A clearing alert fired near a palm oil mill. Does that mean anything?

The trap, in one comparison

A mill buys fruit from growers around it, so the standard way to connect satellite-detected clearing to the palm oil trade is to draw a circle of 50 kilometres around every known mill and ask how much of the clearing falls inside one. For Indonesia the answer is 75%, and stated on its own it sounds close to conclusive.

Here is what was missing. The radar that raises these alerts only looks inside one particular kind of forest, and 43% of that forest is already inside a mill's circle — standing, untouched, before anything happens. So the finding is not "75%". The finding is 75 against 43: clearing is about 1.8 times more concentrated near mills than the forest itself is. Real, and much smaller than the bare share suggests.

Compare instead against all Indonesian land, where 52% is inside some mill's circle, and the same clearing is only 1.4 times concentrated. The denominator does more work than the data. Show me the two maps divided.

How close is close enough?

50 kilometres is a convention, not a measurement, and the answer moves with it. Every bar below is the same clearing and the same forest, asked at a different radius.

Two things fall out of that picture. The wider you draw the circle, the more of the clearing you capture and the less it means — which is exactly backwards from how such a number is usually quoted. And at the tightest radius of 5 kilometres, measured against all land rather than against forest, the ratio drops to 0.75: less clearing than an even spread would give you. The ground immediately around a mill is not forest. It was cleared and planted long ago, and this radar cannot see there at all.

What everyone else attributes to palm

There is a published answer to the question this page's headline seems to be asking, and it is much smaller. Working from sample plots inspected by eye, a 2019 study attributes 23% of Indonesian deforestation between 2001 and 2016 to oil palm — about the same as small-scale agriculture at 22%, and not much more than grass and shrubland at 20%. Mapping the plantations directly instead, a 2022 study puts palm at 32% of forest lost. Globally, a 2018 study attributes 27% of all forest loss to permanent commodity-driven clearing of any kind.

So every published estimate of palm's causal share sits between 23% and 32%, and the 75% at the top of this page is two to three times higher. It is not a contradiction, and it is not a better measurement. It is a different measurement: this page counts clearing that happened near a mill, and those studies asked what the land actually became. Proximity is cheap. That gap, between 23%–32% and 75%, is the size of the mistake available to anyone who quotes the bare share.

One thing this record is genuinely quick at, since it is the question field staff ask: the alerts it uses refresh every 6 to 12 days, and miss about 5% of real clearing while raising about 2% false alarms. Fast enough to matter, on the narrow question of whether something was cleared — which is not the same as knowing who cleared it.

Where mill proximity does carry information

Put the map on its third view and the light moves to the other end of the country. Of the 32 provinces with any detected clearing, only 3 are clearly concentrated near mills:

These are frontier provinces with few mills in them, so a mill nearby genuinely narrows things down. Now the 8 where it does not — the palm heartland, and the provinces every account of this subject names:

Riau is the clearest case. 98% of its clearing was inside a mill's circle — and so is 95% of its standing watched forest. The ratio is 1.03. You could have got that number by throwing darts.

There is a third group worth naming. 7 provinces have no mill at all within 50 kilometres of the forest being watched, and 4 of them have enough clearing to be worth putting a name to: Maluku, West Nusa Tenggara, East Java, North Sulawesi. The largest, Maluku, lost 8,397 hectares. Whatever is happening there, the palm-mill frame does not describe it, and a map that showed only mill-linked clearing would leave it blank.

Why nobody can tell you which mill

This is the part that matters most if you are about to act on a finding like this. Circles overlap. The typical alerted hectare inside a catchment sits inside 6 mills' circles at once; the average is 8.6 and the worst ground sits inside 67. 89% of linked hectares are claimed by two or more mills and 61% by five or more.

The arithmetic that follows is blunt. Add up what every listed mill has inside its own circle over the last twelve months and you get 777,894 hectares. Indonesia actually lost 150,971 hectares in that time. The mills' claims exceed the country's losses 5.2 times over, because the same clearing is inside many circles at once. Any league table of mills built from proximity is counting the same hectares repeatedly.

The mill panel beside the map says this per mill: for each one, how many other mills are close enough for their circles to overlap it. The middle mill has 50 such neighbours, the most crowded has 147, and only 6 of the 1,200 listed mills stand alone.

What this cannot do

Five limits, and the first two change the headline number rather than qualifying it.

