Does a mill catchment tell you anything?

1 The claim under test

Global Forest Watch already publishes the pixels. This case's stated contribution is what comes after them: every alert is clustered into a disturbance event and then tested against mapped oil-palm extent, against the 50 km fresh-fruit-bunch catchment of every mill on the Universal Mill List, and against peat and primary forest. The headline that falls out — 75.0% of alerted hectares linked to palm or a mill — is the number a compliance desk would quote.

Three separate questions hide inside it, and the page answers only the first. Where was forest disturbed? How unusual is it that the disturbance sits inside a catchment? And who, if anyone, caused it? The first is measurement. The second needs a denominator. The third needs a chain of custody that no proximity test can supply. This review supplies the denominator and then asks what is left.

2 What the literature actually attributes to oil palm

Austin et al. (2019), working from visually interpreted sample plots, attribute 23% (90% CI 18%–25%) of Indonesian deforestation between 2001 and 2016 to oil palm, against 20% to grassland and shrubland, 22% to small-scale agriculture and 14% to timber plantations. The palm share is not stable: it peaked near 40% in 2008–09 and fell below 15% by 2014–16.

Gaveau et al. (2022), mapping plantations directly, find 3.09 Mha of old-growth forest converted to oil palm in 2001–2019 — 32% of the 9.79 Mha lost. Globally, Curtis et al. (2018) attribute only 27% of forest loss to permanent commodity-driven deforestation, with 26% to forestry rotations that are not deforestation at all.

Every published estimate of palm's causal share of Indonesian forest loss sits between 23% and 32%. This page's 75% is two to three times higher — because it is not measuring the same thing. It measures proximity, and proximity is cheap.

3 Data, and what the instrument actually is

Alerts. wur_radd_alerts v20260823 — Sentinel-1 C-band radar, 6–12 day revisit, 0.1 ha minimum mapping unit, unaffected by cloud (Reiche et al. 2021, who validate confirmed alerts at 2% false positives and 5% false negatives). The record runs 2020-01-01 → 2026-08-19; 98.9% of the hectares in the event table are high-confidence, so the confidence flag is not where the risk lies.

Events. Alert pixels are clustered 8-connected within a 28-day window and events below 0.5 ha are dropped, leaving 404,287 events and 763,643 ha. A further 166,789 events / 258,962 ha are clipped away as Malaysian, Bruneian, Papua New Guinean or Timorese — the 10-degree radar tiles are not country-clipped, and without that clip a quarter of the "national" total belongs to somebody else.

Linkage. Mapped plantation is gfw_planted_forests v20231128 (SDPT v2), peat is gfw_peatlands v20230315, primary forest is umd_regional_primary_forest_2001 v201901. Mill catchments are computed from 1,396 Universal Mill List points with a KD-tree on the unit sphere. The 50 km radius is the industry convention: fresh fruit bunches must reach a mill within roughly a day of cutting, which historically bounded sourcing at about that distance.

What this review adds. Three tests: (a) a 445.3 m lattice over the instrument's entire detection domain with the mill distance at every sample; (b) 2001–2025 Hansen loss crossed with plantation extent and the primary-forest mask at native 30 m; (c) the catchment share recomputed on unfiltered alert hectares. Nothing here uses data the case did not already have on disk.

4 Finding one — loss is not deforestation, and the gap is measurable

The single most common misreading in this literature is that Hansen tree-cover loss is deforestation. It is not: Hansen et al. (2013) define it as stand-replacement disturbance, and it counts an oil-palm block being replanted or an acacia stand being harvested exactly as it counts primary forest being cleared. The case states that caveat in prose and publishes the KLHK divergence panel. It never measured it. We did.

43.3% of all Indonesian tree-cover loss since 2001 — 14.44 Mha of 33.36 Mha — falls inside land already mapped as plantation: 9.513 Mha oil palm, 2.805 Mha rubber, 2.122 Mha wood fibre. The share is not constant. It rose to 65% in 2010, during the palm expansion boom, and has fallen to 25% by 2025.

