Galamsey Monitor

It ranks all 260 district assemblies (MMDAs) by how strongly galamsey (illegal small-scale mining) pressure shows up in the satellite record, so scarce inspection and enforcement resources can be pointed where the evidence is strongest.

It is a triage and targeting aid — not a determination of galamsey and not enforcement evidence. A high rank means "look here first." It does not, on its own, prove that illegal mining is happening in a district, nor identify who is responsible.

Each district gets a galamsey-pressure score from 0 to 100 — a relative national ranking, not a percentage of land mined. Districts are grouped into action tiers:

• P1 Verify first — score ≥ 70. Satellite evidence is strongest; put these districts at the top of the field-verification queue.
• P2 Investigate — ≥ 50. Schedule verification within the current planning cycle.
• P3 Monitor — ≥ 30. Track through the next imagery release.
• P4 Watch — ≥ 15. Low satellite signal; watch on the next update.
• Baseline — below 15. No detectable galamsey signal (not evidence of absence — see limitations).

Grey districts — "no documented site" is a data-quality flag, not a verdict. A district with zero documented sites in the register can still have real galamsey on its river banks; the register is curated public sources, not an EPA-wide database. To keep the map honest without flat-lining those districts, the grey ramp now carries the same raw satellite signal it always did, split three ways:

• No docs · verify — no documented site, but the raw satellite metric puts the district at score ≥ 50. One of three things is going on: galamsey is happening and the register hasn't caught it yet, a licensed large-scale mine that our subtraction hasn't fully removed, or heavy plantation / logging clearing along river corridors. These districts should be verified before being dismissed — open one and use the Sentinel-2 EO Browser link in the drill panel to look along the rivers.
• No docs · worth checking — no documented site, raw score 30-50. Loss and river-adjacency are non-zero but not striking; the driver could be galamsey, river-bank farming, or logging. Worth a quick look.
• No docs · no signal — no documented site AND raw score below 30. Neither the register nor the satellite metric flags the district. Still not evidence of absence: pits under canopy or activity after the loss window can't be seen.

The tier is an action priority, not a verdict. A district in P1 is not "confirmed galamsey"; it is where officers should look first.

Three components, each normalised across all districts (robust 5th–95th-percentile scaling), then weighted and rescaled to 0–100:

• Forest-loss intensity — 50%. Hansen Global Forest Change tree-cover loss since 2015, as a share of the district's wooded area. Built-up land is excluded (ESA WorldCover) so town growth and urban redevelopment are not counted as galamsey.
• Riverine concentration — 30%. The share of that loss falling within a river buffer (currently the HydroSHEDS Level-8 network used in the riverine-share weight — toggle "Rivers used in the riverine-share weight" on the map to see exactly which reaches are counted). Galamsey is alluvial (it follows rivers for gold and water), so a high riverine share is the single strongest signal separating galamsey from ordinary farm or forest clearing. Known gap: some small tributaries and named local streams (e.g. Wewi through Ayigya, Kumasi) are absent from HydroSHEDS at any level and therefore not seen by the riverine-share weight today. The next producer release unions the network with OSM waterway=stream (17,000+ additional Ghana streams already staged) to close this gap; the fix is coded but requires a GEE re-run to take effect.
• Documented-site density — 20%. Recorded galamsey sites per unit area, from a curated register.

Worked example — a Priority-1 district typically has ~15–25% of its wooded area lost since 2015 with a large share of that loss along rivers, plus 3+ documented sites. The tier is a triage priority; ground truth still requires field or high-resolution verification.

• Hansen/UMD Global Forest Change v1.13 — annual 30 m tree-cover-loss record (UMD, Google, USGS, NASA); the dataset Global Forest Watch and peer-reviewed galamsey studies use.
• ESA WorldCover 2021 — 10 m global land cover; supplies the built-up mask.
• WWF HydroSHEDS — river network for the 1 km riverine buffer.
• Curated galamsey-site register — documented site locations (name, district, year, source).
• Maus et al. (2022) global mining polygons v2 — Sentinel-2-derived visible-mining polygons. Scientific Data 9, 433. DOI 10.1038/s41597-022-01547-4. Dataset 10.5281/zenodo.7211861 · CC BY 4.0. Used to subtract regulated LSM from the score.
• Ghana Chamber of Mines 2024 factsheet — ghanachamberofmines.org. 18 producing-member concessions, emulated dodecagon polygons from public centroids; fallback where Maus 2022 doesn't cover. NOT surveyed licence boundaries.

Earth Engine runs the analysis offline; the server only serves the finished ranking. Reductions run at 100 m (native 30 m times out over 260 districts and does not change the ranking).

Use it to prioritise: which districts to inspect, monitor, or resource first; which regions carry the heaviest cumulative pressure. Before any regulatory or enforcement step against a specific site, confirm with field inspection or high-resolution imagery — the monitor points you to the district, the ground truth comes from verification.

Can: give a consistent, reproducible, nationwide picture of where galamsey-type vegetation loss concentrates, updated each year.

Cannot:

• Attribute loss to a specific cause — it cannot separate unlicensed galamsey from farming, logging or fire with certainty (the riverine-share weight + the built-up exclusion + the licensed-mining subtraction reduce, but do not remove, this).
• Detect small pits hidden under canopy, or loss in the most recent months (Hansen has a ~1-year lag).
• Prove wrongdoing or identify operators — it is not evidence.
• Rank a district highly on documented sites alone — that register is curated, not exhaustive.

Licensed-mining subtraction — coded, not yet applied to the served ranking. The next producer re-run will subtract regulated large-scale-mining polygons before the forest-loss and riverine-share weights are computed, so the score will measure unlicensed vegetation-loss pressure rather than all mining pressure. Sources: Maus et al. (2022) global mining polygons v2 (Scientific Data 9, 433, DOI 10.1038/s41597-022-01547-4; dataset 10.5281/zenodo.7211861, CC BY 4.0) as the primary source; the Ghana Chamber of Mines 2024 factsheet (~18 producing members, emulated dodecagon polygons around publicly-disclosed centroids — not surveyed licence boundaries) as fallback for concessions Maus does not detect. Superseded when the Minerals Commission GCMRS shapefile is obtained. Toggle "Licensed mining concessions" on the map to see exactly which footprints are subtracted.

Weight-calibration caveat. The 50 % / 30 % / 20 % weights on the forest-loss, riverine-share and documented-site components are expert-elicited priors, not empirically calibrated against a held-out validation set. They reflect a defensible reading of the peer-reviewed Ghana galamsey literature (Snapir 2017, