Global Forest Watch: Intelligence Source Guide

Global Forest Watch publishes satellite-derived tree cover loss and near-real-time deforestation alerts alongside concession, protected area and tenure layers. It is the standard open source for detecting forest change fast, and for mistaking tree cover loss for deforestation.

At a glance

Source	Global Forest Watch
Category	Conflict, Crime & Human Security › Environmental & Wildlife Crime
Homepage	https://www.globalforestwatch.org/
Machine interface	`https://data-api.globalforestwatch.org/`
Format	`JSON`
Access	Open — no account required
Disciplines	Environmental Intelligence, Geospatial Intelligence
Mission domains	Environmental Crime

Near-real-time deforestation alerts. — as catalogued in the platform’s own source registry.

Global Forest Watch is a platform run by the World Resources Institute that assembles remote sensing products about forests into a map, a data portal and an API. The two products that matter most are structurally different. The first is annual tree cover loss, derived from Landsat imagery by a university remote sensing laboratory, running from the early 2000s to the most recent complete year at roughly thirty-metre resolution: for each pixel it records the year in which tree cover was lost, against a baseline of percentage canopy cover in a reference year. The second is the near-real-time alert system, which is not one system but several combined – an optical alert product using Landsat, a second optical product using Sentinel-2 at ten-metre resolution, and a radar-based product using Sentinel-1 that sees through cloud. These are published individually and as an integrated alert layer that reconciles them and assigns a confidence level based on how many independent systems detected the same change. Around these sit fire alerts from thermal sensors, tree cover gain, primary forest extent, land cover, and a substantial contextual library: logging, mining, palm oil and other concession boundaries where governments or companies have published them, protected areas, Indigenous and community lands, and administrative units. A separate product line serves company supply chain analysis. Access is through the map interface, an open data portal, and a data API.

The job this source does that nothing else does is to make forest change observable independently of anyone's report of it. Every other route to knowing whether a forest was cleared depends on someone whose interest is engaged – a concession holder, a ministry, a certification body, a community with a grievance. This is a physical measurement made by a satellite that has no view on the outcome, published on a schedule nobody controls. For ENVINT and GEOINT work that is decisive, because it converts a contested claim into a checkable one. The near-real-time alerts do something further: they compress the detection window from years to days, which changes what enforcement is possible. An annual loss statistic tells you a forest was cleared eighteen months ago; a radar alert tells you clearing is happening in a concession this week, in time for a ranger patrol, a supplier suspension or a court application. The contextual layers are what convert detection into attribution: a cluster of alerts is a fact about pixels, but a cluster of alerts inside a named concession, inside a protected area, or inside a titled Indigenous territory is a fact about a party with obligations. That combination – independent detection plus jurisdictional and tenure context – is why this source underpins most modern deforestation litigation, supply chain due diligence and land defender documentation. What it does not do, and cannot, is tell you who cleared the trees or whether they were entitled to.

Who publishes it, and why that matters

The platform is operated by the World Resources Institute, a research organisation funded by governments, foundations, multilateral institutions and corporate partners, working through a partnership that includes the universities and research groups that produce the underlying detection algorithms. Two features of that arrangement matter to an analyst. First, the algorithms are not WRI's; they come from academic laboratories that publish their methods and validation studies in peer-reviewed journals, which means you can read what the omission and commission error rates are for a given product in a given biome instead of taking a vendor's word for it. That is a substantial advantage over commercial monitoring services. Second, the platform is a mission-driven convener, and its relationships include the companies and governments whose behaviour the data describes. It has managed that tension mostly by keeping the data open and letting others do the accusing, which is the right structure but which also means the platform itself will rarely tell you that a specific company deforested a specific area – it gives you the layers and leaves the inference to you. Funding is grant-based, which introduces the ordinary risk that a specific product line depends on a specific grant cycle; the core loss and alert products have proved durable over more than a decade, but individual contextual layers have appeared, aged and stopped updating. Always check when a concession layer was last refreshed before building a finding on it.

Provenance is the first question to ask of any dataset and the one most often skipped. Who collects it, what their incentive is, whether they publish a methodology, and whether they correct the record when they get something wrong all bear directly on how much weight a finding drawn from it can carry.

What a record actually contains

The fields you will be working with, what each one means, and whether it is something you can pivot on. Read the meanings carefully — more analysis is wrecked by misreading a field than by failing to find one, and a field that looks like an observation is often an inference.

Field	Type	What it means	Pivot value
`alert_date`	timestamp	The date of the satellite acquisition in which the change was first detected, not the date the change occurred and not the date it was published. Clearing can precede detection by days to weeks under persistent cloud.	Timeline construction; correlation with concession activity, permits and shipment records.
`confidence`	enum	For integrated alerts, whether the change was flagged by one detection system or corroborated by more than one. Low confidence does not mean false; it means unconfirmed by a second sensor at the time of publication.	Triage priority; decision on whether to task higher-resolution imagery.
`alert_system`	enum	Which detector produced the alert: Landsat-based optical, Sentinel-2 optical, or Sentinel-1 radar. Determines resolution, revisit, cloud behaviour and the characteristic false positive modes.	Expected latency; interpretation of absence; choice of corroborating imagery.
`lat / lon`	array	Alert location, reported at the resolution of the source product – ten metres for Sentinel-2 based alerts, coarser for Landsat-based. A single alert pixel is a small area, and clusters rather than pixels are the analytical unit.	Concession, protected area, administrative unit and tenure overlay; distance to road, river and mill.
`loss_year`	int	For the annual tree cover loss product, the year in which loss was detected for a pixel. Comparability across product versions is not guaranteed because methodology has been revised between releases.	Multi-year trend within a single version; comparison against concession grant dates.
`tree_cover_density`	int	Percentage canopy cover in the baseline year, used as the threshold that defines what counted as tree cover in the first place. Choosing thirty percent versus seventy-five percent changes every downstream number.	Sensitivity analysis; distinguishing dense forest from sparse woodland and plantation.
`primary_forest`	enum	Whether the pixel falls within mapped primary humid tropical forest. Loss inside primary forest is a materially different fact from loss inside a plantation or secondary regrowth, and conflating them is the most common analytical error.	Severity assessment; biodiversity and carbon significance; regulatory relevance under deforestation-free rules.
`area_ha`	int	Area of loss or alerts aggregated over a geography and period. Derived by counting pixels, so it inherits the resolution limits and is not a survey measurement.	Comparison across concessions and periods; scale assessment for reporting.
`concession`	string	The mapped commercial concession the location falls in – logging, oil palm, pulpwood, mining – where a government or company has published boundaries. Coverage is partial and boundary vintage varies enormously.	Company ownership resolution, licence records, EITI disclosure, supply chain mapping.
`protected_area`	string	Protected area designation from the world database, with its IUCN management category where assigned. Category determines what activity is permitted, so the designation alone does not establish illegality.	Management authority, legal instrument establishing the area, permitted use, enforcement responsibility.
`indigenous_or_community_land`	string	Mapped Indigenous or community land where documented. Coverage is incomplete by nature, because much customary tenure is unmapped, and absence of a polygon is not absence of a claim.	Community organisations, tenure litigation, free prior and informed consent records.
`admin_unit`	string	Country and subnational administrative unit containing the location. The basis for aggregating alerts to a jurisdiction with an identifiable enforcement authority.	National forest authority, subnational government, country risk assessment.
`fire_alert`	timestamp	Thermal anomaly detection from a satellite fire sensor. It detects heat, not fire type, and picks up agricultural burning, industrial flares and other heat sources alongside forest fire.	Correlation with clearance alerts; seasonal burning patterns; haze and air quality data.
`driver / land_cover`	enum	Modelled classification of the dominant driver or land cover class. This is a coarse model output, typically produced at low spatial resolution, and it should never be used to attribute a specific clearing event.	Regional pattern analysis only; not usable for site-level attribution.

Coverage — and what is not in it

Coverage is global for the annual tree cover loss product, which is the most spatially complete dataset of its kind and covers every land area imaged by Landsat from the early 2000s onward. The near-real-time alert products are narrower. The optical alert systems were built for the humid tropics and expanded outward from an initial set of tropical countries, so coverage outside the tropics is more recent and in places absent; the radar alert product covers the pantropical belt and is the only one that works reliably through persistent cloud. Practically, this means that the deforestation frontiers of the Amazon, the Congo Basin and Southeast Asia are covered well and quickly, that boreal and temperate forest change is captured annually but with far less near-real-time capability, and that dryland woodland and savanna change is poorly represented by a product designed around tree canopy cover. Update rhythm differs by product: alerts update within days of acquisition, subject to cloud and processing latency; annual loss is released once a year, several months after the year it describes, and each release is a reprocessing of the entire series rather than an increment. Contextual layers have no rhythm at all – concession boundaries, protected areas and tenure layers update when a government or organisation publishes new data, which for some countries means the layer you are using is several years old. Entity coverage is pixels and polygons. There are no companies, no people and no events in the base data; those come from the layers you join to it.

Known blind spots

Absence of evidence here is not evidence of absence. These are the conditions under which Global Forest Watch will not show you something that is nevertheless real:

Tree cover loss is not deforestation. The product records loss of canopy from any cause, including plantation harvest, storm damage, fire, disease and legal selective logging, and a plantation rotation looks identical to clearance of natural forest at the pixel level.
Degradation without canopy removal is largely invisible. Selective logging of high-value timber under an intact canopy, the activity that precedes and finances most frontier clearance, does not reliably trigger these detectors.
The canopy cover threshold defines what counts as forest, so open woodland, savanna, mangrove and dryland ecosystems are systematically under-represented, and change in them is under-reported by construction.
Optical alerts fail under persistent cloud, which in equatorial regions can suppress detection for weeks; the radar product mitigates this but has its own error profile and does not cover everywhere.
Alerts detect change, never cause and never actor. Nothing in the data distinguishes a licensed operation from an illegal one, or a company from a smallholder, and treating a cluster of alerts as evidence of a crime skips the entire legal question.
Concession and tenure layers are incomplete and unevenly dated. Many countries publish nothing, some published once and never updated, and much customary and Indigenous tenure has never been mapped at all – so a clearing outside every polygon may still be on someone's land.
Annual loss releases involve methodological revisions, so numbers from different versions are not comparable, and a change in national loss between two published figures may be an artefact of reprocessing rather than a change on the ground.
Small clearings below the effective detection size, and narrow linear features such as extraction tracks, fall below or between pixels, which biases the data against exactly the incremental smallholder and artisanal patterns that dominate some frontiers.
Regrowth, replanting and recovery are handled by separate and coarser products, so a location can show loss without any indication that the land was subsequently restored, which distorts long-run assessments.

Write the blind spot into the product. A statement that something “was not observed in Global Forest Watch” is defensible; a statement that it “did not happen” is not, and the difference is what survives cross-examination.

Access, licensing and what you may do with it

Access model: Open — no account required

Everything is open. The map requires no account and is genuinely useful rather than decorative – it is the fastest way to see an area's alert history, the concession and protected area context, and the underlying imagery, and it should be the first stop before any programmatic work. Bulk downloads are published through an open data portal in raster and vector form, and these are the correct route for anything covering a large area or a long period. A data API supports programmatic queries over the hosted datasets, including aggregation by geometry, which is how most integrations retrieve alerts for a defined area of interest. There is also a separate product aimed at company supply chain monitoring which operates under different access arrangements. Three practical notes. Datasets are versioned and you should record the version you used, because reprocessing changes historical values. Aggregating alerts by a custom polygon is the workhorse operation and should be done server-side rather than by pulling raw pixels. And the contextual layers each have their own provenance and vintage; before you cite a concession boundary, find out who published it and when, because that metadata is what a respondent will attack first.

Licence

The core forest change products are published openly for public use with attribution, and the platform's stated intent is that the data be freely used including by companies and governments. The important nuance is that the platform aggregates datasets from many producers, and the licence attaches to each dataset rather than to the platform as a whole. The annual loss product comes from a university laboratory with its own citation requirement; protected area boundaries come from an international database with its own terms that restrict some commercial uses; concession layers may come from a government, a company or an NGO, each with different conditions. Before redistributing or building a product on any specific layer, check that layer's own terms and citation requirement rather than assuming a single site-wide licence. Attribution is expected in all cases and, for the academic products, correct citation of the underlying publication is the norm rather than a courtesy.

Rate limits and fair use

Design around bulk retrieval rather than repeated querying. For a defined area of interest, the efficient pattern is a scheduled daily or weekly aggregation query against your polygons, which is cheap; the inefficient pattern is pulling raw alert pixels for a large region and filtering locally, which is expensive for both sides and unnecessary given server-side aggregation. Annual products should be downloaded once per release and held locally. Assume unpublished quotas exist, set a descriptive user agent with a contact address, cache results, and back off rather than retrying tight loops. If your requirement is continuous monitoring of many polygons, batch them into a single scheduled job rather than issuing one request per polygon per day.

Licensing changes, and it changes without warning. A dataset that was free for research this year may not be free for commercial or evidential use next year. Confirm the current terms before you build a dependency on it, and record the terms you relied on alongside the data — the licence in force at the time of collection is part of the provenance.

Collecting it

How Global Forest Watch is actually pulled, in the order you would set it up. Prefer the bulk or export interface over per-item lookups wherever one exists: it is kinder to the publisher, faster for you, and gives a reproducible snapshot rather than a series of point-in-time answers you cannot reconstruct later.

Method	Format	Cadence	Notes
Alert aggregation by geometry	JSON	daily or weekly	Submit your areas of interest as polygons and retrieve alert counts and locations. The core operational pull and the one that scales.
Annual tree cover loss raster	bulk	once per annual release	Download and hold locally with the version recorded. Reprocessing means each release supersedes rather than extends the last.
Contextual vector layers	CSV	per publication; re-check quarterly	Concessions, protected areas, tenure and administrative boundaries. Capture the publisher and vintage of each layer as metadata, because that is what determines its evidential weight.
Fire alerts	JSON	daily during burning seasons	Thermal anomaly detections, useful as a correlate of clearance and as an independent signal in fire-driven conversion. Not a forest-specific product.
Map interface	HTML	ad hoc	Visual verification of any programmatic result against the underlying imagery. Alert clusters that look like clearing usually are; ones that follow a river bend or a cloud edge usually are not.

Ingesting it into the platform

Every step below is idempotent and cursor-based: interrupt one and it resumes from where it stopped rather than duplicating rows or losing progress. Collection is recorded per source, so a feed that quietly stops publishing shows up as a stale timestamp instead of silently thinning your coverage.

Register the products separately — Add Global Forest Watch in sources.php as distinct feeds for annual loss, integrated alerts, fire alerts and contextual layers. They differ in cadence, latency, resolution and error profile, and a single registration hides all four differences from every analyst downstream.
Define and version the areas of interest — Store your monitoring polygons – concessions, protected areas, supply sheds, community lands – as versioned geometries with their own provenance. The polygon is half the finding, and a boundary that silently changed will invalidate a whole time series.
Schedule the alert pull — Run the aggregation query through collect.php on a daily schedule under cron.php. Latency here directly determines whether an alert can support an intervention, and a failed job must raise an alert rather than presenting as a quiet period in the forest.
Cluster before you store — In ingest.php, group contiguous and near-contiguous alerts into events with a first-detection date, an extent and a shape. Individual pixels are noise; the cluster is the unit that corresponds to a real-world clearing and the only unit worth alerting on.
Attach context at ingest — Use enrich.php to stamp each cluster with its concession, protected area, tenure and administrative unit, recording the layer version used. Retrospective overlay against a later boundary file produces findings that cannot be reproduced.
Resolve concessions to companies — Run resolve-everything.php to take concession identifiers out to corporate registries, licence records and, where the country participates, extractive transparency disclosures, producing an organisation entity rather than a polygon label.
Correlate with the wider case picture — Use correlate.php to test clusters against conflict events, mining site data, road development, port and shipment records and adverse media. Deforestation is rarely the whole case; it is the observable part of a supply chain or a land conflict.
Publish to mission views and set thresholds — Expose results in timeline.php and the environmental and human rights mission areas, and configure alert rules on cluster size, protected area intrusion and rate of change rather than on raw alert counts, which will otherwise bury every analyst in seasonal noise.

Registered sources and their last-collected state are listed in sources.php, and the scheduled chain that keeps them current is in automation.php.

How it is wrong, and how to tell

Every dataset is wrong in characteristic ways. Knowing which ways is the difference between using a source and being used by one, and it is the part of source evaluation most often skipped because it is the part that takes work.

The underlying science is strong and, unusually, it is documented well enough that you can quantify how much to trust it. The annual loss product and each alert system have published validation studies reporting omission and commission errors by biome, and those studies are the right basis for a confidence statement rather than a general impression. What the numbers show consistently is that these products are good at detecting large, abrupt, complete canopy removal in humid tropical forest and progressively less good as the change becomes smaller, more gradual, more partial, or occurs in drier or more open vegetation. Radar-based alerts trade cloud immunity for a different error profile, with terrain and inundation producing characteristic artefacts. The integrated alert layer's confidence field is a real quality signal and should be used: multi-system corroboration substantially reduces commission error, at the cost of latency. The two systematic risks are not sensor error at all. The first is definitional: the product measures canopy loss and users read deforestation, and no amount of sensor accuracy fixes an inference the data does not support. The second is contextual: an alert is only as good as the concession or protected area boundary you overlay it against, and those boundaries are frequently the weakest link in a finding. Judge quality by corroboration rate – what proportion of your clusters, when checked against high-resolution imagery or ground reporting, turn out to be what you assumed – and measure that rather than assuming it.

Characteristic false positives

Plantation harvest registers as loss identically to natural forest clearance, so a pulp or oil palm estate on its normal rotation generates a large annual loss figure that says nothing about deforestation.
Cloud shadow, haze and smoke produce optical artefacts that mimic canopy change, and they cluster in exactly the regions and seasons where burning and clearing are most active, so the noise correlates with the signal.
Radar alerts fire on flooding, seasonal inundation and steep terrain effects, producing clusters along rivers and slopes that are hydrology rather than logging.
Natural disturbance – windthrow, landslide, drought dieback, disease and lightning fire – is indistinguishable from anthropogenic clearance in the data and requires context or imagery to separate.
Concession and protected area boundaries are frequently outdated or approximate, so an alert attributed to a company or a park may fall outside the boundary that was legally in force at the time.
Fire alerts detect heat, not deforestation; agricultural stubble burning, gas flares, industrial heat sources and even large fixed installations generate detections that get read as forest fire in aggregate statistics.
Version changes in the annual product alter historical values, so a year-on-year comparison assembled from two releases can manufacture a trend that exists only in the processing.
Small-scale and shifting cultivation produces alert patterns that look like organised clearing when aggregated, which can turn a subsistence practice into a headline about industrial deforestation if nobody looks at the pattern shape.

None of these make the source unusable. They make it a source that requires corroboration before an assertion built on it goes into a product, which is true of every source and admitted by few.

Ageing

The detections themselves are permanent facts about a date and a place and do not decay. Everything else in the analysis does. Concession boundaries change hands, are extended, revoked and redrawn, so an attribution made against last year's layer may name a company that had already sold or surrendered the block. Protected area boundaries are gazetted, degazetted and downgraded – the pattern is well documented and politically significant – so an alert inside a park under today's boundary may have been outside any protection when it occurred, or the reverse. Tenure layers age worst of all, because customary land mapping is episodic and underfunded. Corporate resolution ages at the ordinary rate of mergers and renames. And the base data itself ages in a peculiar direction: the annual product is reprocessed each release, so an old extract diverges from the live dataset over time even though nothing changed on the ground. A stale record looks like a well-formed finding – a cluster of alerts, a named concession, a named company – where the boundary file was three years old and the concession had been transferred. The mitigations are to record the version of every layer used in an attribution, to re-run the overlay rather than trusting the stored result, and to date the boundary against the date of the alert rather than against today.

What this source feeds

A source is only worth what it lets you conclude. These are the disciplines that collect through it, the mission domains it serves and the data points it yields — every one is a tag, so you can follow any thread from here into the rest of the library.

Collected by these intelligence disciplines

Serves these mission domains

Environmental Crime

Yields these data points

How each sector uses Global Forest Watch

The same dataset is worked very differently depending on who you are, what authority you hold, and what you are ultimately producing. A military analyst is supporting a commander’s decision; a journalist is meeting a publication standard; an NGO caseworker is protecting a person. The records are shared — the constraints, thresholds and outputs are not.

🎖 Military and defence

Forest change is a terrain and stability variable before it is an environmental one. For planning and area assessment, alert clusters reveal new access routes, logging roads and clearings that change trafficability and concealment months before any map is updated, and the pattern of clearing frequently maps illicit economies – coca and other illicit cultivation, artisanal mining sites, informal airstrips – that a purely environmental reading would miss. In stabilisation and civil-military work, deforestation is a reliable early indicator of land conflict, displacement and armed-group revenue capture. The constraints are that the products detect change rather than presence, resolution limits what can be seen, and latency is days at best. Use it to cue collection and to maintain a change baseline over an area of interest, not as a surveillance capability, and expect the drier and more open terrain in many theatres to be poorly served by a canopy-based product.

🕵 National intelligence

For ENVINT and for the economic dimension of geopolitical work, this is a low-cost, high-integrity change detection layer over the entire terrestrial surface, with the specific advantage that it is not controlled by any state whose behaviour it might reveal. The standard products are monitoring of concession compliance, detection of clearance in areas where a government asserts none is occurring, and characterisation of illicit economies whose physical footprint shows up as forest change – mining, illicit crops, road building into remote areas. It also supports a class of denial-and-deception finding: where official statistics and satellite observation diverge systematically, the divergence itself is the intelligence. The tradecraft requirement is to be rigorous about the definitional gap between canopy loss and deforestation, because an assessment that elides it invites a technically correct rebuttal that will discredit the whole product.

👮 Law enforcement

For environmental crime units, forest police and prosecutors, the value is detection speed and the fact that the observation is independent of any interested party. Alerts inside a protected area or outside a permitted block generate patrol tasking and inspection priorities, and the annual product supports charging decisions about historical clearance. The evidential position needs early attention in most jurisdictions: satellite-derived alerts are typically introduced through expert evidence establishing the method and its error rates, and the concession or boundary overlay is the element most likely to be challenged, so obtain the authoritative boundary from the national registry rather than relying on the platform's copy. The practical case pattern is alert to inspection to seizure to prosecution, with the satellite record establishing when clearing began and how it progressed, and ground evidence establishing who did it. The data will not identify an offender and any case built as though it does will fail.

🔍 Private investigation and corporate security

In corporate investigations, supply chain due diligence and environmental litigation support, this is now close to mandatory background work. Buyers and financiers are subject to a tightening set of deforestation-free requirements, and the routine question is whether a supplier's landholdings, or the mills and estates in its supply shed, show clearing after a specified cutoff date. That is a well-defined query this source answers directly, and the answer is citable. It also supports the adversarial version: checking a company's sustainability claim against the observed record, which frequently produces a discrepancy worth putting to them. The professional care points are that a cluster inside a concession is not proof the concession holder cleared it, that plantation rotation is not deforestation, and that boundary vintage is the usual weakness. State the method and the cutoff explicitly in any report; the finding is only as strong as the polygon it was computed against.

📰 Journalism and OSINT media

For environmental and investigative reporting this is the workhorse: it is free, global, independent of the governments and companies being reported on, and it produces a visual that readers understand immediately. The durable story forms are clearing inside a protected area, clearing in a supply chain that a consumer brand said was clean, clearing on land claimed by an Indigenous community, and the gap between a government's published deforestation figure and the satellite record. The craft requirements are specific. Do not write deforestation when the data says tree cover loss unless you have established that it was natural forest. Do not attribute clearing to a company without establishing that the boundary was correct and current at the time. Get high-resolution imagery of at least one site before publication, because a single verified image is worth more than any aggregate, and take the finding to the company and the ministry with time to respond. Reporting on land defenders derived from this data should follow the security practices that community and defender organisations already use.

🌍 NGO, humanitarian and human rights

This is the source the conservation and land rights sector was built around, and the range of use is correspondingly wide: protected area monitoring, community-based forest monitoring where local monitors verify alerts on the ground, supply chain campaigning, litigation support, and documentation of encroachment on Indigenous and customary land. The near-real-time products are what enable intervention rather than commemoration – an alert acted on within a week can stop a clearing that an annual statistic can only record. Two considerations govern responsible use. First, incomplete tenure mapping means an area with no polygon is not unclaimed land, and analyses that treat unmapped areas as empty reproduce the exact erasure that customary tenure work exists to correct. Second, publishing precise locations of clearing on contested land can expose the community monitors and land defenders nearest to it, and the sector's own documented casualty figures make that a real rather than theoretical risk. Coordinate location disclosure with the people who live there.

🎓 University and research

The forest change products are among the most heavily used open geospatial datasets in environmental science, and the research programmes they support are mature: global and regional deforestation accounting, drivers of forest loss, carbon flux estimation, protected area effectiveness, the effect of certification and zero-deforestation commitments, and the methodological literature on the products themselves. Using them well requires reading that methodological literature rather than the platform's summary. The canopy cover threshold is a researcher decision with large downstream effects and must be stated. Product versions differ and must be cited by version, not by name. Validation error rates are biome-specific and should be propagated rather than assumed uniform. And the distinction between tree cover loss, deforestation, degradation and forest conversion is where most citation errors in the applied literature originate, so it is worth defining precisely in any paper that uses the data. The platform is also a legitimate object of study in its own right, as an infrastructure that shapes what forest governance can see.

