CASE H · FOREST & COMMODITY WATCH

The forest moved last week.
How close did it come to a supply chain?

Radar sees through the monsoon. Every week, Sentinel-1 flags forest that has just been cleared somewhere across the archipelago. Global Forest Watch shows you the pixels. This page asks the question a compliance desk actually has to answer — was that clearing inside an oil-palm estate, on its edge, inside a mill's sourcing radius, on peat, in primary forest? — and then asks how unusual the answer is, against a country where half the alertable forest already sits inside somebody's 50 km radius.

—week —
0hectares alerted, cumulative
—inside a catchment or mapped palm
linked unlinked RADD —
01

Two decades of loss, and two ways to count it

Before the weekly alerts mean anything, the long record has to be honest. Indonesia has two official-looking numbers for the same year, and they differ by a factor of six. Neither is wrong.

—hectares of tree-cover loss, 2024Hansen / UMD, ≥30% canopy
vs
175,400hectares of net deforestation, 2024KLHK, Indonesian government

Gross tree-cover loss against net deforestation

Two definitions, plotted on one axis so the gap is visible rather than argued about.

Where the loss lands, province by province

Annual tree-cover loss 2001–2024 at ≥30 % canopy. Hover a spark to read the year.

So what. A single headline hectare figure is unusable for compliance. The rest of this page therefore never reports a number without the definition, the sensor and the version string attached to it. We have since put a number on the gap: — of Indonesia's tree-cover loss since 2001 falls inside land already mapped as plantation, which is loss but not deforestation. The review article shows the year-by-year split.

02

What the radar saw

RADD reads Sentinel-1 backscatter, so cloud is irrelevant — the reason a radar-first alert layer is the only credible weekly monitor in the wet tropics. We decode the raw date_conf tiles ourselves, group touching alert pixels into disturbance events, and stitch events that straddle a processing seam so nothing is counted twice.

The weekly ledger

Alerted hectares per week, split by whether the event links to the palm-oil system. Switch province to see a different frontier.

Both ends of the series are artefacts and are left in rather than smoothed away. The first two weeks of 2020 carry 10,619 and 9,907 ha against a median week of 1,815 ha, because RADD's Indonesian archive starts on 2020-01-01 and the initial detection batch all carries that date. The final weeks fall away because the tile version was cut on 2026-08-19 and detections near the cut have not yet had a second pass to confirm them. Events are dated by their first alert; the counts are clipped to Indonesian province polygons, because the 10-degree radar tiles also cover Malaysia, Brunei, Papua New Guinea and Timor-Leste.

So what. Alerts arrive as a continuous stream, not an annual report. The unit of action is the event — and an event is only actionable once you know whose land it sits on.

03

Linked, or unlinked

This is the part the public dashboards stop short of. Every event is tested against mapped oil-palm plantation, against the 50 km fresh-fruit-bunch catchment of every mill on the Universal Mill List, and against peat and primary forest. Four mutually exclusive classes, evaluated in order.

Linkage composition by quarter

Share of alerted hectares by class. The frontier and the estate behave differently — and they move differently over time.

So what. The linked share is the number an EUDR desk, a trader and an NGO all need and none of them can get off a public map. But a share is only informative against its base rate, and half of Indonesia's alertable forest already sits inside somebody's 50 km radius — so read the lift, not the share. The review article computes both, province by province, and finds the mill-catchment flag is strongly informative on the Kalimantan and Papua frontier and worth almost nothing in Riau.

04

The mills, as a screening order

A mill cannot be blamed for a clearing inside its catchment — a 50 km radius is a sourcing hypothesis, not a chain of custody, and the catchments overlap so heavily that most alerted hectares sit inside several at once. It is, however, exactly the screen a due-diligence team runs first. Alert pressure = alerted hectares in the catchment over the trailing twelve months.

Alert pressure by mill

Circle area is alerted hectares in the catchment; a ring marks RSPO-certified mills. Hover for the mill card.

So what. Pressure ranking turns 1,000-odd mills into a shortlist of a dozen worth a field visit — which is the entire value of the exercise, and the limit of it. Under the EU Deforestation Regulation an operator needs the plot polygon, not a radius; this is the triage that decides which plots to ask for first.

05

Explore every event

Click any event for its dates, size, flags and nearest mill. Filters and the selected event live in the URL, so a finding can be sent as a link.

— Download shown events (JSON)
06

Methodology, gates and licences

Everything below is reproducible from public sources with one free API key. Where a check did not pass, it says so.

Validation gates

Our hectares against Global Forest Watch's own server-side aggregation of the same tiles over the same geometry — so the comparison isolates our plumbing.

How an event is built

  1. Decode. RADD publishes date_conf as confidence × 10000 + days since 2014-12-31; 2 = low, 3 = high confidence. We read the 10 m tiles in 10,000 px blocks so a 10-degree tile (20 GB uncompressed) never enters memory whole.
  2. Cluster. 8-connected components over alert pixels; events smaller than 0.5 ha are dropped. Events crossing a block seam are re-joined with a union-find over the block edges, so a single clearing is one event, not two.
  3. Measure. Hectares use the true geodetic pixel area at each pixel's latitude, not a nominal constant.
  4. Link. Palm, peat and primary forest are sampled in the same pass — every layer sits on the identical 10-degree grid, so co-location is exact rather than a spatial join. Mill catchments come from the Universal Mill List points via a KD-tree on the unit sphere, which also yields the distance and identity of the nearest mill.
  5. No forest mask is applied to the alerts. RADD is already a forest-disturbance product, and the alert-reconciliation check compares against GFW's own unmasked aggregation — masking here would compare unlike with unlike.

Sources and licences

Every stored layer is CC BY 4.0 with no country carve-out. Licence strings are read live from the API's own dataset metadata, not assumed.

The concession problem

There is no commercially-licensed, redistributable Indonesian concession vector. Global Forest Watch's API marks its oil-palm and wood-fibre concession sets “CC BY 4.0 (excluding Indonesia)”; the RSPO layer carries a map-publication disclaimer rather than an open licence; the ministry's own idn_wood_fiber is published as “View Only, Not Downloadable.” Indonesian law bars non-government publication of estate boundaries. Therefore this page stores and redistributes no concession boundary at all. The stored linkage uses mapped plantation extent, mill catchments, peat and primary forest — all openly licensed. Government concession services remain available as a live, labelled reference overlay at view time.

07

We had this instrument reviewed, adversarially.

An independent read of the same data against the published literature. It computes the base rate this page never had — half of Indonesia's alertable forest is already inside a mill catchment, so the linked share is a lift of 1.76×, not a finding of near-universal palm linkage. It shows the mill-catchment flag is worth 2.4× on the Kalimantan frontier and nothing at all in Riau, measures how much of Indonesia's "tree-cover loss" is plantation rather than deforestation, and corrects three claims this page used to make.

Read the review article →