Keystone Watch (The 162 Africa Keystones, watched from orbit)

Which of Africa's keystone protected areas need a look right now, each park read from space against its own history.
162 keystones, flagged on fire, forest loss, vegetation and surface water.

Not the PACE list: PACE rates how 200 parks are run, from published evidence. This page reads the 162 keystones from satellites alone, and neither feeds the other.

Methodology →
Who runs the estatefilter the board by manager
What needs a look
Land use
Portfolio trend
Evidence base

Where they are

red = needs a look · click a park
scroll or pinch to zoom for park boundaries
  • Each dot is a keystone park, placed roughly where it sits in Africa, sized by area. Zoom past 3x and a park becomes its real boundary. Click any park to open its full readings.
  • The pills above change what the colour means. All feeds is the board's own reading, red if a park is past normal on any of the four. Pick a single feed and every park is re-graded on that one alone, through the same test the list below uses, and the count on each pill is how many of the 162 that feed has actually read. A grey park is one it has not read, which is never an all-clear.
  • Only fire fills the boundary with dots. Vegetation and surface water reach this board as one reading for the whole park, so there is no located ground to draw inside it. Forest loss does have located alerts, but its scan reaches a rolling handful of parks a day, and an estate map of them would show one park's clearing and read as an estate almost free of it. Open a park to see its own forest-loss dots.
  • Keep zooming and the fires appear inside the boundary. Each small dot is a spot where satellites picked up active fire heat on the park's own ground over the last fortnight (NASA VIIRS). Size and colour are fire power, on one scale across every park, so two parks side by side can be read against each other.
  • Read the fire dots with the season in mind. Fire is a natural part of most of these landscapes and much of it is deliberate early-dry-season burning, which is a sign of management rather than of trouble. They are heat detections, not mapped fire perimeters, and the feed shows each park's strongest 150, so a busy park is showing a sample and not a count. Nothing on this board is scored from them.

What needs a look

Show every park, every feed
  • One row per park, showing its current reading on each feed.
  • A blank cell is not an all-clear. It means no usable series yet for that feed, from cloud cover or a still-building baseline.

Land-use conversion

annual · measured inside every boundary

Is cropland or clearing spreading inside the park?

Loading the land-use screen…
  • A cropland flag on its own is only a suspicion. Land-cover maps mistake dry grass, fire scars and bare ground for crops.
  • So every flag gets a second, independent read of the same ground (Sentinel-2 spectral unmixing), which separates living plants, dry plant matter and bare soil.
  • Real cropland leaves bare, tilled soil. Dry grass does not. A park flagged as cropland with no bare soil is stood down as grass. A park is shown once both reads are in.
  • Annual context only: never a score input, never an alert on this watch, and it abstains rather than guesses when a park has no usable read.
  • A separate forest-loss read catches clearing the cropland map cannot see.
  • Where cleared ground has re-covered as natural savanna, confirmed by an independent land-cover map, it is shown as forest loss, natural re-cover rather than raised as a watch.
  • A barren salt pan or gravel desert the map reads as built-up is shown as bare desert/pan, not conversion.
  • Each park's full card, with the ground-cover bar and the crop-flag map, is on its own page.

How the screen and the second read work → · Plain-English explainer →

Portfolio readings over time

measurement-of-record · deepens monthly

Each month, how many of the measured parks were within their norms, a little elevated, or clearly beyond band, feed by feed.

Within norms Elevated Beyond band each bar = keystones measured that month (n), split by reading level · faint = thin coverage
  • A bar past the norm is not an alert. It means the reading was unusual that month, not that it cleared both gates on the tab before this one.
  • Fire and forest loss rise every dry season as it sweeps the continent. That is the signal working, not the estate failing.
  • Vegetation, water and night-lights are the ones to read for whether the wider estate is holding up.
  • The window is the last 24 months. Vegetation, surface water and night-lights carry the full two years, because their record includes archive readings scored retroactively under the same code. Fire and forest loss are shorter, because theirs does not reach as far.
  • A month covering under a third of the estate is dropped, not drawn faint. Fire's pre-2026 archive holds 27 of the 162 keystones, and 27 parks should not be shown wearing the estate's clothes.

The evidence base

measurement of record

Every number on this watch is saved with a full receipt, so any figure traces back to where it came from.

  • Each reading is stored as a measurement of record, kept with the exact baseline it was compared against, its anomaly, its observation date, a schema version, and the code commit that produced it.
  • Nothing is recomputed at display time, so any figure on the portal traces back to the reading behind it. This is the depth an evaluation can audit.

How Keystone Watch worksthe two gates, the eleven feeds, and what each one can and cannot raise

Two gates. A flag needs both.

  • Gate 1, the level. The park's latest reading has left its own normal range.
  • Gate 2, the trend. The trend through that reading is still getting worse.
  • Cleared one gate only? It is listed and marked, never counted as an alarm.

Which norm depends on the feed.

  • Vegetation, surface water, night-lights: a multi-year baseline for that time of year, so ordinary seasonal swing is expected.
  • Fire and forest loss: the park's own recent weeks, roughly the last 60 days, which suits fast, event-like signals.
  • Fire also carries its multi-year fire season on the chart as context. The threshold stays the recent-history one.

Eleven feeds, two jobs. Five can raise an alert on their own. Six are context only: they sharpen how an alert is read, but never raise one.

5 alerting feeds
  • Fire: active-fire thermal detections inside the boundary (VIIRS)
  • Forest loss: tree-cover disturbance alerts (Landsat and Sentinel-2)
  • Vegetation: NDVI greenness, where dips flag dieback (MODIS)
  • Surface water: share of the park that is open water, read monthly over the whole park, where declines flag drawdown (Sentinel-2)
  • Night-lights: night-time radiance, a human-pressure proxy (VIIRS Day/Night Band)
6 context feeds
  • Rainfall: week-of-year rainfall normal, a drought cross-check (CHIRPS)
  • Soil moisture: root-zone week-of-year normal, fused with rainfall as one drought cross-check (NASA SMAP L4)
  • Burned area: square kilometres burned inside the boundary each month (MODIS)
  • Land-use conversion: an annual cropland and clearing screen, ground-cross-checked, with a near-real-time check on all 200 parks that qualifies a clear the annual window can no longer vouch for, naming the sensor that spoke (Dynamic World, Sentinel-2 and Hansen; RADD radar and optical DIST-ALERT for currency)
  • Biomass: above-ground biomass trend year on year, a woody-carbon proxy (ESA CCI)
  • Forest structure: whole-mission canopy height, cover and biomass fingerprint (NASA GEDI)

Attribute before you alarm.

  • Low rainfall sharpens a browning, it never quiets one. Here the dry season is when rainfall falls and clearing peaks together, so suppressing on drought would blind the watch exactly when destruction is most likely.
  • Habitat is used to raise a fire where fire does not belong, a rainforest or a wetland. It is never used to mute an expected savanna fire, because abnormal savanna fire is often arson, charcoal or clearing.

The four tabs below.

  • What needs a look: every park that has moved, on any alerting feed, grouped and mapped.
  • Land use: an annual screen of each park's interior for cropland conversion, cross-checked by spectral unmixing so dry savanna grass is not mistaken for tillage.
  • Portfolio trend: how the estate is holding over time.
  • Evidence base: the audit trail behind every number.

Full methodology →