Playbook: working Global Forest Watch end to end

A repeatable sequence from first pull to finished product. Each phase states what you are trying to establish, not merely what to click — the objective is a defensible chain of reasoning, not a completed checklist.

Phase 1 — Define the question in the data's own terms

Decide whether you are asking about canopy loss, about deforestation of natural forest, about degradation, or about a specific actor's conduct. Only the first is measured directly. The others require additional layers or additional evidence, and choosing the wrong one at the start produces an analysis that cannot be defended at the end.

Phase 2 — Fix the geometry before the data

Obtain the authoritative boundary for your area of interest – the concession from the national cadastre, the protected area from the gazetting instrument, the community land from the tenure record – and note who published it and when. The polygon is half the finding and it is the half most likely to be attacked.

Phase 3 — Establish the baseline

Pull the annual loss history for the area under a stated canopy threshold, and characterise what normal looks like: how much loss per year, in what pattern, in what season. Without a baseline you cannot say whether current alerts represent a change, and most alert clusters in an actively managed landscape are entirely routine.

Phase 4 — Separate natural forest from plantation and regrowth

Overlay primary forest extent and any available plantation mapping. Loss inside primary humid tropical forest is a different fact, legally and analytically, from a pulpwood harvest, and the number of published analyses that conflate the two is the reason this step is non-negotiable.

Phase 5 — Cluster the alerts and characterise the shape

Group alerts into events and look at the geometry. Straight edges and rectangular blocks indicate mechanised commercial clearing. Fingers along a road indicate incremental frontier expansion. Scattered small patches indicate smallholder or shifting cultivation. Riverine and slope-aligned patterns often indicate a radar artefact rather than clearing at all.

Phase 6 — Verify a sample against higher-resolution imagery

Take a representative subset of clusters and check them against the best available imagery. This is the step that catches systematic false positives, and it converts a modelled result into an observed one for at least part of your sample. Never publish an aggregate you have not sampled.

Phase 7 — Attribute to a boundary, not to a company

Establish which polygon the cluster falls in and whether that polygon was in force on the alert date. Then, and only then, resolve the polygon to a licence holder through the cadastre and corporate registries. The chain from pixel to company runs through two documentary steps and skipping either is where attribution errors happen.

Phase 8 — Test legality separately from occurrence

Determine what the applicable permit, management plan or protected area category actually allowed. Legal selective logging, legal conversion under a valid permit and legal plantation harvest all produce alerts. Illegality is a legal conclusion about a licensing regime, not an inference from a satellite.

Phase 9 — Bring in the adjacent evidence

Correlate with road construction, mining activity, mill and processing facility locations, export shipments, conflict incidents and adverse media. Clearing is the visible end of a chain, and the parts of that chain with named parties are where the case actually gets made.

Phase 10 — Assess the human dimension before publishing

Check tenure layers and consult community organisations where the clearing is on or near claimed land. Determine whether publication could expose local monitors or defenders, and adjust the granularity of location disclosure accordingly. This is a documented physical risk in several of the countries where the data is most useful.

Phase 11 — Record versions and archive the extract

Save the dataset version, the boundary layer version, the query geometry and the raw result. Annual reprocessing and boundary updates mean that a finding you cannot reproduce in a year is a finding you cannot defend in a year.

Phase 12 — Operationalise the watch

Register the polygons in watchlist.php with thresholds keyed to cluster size, protected area intrusion and deviation from the seasonal baseline, and route alerts to the party who can actually act – a patrol, a supplier compliance team, a regulator – within the window in which action is still possible.

The platform ships this as a step-checked workflow in playbooks.php, so progress is recorded against a case rather than held in someone’s head.

What to pair it with

No single source carries a finding. These are the datasets that corroborate, extend or contradict this one — and a source that contradicts is worth more than one that agrees, because it is the only thing that will tell you when you are wrong.

Source	Relationship	What it adds
Protected Planet (WDPA)	prerequisite	Authoritative protected area boundaries and IUCN management categories, which determine what activity was permitted at a location.
Copernicus and Sentinel imagery	corroborates	The underlying imagery for independent verification of a cluster, and the source of the radar data behind the cloud-immune alerts.
NASA FIRMS	extends	Active fire detections from thermal sensors, useful for separating fire-driven conversion from mechanical clearing and for timing.
LandMark	extends	Mapped Indigenous and community lands, essential context for any finding about clearing on customary territory.
OpenStreetMap	corroborates	Road, track and settlement data that explains access and frequently dates the arrival of infrastructure preceding clearance.
Global Witness	extends	Investigative documentation of who is behind resource-linked land conversion, and the standard reference on risks to land and environmental defenders.
EITI	corroborates	Licence registers and payment disclosures for extractive concessions, giving the documentary counterpart to observed disturbance in mining areas.
Trase and supply chain mapping initiatives	extends	Links between production regions, traders and consumer markets, converting a location-based finding into a supply chain finding.
UN Comtrade	corroborates	Timber and agricultural commodity trade flows for testing whether observed production is consistent with declared exports.