That fall is itself the most interesting number in the figure, and it agrees with Austin et al.'s finding that palm's share of deforestation collapsed after 2012 and with Gaveau et al.'s finding that expansion had returned to pre-2004 levels by 2019. Three independent methods — theirs, theirs, and a raster crossing on this box — see the same turn.

What this measurement cannot do. SDPT is a single 2023-vintage extent, so loss inside today's plantation in 2004 is the conversion that created the estate, while the same pixel class in 2024 is far more likely to be a rotation. The number bounds the definitional gap; it does not partition it into conversion and harvest. Doing that properly needs a plantation-establishment year, which no openly licensed Indonesian layer supplies.

5 Finding two — half the country is already inside a catchment

A 50 km circle drawn around each of 1,396 mills is a large object. The question the page never asks is how much of Indonesia those circles already cover. We answered it by sampling the instrument's own detection domain — the 2001 primary humid-tropical forest mask, which is exactly where RADD is permitted to raise an alert — on a 445.3 m lattice, and measuring the great-circle distance to the nearest mill at every one of those samples.

So the headline is not empty — alerted hectares really are 1.76 times more likely to sit inside a catchment than a random hectare of alertable forest. But the informative statement is the lift, not the share. Quoting 74.8% without 42.5% beside it invites a reader to conclude that three-quarters of Indonesian forest disturbance is palm-related, which is roughly three times what the literature attributes to palm (§2) and is not what the number says.

The signal is strongest exactly where mills are scarce, and vanishes where they are dense. In North Kalimantan the lift is 2.4317×; in West Papua, 2.1764×. In Jambi it is 0.9899× and in Riau 1.0335× — which is to say, none.

This inverts the product. A mill-catchment flag in Riau distinguishes nothing, because 94.8% of Riau's alertable forest is inside a catchment already. The same flag on the Kalimantan and Papua frontier is genuinely discriminating. The case's own linkage check — Riau's linked share must exceed 25%, and it reaches 97.9% — is therefore a check that could not have failed. It also compares a proximity share against a floor drawn from Gaveau et al.'s direct conversion figure, which is a different quantity. Both problems are ours to fix, and §13 records what we changed.

6 Finding three — the published radius is near the bottom of the range

If proximity to a mill carries information, it should carry more of it close in. The base rate makes that testable at any radius, because both numerator and denominator are recomputed together.

The lift peaks at 3.01× at 10 km — where 5.3% of the domain sits but 16.0% of alerted hectares do — and decays monotonically to 1.76× at 50 km and 1.3795× at 100 km. Of the 8 radii tested, only 75 km and 100 km discriminate less than the one the case publishes.

That is not an argument for abandoning 50 km, which is the industry's own sourcing convention and the number a trader recognises. It is an argument for publishing the 10 km class alongside it. A hectare within 10 km of a mill is three times more surprising than chance; a hectare within 50 km is 1.76 times. Those are different signals and they should not share a colour.

7 Finding four — the headline is partly a product of the size floor

The event table keeps only clusters of at least 0.5 ha. The page states the consequence for the reconciliation provinces — the floor keeps 22% of alerted hectares there. Nationally it keeps 32.8%: 763,643 ha survive out of 2,326,554 ha of alert. That is a large filter, and filters are only harmless if they are random.

It is not random. Mill distance falls steadily with event size: the median for 0.5-1 ha events is 30.6 km, for 20–100 ha events 18.2 km. Small clearings — smallholder plots, roadside encroachment, staged clearing — sit further from mills than large ones, and the floor removes them preferentially.

Recomputing the catchment share on the unfiltered alert grid, with no size floor at all, gives 59.2% against the published 74.8% — a gap of 15.6 percentage points, and a lift over the base rate of 1.39× rather than 1.76×. Roughly a third of the headline's distance above chance is the size floor, not the landscape.