1. The size filter flatters the result. This case only counts clearings of 0.5 hectares or more, which keeps 33% of the alerted hectares. Counted without that filter, the share near a mill is 59%, not 75%. Small clearings sit further from mills: the smallest band has 66% within reach and a middle distance of 30.6 kilometres, against 82% and 19.7 kilometres for the largest. The filter is stated, not netted out.
2. "All of it was primary forest" is not a finding. Every alerted hectare here is inside primary forest as mapped at the start of the century — 100% of them. That is because this radar is only pointed at that forest. It is a property of the instrument, and it also means clearing inside an estate that was already planted then is invisible to this case entirely. The palm-linked share is a floor.
3. The plantation map is the weak link. The layer used to decide whether clearing was inside an existing estate maps 19,072,149 hectares of oil palm in Indonesia. Two published satellite surveys put it at 11.54 and 16.24 million hectares. Only 14% of the mapped palm lies inside the area this radar watches at all.
4. Small alerts are mostly unconfirmed. A second, optical satellite agrees with only 43% of the smallest clearings, rising to 87% of the largest. Radar sees them first — a middle lead of 20 days, and radar was first for 78% of matched pairs — which is the point of using it, but early is not the same as confirmed.
5. One check was set in advance and failed. The province-by-year clearing totals were compared against the published source in 335 comparisons with a 5% tolerance, and the worst gap was 5.9%. That is a fail, and it is published as one. It happens in Jakarta Special Capital Region on 1.08 hectares — a disagreement about a patch smaller than a football pitch. Restricted to the 324 comparisons above 100 hectares the worst gap is 4%. Both are reported; the failure is not rewritten. The other three checks passed: the alert totals reconcile to within 8.2% against a 10% tolerance, and the optical satellite agrees on 79% of large clearings against a 60% floor.

And "forest loss" is not what most people think it is

One more denominator, because it wrecks more public arguments than any other number in this field. The most-quoted Indonesian deforestation figures measure tree cover removed, from satellite, and they count an oil palm block being replanted exactly the same way they count primary forest being cleared.

Crossed against the map of planted estates, 43% of all Indonesian tree-cover loss since 2001 happened inside land that was already a plantation. It is not a constant: it peaked at 65% in 2010, when the estates planted in the boom came round for harvest, and it is 25% in 2025.

This is also why the government's own deforestation figures and the satellite ones disagree so badly, and why neither is the corrected version of the other. They are answering different questions. A number in this field is only meaningful with its definition attached, which is the whole lesson of this page in one sentence.

What a government or an NGO can actually do with it

1. Publish the base rate next to every share, always. That is the whole method here and it costs nothing. A share of 75% against a base of 43% is a finding; 75% alone is a press release.
2. Spend the attention on the frontier. Mill proximity carries real information in Papua, West Papua, North Kalimantan and almost none in the heartland. The heartland still needs watching — it has the most clearing — but it needs a different instrument, because this one cannot separate a mill's influence from the fact that mills are everywhere there.
3. Ask for the denominator before you act on someone else's number. If a supplier list, a risk score or a league table does not tell you what it is comparing against, it is not something you can regulate on. This case's own pre-registered check is the worked example of getting that wrong.

Could this run in daily operations?

DEPLOYABLE WITH LIMITS — the alert stream and the mill-pressure ranking

Every pipeline here is a batch backfill: it fetches a season or an archive, not a live feed. Anything operational means rebuilding the ingest for near-real-time arrival, whatever the verdict below says.

What stops it. The reconciliation check failed on its single worst comparison, just outside a tight tolerance — while almost every comparison sits inside it, and every area above the materiality floor passes. The failure is on an area too small to matter, and it is published rather than waived.

What it would need. The least of any case here: RADD alerts already arrive weekly, so the ingest is closest to operational. What is missing is alerting and an on-call owner, not science.

If you want the detail

The methods, the four checks set in advance and how each came out, the reconciliation against the published source, and the 8 data sources that were rejected on their licence terms are on the technical article, and the technical dashboard carries the full event table and the statistics behind it. Every figure here is read from the same record as those, so they cannot drift apart.

All questions · openstudy.id

Clearing alerts from wur_radd_alerts v20260823; mills from gfw_universal_mill_list v202508. 404,287 separate clearings totalling 763,643 hectares between 2020-01-01 and 2026-08-19, clipped to Indonesian province polygons — 258,962 hectares fell in neighbouring countries and were dropped rather than counted as Indonesian. The base rate is a systematic sample of the watched forest at 445.3 m spacing. Province outlines simplified to 2,500 m for the web from the same geometry the analysis used: 18,241 points to 8,571. Written to be read without a background in radar or statistics; nothing was rounded to make a point.

Found something wrong on this page? Report a correction — it opens a pre-filled issue — or email taufik.adi@openstudy.id. Corrections are credited by name in the errata, and one that changes a finding says so on the page.