Legal, ethical and operational constraints

Collection and use of the data is unproblematic in itself: it is published openly for public use and the underlying observations are made by public earth observation programmes. The legal weight sits entirely in what you assert with it. Attributing clearing to a named company or individual is a defamation exposure in every jurisdiction, and the exposure is highest in precisely the countries where the reporting matters most, several of which combine claimant-friendly libel law with an established pattern of strategic litigation against public participation. The defensible attribution chain has three documented links: the alert cluster, the boundary that was legally in force on the alert date, and the registry record of who held that boundary then. Skipping any link converts a strong finding into an actionable allegation. Separately, an alert is not evidence of illegality, because a great deal of forest clearance is lawful under a permit; establishing that a clearing was unlawful requires the licensing regime, not the imagery. Where analysis concerns Indigenous or community land, free prior and informed consent norms apply to how you handle information about that land and its occupants, not just to the company operating there. And publication decisions carry a physical safety dimension: land and environmental defenders are killed at a documented and sustained rate, and precise location disclosure can raise rather than lower risk for people living at the frontier.

Operational security

Queries to an open environmental data platform are unremarkable and reveal little on their own. Two exposures deserve attention. The first is pattern: a sustained, narrowly focused collection against a single concession, from an attributable corporate or governmental network range, tells anyone with access to those logs which company you are examining, before you have published anything. The mitigation is to collect a wider region than you need and filter locally, which costs almost nothing given the aggregation endpoints. The second is downstream and more serious. Findings derived from this data are frequently about powerful actors operating in remote areas, and the people best placed to verify a cluster on the ground are also the people least able to protect themselves from the response. Any workflow that involves local verification should be designed with the verifier's security first: minimum necessary specificity in communications, no unnecessary retention of who checked what, and disclosure timing agreed with them rather than driven by your publication schedule. Treat the identity of ground verifiers as the most sensitive element of the entire investigation, because it is.

Two rules that hold regardless of jurisdiction. Collection that is lawful is not automatically proportionate, and a dataset assembled for one purpose does not carry consent for another. Where the records concern identifiable people, the question is not only whether you may hold the data but whether holding it serves the purpose you are accountable for.

Is it earning its place?

Sources accumulate. Feeds get added during an incident and are never reviewed again, and a decade later the pipeline is carrying dead weight that nobody dares remove. These are the measures that show whether Global Forest Watch is contributing anything, and they are worth baselining now so the answer is available later.

Alert-to-verification ratio: the proportion of clusters you triaged that were checked against high-resolution imagery or ground reporting, which measures whether your pipeline is producing findings or just volume.
False positive rate of your own triage rules, measured on the verified sample and tracked over time, per biome and per alert system.
Median time from satellite acquisition to an actionable notification reaching the party who can respond, which is the only metric that determines whether near-real-time capability is being realised.
Proportion of alert clusters that could be attributed to a boundary whose vintage was current at the alert date, which measures the health of your contextual layers rather than the satellite data.
Concession layer freshness: the age distribution of the boundary files you are relying on, reviewed quarterly, because this is where attributions silently rot.
Number of cases in cases.php where a forest change record supplied the initiating observation, as distinct from illustrative background.
Reproducibility of published figures against the current dataset version, which tells you how exposed your past outputs are to annual reprocessing.
Share of monitored area that has near-real-time alert coverage at all, so that a quiet period can be distinguished from an unmonitored one.

Beware of volume. Indicator counts rise easily and say almost nothing. Unique contribution — findings this source produced that no other source in your stack would have — is the measure that matters, and it is usually far lower than anyone expects.

Tradecraft notes

The distinctions that separate a competent analyst from a fast one:

Tree cover loss is not deforestation and the distinction is not pedantry – it is the single most common way analyses using this source are successfully rebutted. Establish forest type before you use the word.
The canopy cover threshold is a choice you are making, whether or not you know you are making it. State it, and run the analysis at a second threshold to see how much your conclusion depends on it.
Cluster shape carries information no attribute field contains. Rectangles mean machines, fingers mean roads, scatter means smallholders, river-aligned bands usually mean water rather than logging.
An alert is dated to acquisition, not to the event. Under persistent cloud the true start of clearing can precede the first optical alert by weeks, which matters enormously when the question is whether clearing began before or after a cutoff date.
Check whether the boundary was in force on the alert date, not today. Concessions transfer and protected areas are degazetted, and an attribution against a current layer can name entirely the wrong party.
Use the confidence field rather than ignoring it. Multi-system corroboration materially reduces commission error, and the right posture is to alert on low confidence but act on high.
Never compare figures across annual product versions. Each release reprocesses the whole series, and cross-version comparison is the standard way to manufacture a trend that does not exist.
Verify a sample with high-resolution imagery before every publication. One confirmed site does more for a finding's credibility than a hundred thousand hectares of unchecked aggregate.
Where the finding touches contested land, the security of local verifiers is part of the analysis, not an afterthought to it. Decide location granularity with them, not for them.

Questions analysts actually ask

Does a tree cover loss pixel mean deforestation?

No. It means canopy cover was lost from any cause – harvest, fire, storm, disease, legal logging or clearance. To make a deforestation claim you need to establish that the area was natural forest and that the loss was conversion rather than a management cycle.