The unfiltered comparator is coarse: the raw grid the pipeline writes for its own alert-reconciliation check is 0.05° (about 5.5 km), so each cell's mill distance is its centre's. Against a 50 km radius that blurs the boundary symmetrically rather than shifting it, so the direction of the finding is safe even though the second decimal is not.

8 Finding five — eight mills claim every hectare

The case is careful in prose: "a mill cannot be blamed for a clearing inside its catchment — a 50 km radius is a sourcing hypothesis, not a chain of custody." The arithmetic underneath that sentence deserves the same care.

An alerted hectare inside a catchment is inside a mean of 8.58 catchments (median 6, 90th percentile 19, maximum 67). 88.6% of in-catchment hectares are claimed by two or more mills and 61.3% by five or more. "Linked to a mill" therefore names, on average, eight or nine suspects.

This propagates into the ranking. Summed over all 1,396 mills, the alert pressure the page publishes comes to 780,633 ha for the trailing twelve months, against 109,384 ha of alert actually inside any catchment in that window — an inflation of 7.14×. Each mill's number is a defensible statement about that mill's neighbourhood; the column is not additive, and a reader who sums it, or who compares two overlapping mills as though the hectares were disjoint, will be wrong.

9 Finding six — the sensor's domain decides most of the answer

The case already publishes the most important structural fact about RADD in Indonesia: 100% of alerted hectares are inside 2001 primary forest, because that mask is the detection domain. What it does not quantify is how much of the palm system that excludes.

Only 13.7% of the oil palm this case maps — 2.61 Mha of 19.07 Mha — sits inside the domain. An estate that was already plantation in 2001 cannot raise a RADD alert at all, so the PALM-INTERNAL class (1.8% of hectares) is not a measure of replanting; it is a measure of the small part of the estate that used to be primary forest and still carries the mask.

A claim in the case's own notes that its data contradicts. The repository records a decision that SDPT v2 "under-maps smallholder palm relative to Descals' 10 m remote-sensed extent", and concludes the published palm-linked share is therefore conservative. Measured on the same lattice, SDPT maps 19.07 Mha of Indonesian oil palm against Descals et al. (2021)'s 11.54 Mha mapped / 12.05 Mha estimated and Gaveau et al. (2022)'s 16.24 Mha mapped / 18.83 Mha adjusted. It is 1.653× Descals, not a fraction of it. Whatever else is true, the conservatism argument does not hold, and the note has been corrected.

A claim on the page itself that its data contradicts. The page states that Java and Nusa Tenggara "return no alerts at all". They return 2,590.2 ha and 6,977.5 ha respectively — 0.34% and 0.91% of the national total, from 32 of 34 provinces that produce any. Vanishingly small, which is the point being made, but not zero. Corrected in §13.

10 Finding seven — what "two-sensor agreement" is measuring

The case's two-sensor check reports that 78.6% of high-confidence events of at least 5 ha carry a GLAD-L optical alert within ±60 days, and passes against a 60% threshold. That subset is 17,062 events — 4.2% of the record and 34.6% of its hectares.

Confirmation rises monotonically with size, from 42.7% for 0.5-1 ha events to 86.9% for the largest. Across all 404,287 events it is 51.4% by count and 65.0% by hectare. Two things are tangled here and the check does not separate them: a bigger clearing genuinely is easier for a 30 m optical sensor to see, and a bigger footprint contains more GLAD pixels spanning more dates, so the interval-overlap rule the check uses becomes easier to satisfy. The published 78.6% is therefore partly a property of the ≥5 ha filter.

The check is also one-directional. It asks how many radar alerts optical confirms; it cannot ask how many optical alerts radar missed, because GLAD is only sampled inside RADD footprints. A shortfall here is evidence about optical, never about radar.

