What it is
Land-Use is one of Keystone Watch's satellite feeds. The others raise alerts when something happens this week (fire, forest loss, vegetation, surface water, night-lights) or add seasonal context (rainfall, burned area, biomass). Land-Use is the annual, structural read: what is happening to the land itself, year on year.
A fire is an event. Turning a wilderness core into farmland is a conversion, and it is often what quietly kills a protected area over years while the weekly alerts stay calm.
It appears as a card on each park page and in Keystone Watch: one plain-English verdict (“No clearing or new farmland”, “Dry grass, not cropland”, “Confirmed: land being farmed”) with the evidence behind it.
The hard problem
Anyone can download free satellite imagery. The trap is what that imagery looks like in savanna.
The savanna false positive. African savanna spends half the year as green grass and half as dry, brown grass. To an automated land-cover classifier, dry grass looks like a harvested crop field. Flag everything the classifier calls cropland and half of Africa lights up every dry season. No one can act on that.
So the whole task is telling dry grass from cropland. That is what this feed does, and why off-the-shelf tools do not work here.
Why it exists
The feed was built to answer one question: can African protected areas be watched from orbit for real, human-driven conversion, without drowning in the savanna false positive? The method at its core, spectral unmixing, is a standard, peer-reviewed technique. It runs on every keystone park today.
How it works: five layers, each allowed to overrule the last
A cropland map raises the flag. A second read of the same ground checks for bare, tilled soil and decides whether the flag is real. A forest-loss map catches clearing the cropland map cannot see, radar covers the parks under cloud, and a map from another team checks the result.
Google Dynamic World asks: is cropland or built-up ground spreading? It is a 10-metre land-cover map, and the feed compares the park's interior in an early year with a recent one (currently 2019 vs 2024). If the share of the interior read as cropland or built-up ground has risen by at least 1 percentage point, the flag goes up. On its own that flag is never a verdict, because in savanna the map flags dry grass.
Sentinel-2 asks: is that bare, tilled soil, or just dry grass? Every patch of ground reflects light as a mix of three signatures: green plants, dry plants and bare soil. The second read works out how much of each is present (spectral unmixing, in the jargon). A ploughed field exposes bare earth; a field of dried grass does not. So the rule is a bare-soil floor of 3%: below it, the flagged ground is grass and the flag is stood down.
Hansen Global Forest Change catches forest cleared to grass or settlement, which the cropland map cannot see. It is the standard annual map of tree-cover loss. Loss counts only where there was real forest to begin with (25% canopy or more) and only where the trees have not grown back in the latest imagery. This layer recovered the four forest-clearing cases the cropland map had missed.
Why regrowth is the key test, and how natural loss is told apart
A savanna fire or a storm gap regrows within a season or two and is not counted; a cleared plot stays cleared. The test works even in the humid tropics, where a cleared plot is as green as the forest it replaced. Loss that lines up almost entirely with recorded fire is set aside as the natural burn cycle, and a documented disaster (Cyclone Freddy’s landslides on Mount Mulanje) is excluded by name.
Forest can also be lost naturally and re-cover as open savanna. So a flagged loss gets its own second read over the cleared ground: where that ground reads as living dry grass and an independent land-cover map agrees it is open savanna, the loss is shown as trees lost, grown back as savanna rather than raised as a watch. The check only ever adds an explanation; it never hides a loss.
Sentinel-1 radar covers the parks where the cameras are blind. Optical satellites cannot see through cloud. Radar bounces its own microwaves off the ground and can. Where the optical read is unusable, radar gives a lower-confidence read, labelled as radar only. It is never presented as agreement between the optical reads.
ESA WorldCover, a cropland map built by another organisation from other satellites, is compared against every verdict. It agrees with Canopy on 88.8% of parks, and it puts the parks called grass at near-zero cropland. This is agreement between two maps, not proof that either one is ground truth.
What comes out: one verdict per park, in plain words
The layers combine into one verdict per park, and the card under it shows the evidence. These are the words every card and chip uses, so a verdict reads the same wherever it appears.
Tarangire, worked through
Tarangire shows the second read doing its job. The cropland map flagged a large slice of the park as crop-like, the second read found almost no bare soil under it, and the verdict was dry grass, not cropland.
The numbers, the dry-year twist, and what else the card shows
The independent WorldCover map agrees: it sees just 0.324% cropland there. A naive tool would have raised a false alarm on 14% of the park. In the dry year that followed both reads moved together, but because the rise sits almost entirely in that final year, a persistence guard holds the verdict at possible clearing, not confirmed.