How quickly do alerts appear after clearing?

Days, in good conditions. Under persistent cloud, optical alerts can take weeks; the radar-based product is far less affected and is the reason the integrated layer works at all in equatorial regions. Always distinguish the acquisition date from the date the clearing began.

Can I identify who cleared the forest?

Not from this data. You can identify which mapped boundary the clearing falls inside, and then resolve that boundary to a licence holder through a cadastre and a corporate registry. Both of those are documentary steps outside this source, and both must be dated to the time of the clearing.

Why do my numbers differ from last year's published figure?

Most likely because the annual product was reprocessed. Each release revises the whole series rather than appending to it. Cite the version you used, archive the extract, and never compare figures across versions.

Are alerts inside a protected area evidence of a crime?

No. Protected area categories permit different activities, some including extraction and community use, and some clearing is lawfully permitted. Establish the category, the management plan and any permits in force before treating an alert as an offence.

Why does an area I know is being logged show no alerts?

Selective logging under an intact canopy is largely invisible to these detectors, small clearings can fall below detection size, and some regions lack near-real-time coverage entirely. Absence of alerts is weak evidence of absence of activity and should never be reported as a clean bill of health.

Which alert product should I rely on?

Use the integrated layer for triage, because it reconciles the individual systems and carries a confidence field. When a specific finding matters, go back to the contributing system, because the error modes differ – optical products fail on cloud, radar products fail on water and terrain.

How do I use this for deforestation-free supply chain checks?

Define the supply shed or estate boundaries, fix a cutoff date from the applicable regulation or commitment, compute loss after that date within those boundaries under a stated forest definition, and verify a sample. Document all four choices, because each of them changes the answer.

Does the platform generate any of this with a model?

The forest change products are remote sensing algorithm outputs from the original producers and are ingested as published, with their derived status recorded. Within the platform, the only language model involvement is the Summarise skill in copilot.php, which writes prose over existing records and creates no detection, relationship or attribution.

Standards, formats and interoperability

What this source speaks natively, and what it has to be translated into before a partner can consume it. Work that arrives in a recognised format is easier to defend, easier to hand over and easier to automate against:

GeoTIFF and Cloud Optimized GeoTIFF for the raster loss and alert products, which is how the pixel-level data is actually distributed and consumed.
GeoJSON, shapefile and vector tile formats for concessions, protected areas, tenure and administrative boundaries.
OGC web service standards for map and feature services, which is how the layers are consumed by external GIS clients.
The IUCN protected area management categories, which define what activity a designation actually permits and therefore whether an alert is anomalous.
FAO forest definitions and national forest definitions, which differ from the canopy-cover threshold used here and are the source of most reconciliation disputes with official statistics.
Emerging deforestation-free regulatory frameworks in major consumer markets, which set cutoff dates and geolocation requirements that this data is routinely used to test against.
STIX 2.1 and MISP as platform export formats, with clusters mapped to location and observed-data objects for sharing with partners.
Accuracy assessment conventions for land cover change products, which define the omission and commission statistics you should be quoting rather than describing accuracy qualitatively.

References

Primary documentation and authoritative references for this source. Publishers revise and retire material, so treat the retrieval date as part of the citation and re-check before relying on any of it in a formal product.

Global Forest Watch — World Resources Institute. The map, data portal and methodology documentation. Read the per-dataset methodology pages before using any layer in a finding.
Global Forest Watch Data API — World Resources Institute. The programmatic interface, including aggregation of alerts by submitted geometry, which is the core operational query.
World Resources Institute — WRI. The operating institution, its research on forests and land use, and the annual analyses that interpret the loss data.
Protected Planet — UNEP-WCMC and IUCN. The World Database on Protected Areas, the authoritative source for protected area boundaries and management categories.
Copernicus Programme — European Union. Sentinel-1 radar and Sentinel-2 optical imagery, which underpin the near-real-time alert products and support independent verification.
NASA FIRMS — NASA. Active fire detections, the source of the fire alert layer and a useful independent signal for fire-driven conversion.
LandMark — LandMark partnership. Global mapping of Indigenous and community lands, the essential tenure context for clearing on customary territory.
Trase — Trase. Commodity supply chain mapping linking production regions to traders and consumer markets, extending location findings into supply chains.
Food and Agriculture Organization of the United Nations — FAO. Global forest resources assessments and the forest definitions that national statistics use, which is where most reconciliation disputes originate.
Global Witness — Global Witness. Investigations into who profits from forest conversion and the annual documentation of killings of land and environmental defenders.

Link integrity: every reference above was verified with a live request when this page was generated. Where a publisher had moved or withdrawn a document, the link was repointed at a preserved copy in the Internet Archive and marked as archived. Anything with no reachable copy anywhere had its link removed rather than left to rot — the source is still credited, it simply cannot be linked.

Put it into practice

The Quantus Intel threat intelligence platform operationalises this source: Forest change alerts are clustered into events, stamped with versioned concession, protected area and tenure context at ingest, resolved to licence holders through registry and disclosure sources, and watched against thresholds that route to whoever can still act.. Browse the full source catalogue, or follow any tag above into the rest of the library.