Where both sensors do fire, the radar-first case is what the case is built on and it holds: radar is first for 78% of matched events, with a median lead of 20 days (interquartile range 3 to 38). In the wet tropics that is the difference between acting this month and acting next quarter, and it is the strongest single justification for a radar-first monitor. The remaining 41% of matched pairs sit outside the ±60 day window entirely — usually because the same ground was disturbed in an earlier episode that GLAD recorded — so they are co-location, not concurrence, and are excluded from the lead statistic.

11 Finding eight — what does reconcile, and what the checks are worth

Adversarial review is only credible if it also reports what survives. The plumbing here does.

The case's tree-cover-loss reconciliation lands within 0.71% of GFW's published 2024 figure, and this review's independent crossing reproduces the case's own published national loss series to within rounding. More usefully, the primary-forest loss computed here for the first time — Hansen loss year crossed with the 2001 primary mask at 30 m — lands within 0.71% and 0.75% of GFW's published Indonesian primary-forest loss for 2023 and 2024. That is a genuinely independent check on the geometry, the pixel-area weighting and the masking, and it passes.

The four published checks are a mixed set once the base rate is in hand. The alert reconciliation (worst deviation 8.17% against a ±10% tolerance) is a real test of the raster plumbing and it earned its keep — the repository records that it is what caught a clustering bug that had back-dated most of the archive into 2020. The tree-cover-loss reconciliation is reported failed, on a single province-year of 1.08 ha against 1.15 ha in Jakarta; 99.7% of 335 comparisons pass and every one of the 324 above 100 ha does. Publishing that as a failure rather than re-cutting the tolerance is the right call and we would not change it.

The two remaining checks are weaker than they look. The two-sensor check measures a filtered subset whose pass rate is partly the filter (§10). The linkage check cannot fail, for the reason in §5. Neither is wrong; both need their denominator printed beside them, which is what §13 does.

12 What follows for decisions

Evidence is only worth gathering if it changes an action. Read strictly, this instrument supports four uses and forbids a fifth.

1. Weekly triage on the frontier. Where the lift is 2.4317× or 2.1764× — North Kalimantan, West Papua, Papua — a catchment flag genuinely narrows the search. For a provincial forestry office in Kalimantan or an NGO monitoring team, the ranked weekly list is the product, and the 20-day radar lead over optical is what makes it a weekly list rather than a quarterly one.
2. Peat exposure, which is the highest-consequence flag on the page. 13.6% of alerted hectares are on peat and 85.5% of those are inside a catchment — a higher rate than alerts as a whole. Indonesia's peat moratorium and the EUDR both make peat clearing categorically different from other clearing, and this is the one flag where a spatial coincidence maps onto a legal boundary rather than a commercial hypothesis.
3. Due-diligence triage, explicitly labelled as triage. Regulation (EU) 2023/1115 requires plot-level geolocation — a polygon for any plot over 4 ha — and a cut-off of 31 December 2020, which is almost exactly where this alert record begins. A 50 km circle is the pre-EUDR proxy the regulation was written to replace. It cannot discharge due diligence; it can tell an operator which of its declared plots to check first, and which weeks to check them in.
4. The definitional argument, which this case wins. Chapter 01 already refuses to publish one headline hectare figure. §4 now puts a number on why: 43.3% of the loss is inside land already mapped as plantation. For KLHK, for a ministry press office, and for anyone arguing about whether Indonesian deforestation is 175,400 ha or 1.1 Mha, that decomposition is more useful than either number alone.
5. Not: attributing clearing to a named mill or group. With 8.58 mills per hectare, a 1.76× lift and a pressure column that sums to 7.14× the hectares that exist, the ranking cannot support a statement about any individual mill's conduct. It is a screening order. Naming groups next to it invites exactly the reading the page's own caveat disclaims.

The honest reframing is small and it makes the case stronger: this is not "deforestation as a supply-chain question", it is a weekly triage instrument whose value is highest exactly where the supply chain is least mapped. That is the gap between having data and acting on it — which is the axis the work is meant to sit on.