Beyond the verdict, the card shows its working:
- A corroboration ratio, bare soil divided by the crop flag. Tarangire's ~1.8 over 14 (its 2019–2024 reading) visibly says “mostly grass”.
- An agreement line, “Independent checks: N of M agree” (for example “2 of 2 agree”, naming the second read and WorldCover), so you can see how many separate lines of evidence line up.
- A full audit trail: every verdict carries its datasets, dates, geometry source and thresholds, so a reviewer can trace exactly how it was produced.
What the validation found
The feed was checked against park interiors labelled by eye, plus documented conversions. It caught 88.2% of the labelled conversions (15 of 17), and it was right on every site (100%) where it called dry grass rather than cropland. The two conversions still missed, mining pits in Sapo and oil pads in Murchison Falls, are too small and non-forested for a tree-cover read. Precision is about 54%, because the watch tier flags anything worth a look on purpose.
How the set was built, what is still missed and why, and what the forest-loss layer costs
Fifty-five park interiors were hand-labelled from 2019-versus-2025 imagery. Six more were added as documented conversions, interiors where an independent source (investigations, NGO monitoring, peer-reviewed work) records new conversion between 2019 and 2025: illegal gold mining in Sapo, oil-project pads and roads in Murchison Falls, displaced-population clearing in Virunga and Kahuzi-Biega, and farming and settlement expansion in Niassa and Bale Mountains. The set has since grown to 87 interiors, all 87 now labelled; 17 too ambiguous to call were excluded, leaving 70 rated, 53 intact and 17 converted.
The nine-park challenge set held: interiors WorldCover reads as cropland but the feed cleared were confirmed not converted (five intact, four too cloudy to call), so the disagreements were about cropland presence, not missed change.
Recall was first measured at 1 of 7. The cropland map alone had caught only the one conversion it had already surfaced. Every one of the six misses is a localized conversion: mining pits, oil pads and access roads, dispersed fields around interior villages, scattered settlement, or forest clearing by displaced populations. Each is real but too small to move the whole-interior fraction this annual read measures, or is forest loss the cropland map cannot see. A threshold sweep found no setting that recovers them, and a test of sharper resolution, sub-tiling each interior and adding a radar and a per-pixel change read, did not recover them either. The limit is the detection floor of the underlying data at these sites, not thresholds or resolution. Four of the six, Virunga, Kahuzi-Biega, Niassa and Bale Mountains, were forest clearing, and the forest-loss layer recovered them, lifting recall to 88.2%.
It is not free. Reading forest loss at whole-park scale also flags parks labelled intact that carry real tree-cover loss (the Kilombero valley, the Mount Gorongosa sector, a smaller Rwenzori loss), whose signal is as strong as the true misses. A footprint check now clears the ones that are natural savanna re-cover: it reads the cleared ground and requires an independent land-cover map to agree it is open savanna. That stood down four intact false positives (Banhine, Zinave, Lower Zambezi and Mangochi) while keeping the real conversions flagged. A desert-pan check does the same at the other extreme, standing down barren salt pans misread as built-up (Etosha, Kgalagadi). The rest are shown as a forest-loss watch, a prompt to look, and those labels are being re-checked. A fuller adjudication of the once-ambiguous alerts confirmed more of the low-confidence watch flags as intact, which is why precision fell from an earlier 71% on the smaller labelled set, with the miss rate unchanged.
WorldCover independently agreed on 88.8% of 190 parks. The methods are not homegrown: Dynamic World for the cropland map, and the established, peer-reviewed literature on separating green plants, dry plants and bare soil for the second read.
The caveat: recall rests on a still-modest sample (17 labelled conversions). The six documentation-derived sites have since been confirmed in imagery, so they are genuine conversions the data cannot resolve, not mislabels. At this size it is a first estimate, not a gold standard, and the point estimates carry wide error bars. The direction is clear: the feed's validated strength is that it does not misread grass and holds its clears against an adversarial check; its measured weakness is missing localized, non-savanna conversion.
What it is not
- It never moves a score. Land-Use is context only: never an input to PACE, Canopy's park rating, and never a Keystone Watch alert. It informs judgement; it does not grade.
- It abstains rather than guesses. No data means “not enough clear imagery”, not a made-up number. Three parks still abstain even after radar: Central Kalahari, Ouadi Rime-Ouadi Achim, and Termit et Tin-Toumma.
- It ranks parks; it is not a soil map. The green, dry and bare split reliably orders parks. It is not an exact percentage of soil.
- The watch tier is sensitive, not precise. It over-flags on purpose: none of the low-confidence watch flags in the validation set is a confirmed conversion. They are prompts to look, and they are labelled that way.
