is. Past a tenth of the ring, the park is held out. Some parks are affected. How much overlap a park has does not predict which way its difference runs.

Holding those parks out makes the whole cohort read better, and we are telling you because it does.

Why they are held out rather than corrected, and why this does not make the page conservative

A polygon that covers most of the park is treated as the park itself and is not subtracted, which is right: a park must not be masked out of its own ring. Ahaggar is the clearest case. The boundary we hold covers 21,141km², and it sits inside a protected area of 542,456km² carrying the same name, so the entire ring around it is protected land.

A ring that is partly the park behaves like the park, so the two sides move together. Above the threshold there is nothing left to compare: a park differenced against itself is not a conservative result, it is not a result. Those parks are held out, exactly as a park with a circle for a boundary is held out. Nothing is adjusted, re-weighted or corrected, and no number is moved. What survives is published band by band, so you can see whether the residual overlap tracks anything: the surviving bands are published.

It does not make this page conservative. That claim needs both known biases sized, and neither is. This one understates protection; parks being sited on high and remote land overstates it. We have measured the direction of both and the magnitude of neither.

The fix on the data side is to anchor each park to the protected-area record that actually contains it, which changes the boundary itself and so reaches every page that uses it. That is open work, and it is tracked rather than done quietly here.

Whether the ring is a fair control

The design assumes the ring shows what the park would have looked like unprotected. There is a good reason to doubt that. Parks tend to sit on high, steep, remote land nobody wanted to farm anyway, while the land just outside is flatter, easier to reach and more populated. That tilts this comparison towards making protection look good.

So we measure the gap rather than assume it away, and publish it per park without adjusting anything.

What is measured, and why nothing is corrected for it

The pattern is documented in the protected-area literature as "high and far". For each park, four things are read over the park and over the same ring the difference is taken against: ground height and steepness from a digital elevation model, travel time to the nearest city, and modelled population density. Each becomes a standardised mean difference, the conventional balance statistic, published per park and summarised for the cohort above.

It is disclosed, not corrected. The field standard for handling this is covariate matching, which pairs protected land with comparable unprotected land before differencing. We do not do that, and we do not publish an adjusted number that pretends we did. Every difference on this page is the raw comparison. What has changed is that you can now see which parks have a credible control and which do not.

What gets held out, and why

That last guard removes results that would have been flattering. It is the reason to trust the ones that remain.

Known limits

Each measure rests on one satellite product. The land-use product mistakes dry grassland for farmland in some years, which is why some parks are held out. The forest product has the same weakness on fire and storm damage. Where a product is wrong about a place, that measure is wrong about that place. Having two helps because they are wrong about different places, not because either is safe alone.

We cannot show the two sides started out alike. The readings begin in 2019, decades after most of these parks were created, so there is no way to check. What we can check is whether a halfway reading in 2021 agrees with the direction each side ended up going, which catches a park flattered by two lucky dates. That is a smaller claim, and it is the one the data supports.

These numbers changed once we found the ring was partly the park. An earlier version scored 139 parks; 27 had a ring more than a tenth their own protected land and are now held out. The cohort went from 39 working of 139 to 36 of 112. Removing them makes protection look better, and we are telling you the direction because it flatters us, not despite it.

Two known biases, pointing opposite ways, and neither is sized. Protected land left inside the ring understates protection. Parks being sited on high, far, unwanted ground overstates it. We have measured neither one's magnitude. So nothing on this page may be called a conservative estimate: that would mean assuming one bias beats the other, and we do not know that it does.

Two dates are a before and an after, not a trend. Each difference is a park’s first reading against its last, and says nothing about the path between them. Fitting a line through every year instead was tried and moves almost nothing, so the simpler version stands. The in-between years are used to measure how much each park’s reading wobbles, set out above.

A difference is not a cause. This can show that pressure grew more slowly inside a boundary. It cannot say which patrol, community agreement or budget line did it, and it does not try to.

Part of the ring is sometimes part of the park. Where a park's boundary is one unit of a larger protected area, the rest of that area is read as the park and stays in the ring it is compared against. Past a tenth of the ring the park is held out; below it the share is measured per park and published beside the number. Set out above.

Four of the largest parks have no halfway reading. Termit, Ouadi Rime-Ouadi Achim, Namib-Naukluft and Skeleton Coast are too large to compute at full resolution. Rather than measure them more crudely, which would make them incomparable with every other year, they are marked lower confidence and the gap is left visible.

What this still needs

These would make this materially harder to argue with, in the order they are worth doing. Where one has since been done the entry says so and links to the answer, including where the answer was negative.

  1. A second opinion, from tree cover. This one is done, and it is above. A different satellite record, measuring loss of trees rather than spread of farmland, and wrong in different ways. It reaches back far enough to look at a park and its neighbour before the comparison window opens, which is a check the land-use leg cannot run at all. What it has not done is confirm the first leg: on its own terms its cohort and its placebo are not separated, so it is published as a second reading with its own limits rather than as corroboration of the first.
  2. What each park costs to run. No spending figure for any individual park exists here yet. With it, the question changes from whether protection worked to whether it was worth what was spent, which is what funders actually need to decide. This is waiting on operators sharing real numbers, not on any technical work.
  3. Fire, if it can be handled honestly. Savannas need fire, and some burning is deliberate management. A naive comparison would read good management as failure. Either that gets designed around properly or fire stays out, with the reason stated.
  4. Matching each park to comparable land. This one is done, it is above, and it failed. Measuring how far each park differs from its ring was the cheap version, and it came back with an awkward pair of answers: the imbalance is large, and it does not explain the results. The full version, pairing each park with genuinely comparable unprotected land before differencing, has since been built. It does not rescue the per-park question, because the estate does not contain the land to match against: protected ground in Africa is largely the high, steep, inaccessible ground, and there is very little unprotected ground like it. No published effect moves. The attempt is reported in full rather than dropped, because what it rules out is worth as much here as a result would have been.

Greenness, surface water and night lights were considered and dropped. The first two mostly track rainfall, so a wetter strip of land beside a drier park would show up as a protection effect, which is nonsense. Night lights are near zero inside protected areas, leaving nothing to compare. Dropping them removes results that would have needed so much qualification they would have meant nothing.

Where the numbers come from