13 Corrections applied to the case page

Everything in this section has been changed on the instrument itself, not merely recommended.

1. The headline now carries its denominator. The linkage tile reads 74.8% against a base rate of 42.5% and a lift of 1.76×, and the 10 km class is published beside the 50 km one.
2. The linkage check is relabelled. It was presented as a literature-floor test against Gaveau et al.'s 32% direct-conversion figure. A proximity share cannot be tested against a conversion share, and in Riau the check could not fail at any threshold below 95%. It is now stated as a lift test with Riau's base rate printed beside it.
3. "Java and Nusa Tenggara return no alerts at all" is corrected to the true figures — 2,590.2 ha and 6,977.5 ha, together 1.25% of the national total.
4. The national event-floor share is published. The page previously gave only the three-province figure (22%); the national figure is 32.8%, and the unfiltered catchment share 59.2% is now stated next to the filtered one.
5. The mill ranking is relabelled a screening order and carries the number of mills sharing each hectare, so the column is not read as additive hectares.
6. The SDPT extent note is corrected in the repository: SDPT maps 1.653× Descals' Indonesian oil-palm extent, so the palm-linked share is not conservative for the reason previously given.

Two things we deliberately did not change. The failed tree-cover-loss reconciliation stays failed and stays published. And the 50 km radius stays, because it is the convention the industry and Trase use; it now simply travels with its base rate.

14 References and reproducibility

1. Hansen, M.C. et al. (2013). High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 342(6160), 850–853. doi:10.1126/science.1244693
2. Austin, K.G., Schwantes, A., Gu, Y. & Kasibhatla, P.S. (2019). What causes deforestation in Indonesia? Environmental Research Letters 14(2), 024007. doi:10.1088/1748-9326/aaf6db
3. Gaveau, D.L.A. et al. (2022). Slowing deforestation in Indonesia follows declining oil palm expansion and lower oil prices. PLOS ONE 17(3), e0266178. doi:10.1371/journal.pone.0266178
4. Curtis, P.G., Slay, C.M., Harris, N.L., Tyukavina, A. & Hansen, M.C. (2018). Classifying drivers of global forest loss. Science 361(6407), 1108–1111. doi:10.1126/science.aau3445
5. Reiche, J. et al. (2021). Forest disturbance alerts for the Congo Basin using Sentinel-1. Environmental Research Letters 16, 024005. doi:10.1088/1748-9326/abd0a8
6. Descals, A. et al. (2021). High-resolution global map of smallholder and industrial closed-canopy oil palm plantations. Earth System Science Data 13, 1211–1231. doi:10.5194/essd-13-1211-2021
7. Turubanova, S., Potapov, P.V., Tyukavina, A. & Hansen, M.C. (2018). Ongoing primary forest loss in Brazil, Democratic Republic of the Congo, and Indonesia. Environmental Research Letters 13, 074028. doi:10.1088/1748-9326/aacd1c
8. Richter, J. et al. (2024). Spatial Database of Planted Trees (SDPT) version 2. World Resources Institute. doi:10.46830/writn.23.00073
9. Regulation (EU) 2023/1115 of the European Parliament and of the Council of 31 May 2023 on the making available on the Union market and the export from the Union of certain commodities and products associated with deforestation and forest degradation. OJ L 150, 9.6.2023, 206–247.
10. World Resources Institute et al. Universal Mill List, gfw_universal_mill_list v202508. Accessed through Global Nature Watch.

Data vintages: alerts wur_radd_alerts v20260823, optical umd_glad_landsat_alerts v20260829, loss umd_tree_cover_loss v1.13, plantation gfw_planted_forests v20231128, peat gfw_peatlands v20230315, primary forest umd_regional_primary_forest_2001 v201901, mills gfw_universal_mill_list v202508. Every stored layer is CC BY 4.0; no concession boundary is stored or redistributed anywhere in this case.