- Its figures are a first estimate. Recall is 88.2%, up from a first measurement of 1 of 7; the two still missed, mining in Sapo and oil infrastructure in Murchison Falls, are too small and non-forested for a tree-cover read. Precision is about 54%, revised down from 71% as more of the once-ambiguous alerts were adjudicated. Both rest on a small, expert-judged sample, so read them with wide error bars.
- It is annual, not real-time. Permanent conversion is best confirmed over several clean years, not chased quarterly.
What’s coming: reading shape, not just colour
Everything above is live today; this is not. It is specified but not yet built, on the order of one to two months of work, and it is designed to run alongside the current method for a full season rather than replace it.
Every layer above reads colour, the light a patch of ground reflects, and that is exactly where dry grass and tilled soil look alike. Shape is a second, independent clue. A crop field is a made object: straight edges, corners near ninety degrees, boundaries shared with the next field. A natural clearing has none of that.
The design, a worked example, and what has to be settled first
Such regularities resist hand-written rules but suit a supervised deep-learning model: a segmentation network trained on labelled imagery to draw field boundaries, and in effect to learn what a human-made parcel looks like. It would not replace the second read. It would add another independent line of evidence to the same verdict, drawn from the shape of the land rather than its colour.
Worked example
Two interior patches both read as bare and brown. By colour they are near-identical; by shape they are separable.
| Criterion | Patch A | Patch B |
|---|---|---|
| Colour (spectrum) | Bare soil, brown | Bare soil, brown |
| Edge geometry | Irregular | Rectilinear |
| Corner angles | None near 90° | Four near 90° |
| Shared boundaries | None | Three |
| Classification | Natural clearing | Cropland |
Illustrative, not a result. The colour row is identical across both patches; every shape criterion separates them. That separability is the whole rationale for the approach.
What has to be settled first
This is the only part of the feed that needs a bespoke model rather than an established product, which brings three dependencies at once: labelled training data, GPU compute, and time to validate. Two constraints in particular must be resolved before committing to it.
- Domain shift. A public benchmark for this task, Fields of the World, exists, but datasets of its kind lean towards European and North American farmland and under-represent African smallholdings, whose parcel size and shape differ markedly. Whether a model trained on that distribution works on the parks we monitor is an open question, most plausibly answered by in-region labelling or domain adaptation.
- Seasons, not snapshots. A single image cannot separate a ploughed field from a burn scar or a harvested plot; each is bare soil at one moment. Telling them apart needs a sequence across seasons, so the model reads a trajectory rather than a single frame.
Running it alongside the current method, rather than in place of it, is deliberate: two independent methods that agree are a validation, and where they diverge, the divergence is itself the finding. That is the standard the layers above already hold themselves to.
Where it fits in Canopy
Keystone Watch monitors; it does not rate. Land-Use is its year-on-year layer, reading the ground inside the boundary while the other feeds watch for weekly events. It is kept strictly separate from PACE, the park rating, and from the alert feeds. Every verdict is flag-and-review with an audit trail, never a black-box number.
Common questions
How is this different from Global Forest Watch or existing deforestation alerts?›
Those track tree-cover loss, and they are built for forests and for events. Land-Use tracks conversion to human use inside savanna and mixed landscapes, where most African keystone parks are, and where a raw tree-loss alert is blind (no trees to lose) or noisy (every dry-season fire). The savanna false positive is what makes generic cropland maps unusable here, and it is the problem this feed is built around. A guarded forest-loss read is folded into the verdict for the one case the cropland map cannot see, forest cleared to grass or settlement: only inside real forest, only where it has not grown back, and with fire and documented natural disasters filtered out. That is what separates it from a raw deforestation alert.
Why annual, not real-time? Isn't faster better?›
Not for this signal. A fire is an event you want in hours; permanent conversion is a multi-year trend you want to confirm, not react to on a single cloudy image. Annual composites are more correct and far less noisy. The five alert feeds cover the fast events; Land-Use is deliberately the slow, structural read.
How do you know you're not missing conversions (false negatives)?›
It was measured, then improved in the open. The cropland map alone caught only 1 of 7 (about 14%) in the first measurement: six misses, all localized (mining, oil infrastructure, dispersed fields, scattered settlement) or forest clearing the cropland map cannot see. An independent forest-loss layer was added for that forest-clearing gap, and on the current, larger validation set recall is 88.2%; the two still missed (Sapo mining, Murchison oil pads) are too small and non-forested for a tree-cover read. The trade: whole-park forest-loss detection also flags intact-labelled forest parks with real loss (Kilombero, Gorongosa, Rwenzori), whose labels are being re-checked; the footprint and desert-pan checks then cleared the natural savanna re-cover cases (Banhine, Zinave, Lower Zambezi, Mangochi) and the salt-pan cases (Etosha, Kgalagadi) while keeping the real conversions flagged. A fuller adjudication of the once-ambiguous alerts confirmed more low-confidence watch flags as intact, so precision is about 54% (revised down from 71% on the smaller labelled set). What held throughout: all three checks, and the clears on a nine-park WorldCover-disagreement challenge set, all still 100% correct. The sample is still modest (17 labelled conversions, expert-judged), so the point estimates carry wide error bars. Sub-tiling for sharper resolution was tested first and did not recover the misses, which is why the fix was a new signal.
What's your false-positive rate? How do you avoid crying wolf?›
The second read exists for this. It was correct on every grass site it stood down (100%), including Tarangire, where 14% “crops” became dry grass. The low-confidence watch tier over-flags on purpose: none of its flags in the validation set is a confirmed conversion, which is why those verdicts say “worth a look”, not “confirmed”. The confident tier requires two independent sensors agreeing over multiple years, and it has been stress-tested three times: a tiny palm-forest island (Vallée de Mai) once reached it on a mis-recorded area, now fixed; a savanna dry-year spike (Tarangire, 2025) reached it when one dry season pushed both reads at once, now held by a persistence guard to “possible clearing, not confirmed”; and a wetland (Ichkeul) whose cropland read dipped far below its own starting point in between and was held back by a series-instability guard until the dip turned out to be ours: half its years had been read over a circle drawn round the park and half over the true boundary. Re-read over the boundary alone, the dip is gone, the guard has nothing to hold, and Ichkeul is the one park the confident tier holds.
Why trust a “relative index” instead of an absolute measurement?›
Because for the decision that matters, is this park changing, and more than its neighbours?, a well-ordered relative index is the right tool. Bare soil and dry vegetation are near-identical to an optical satellite, so claiming an exact soil percentage would be false precision. The feed ranks and flags; it does not pretend to a lab measurement.
How small a change can you detect? What's the resolution?›
The cropland map is 10-metre pixels. It starts to flag at a 1 percentage-point rise in the crop-or-built share of the interior (high at 3pp). So the feed is built to catch meaningful, spreading conversion, not a single field, which is the right altitude for a park-scale early warning.
What about cloud? Half of Central Africa is under cloud.›
That is why there is a radar layer. Sentinel-1 sees through cloud. On the 11 park interiors the optical pair could not judge, radar cleared 8 as quiet, found 0 conversions, and 3 still abstain. Radar reads are labelled as lower-confidence, single-sensor reads.
What if a data source disappears?›
No single source is load-bearing. The cropland map (Dynamic World), the second read (Sentinel-2, ESA-run), the forest-loss layer (Hansen Global Forest Change), the fallback (Sentinel-1, ESA-run) and the check (ESA WorldCover) are five different products from four providers. The second read, the actual method, is Canopy's own analysis of raw Sentinel-2 imagery, a public, funded, long-term European mission. The cropland map could be re-implemented on other land-cover products if needed.
How many parks does it cover, and can it scale beyond Africa?›
It runs on the 162 Africa Keystone parks, with verdicts computed across ~190 park interiors. The method is not Africa-specific; the second read is calibrated for the hardest case, savanna, so it generalises. Scaling globally is a compute-and-calibration exercise, not a new invention.
Numbers and facts, at a glance
- What: the annual land-use conversion feed, “is this park being converted from the inside?”
- The trick: a bare-soil floor of 3% separates real tillage from dry savanna grass.
- Coverage: 162 Africa Keystone parks; verdicts across ~190 interiors; updates annually.
- Five inputs, four providers: Dynamic World (cropland map), Sentinel-2 (second read), Hansen (forest-loss layer), Sentinel-1 (radar fallback), ESA WorldCover (independent check).
- Validation: all three checks (dry grass, natural re-cover, desert pan) 100% correct on their labelled cases; nine-park challenge set held; recall 88.2% after the forest-loss layer (localized mining and oil still missed); precision about 54% after a fuller adjudication of the once-ambiguous alerts; WorldCover agrees 88.8%.
- Tarangire example (2019–2024): 14% “crops” became 1.8% bare, so dry grass, not cropland (WorldCover 0.324%); in the 2025 dry year its bare share rose past the floor and the persistence guard held it to possible clearing, not confirmed.
- Limits: context only, never moves a PACE score; abstains rather than guesses; the watch tier over-flags on purpose (a prompt, not a verdict).
- See it live: the card on any park page, and in Keystone Watch.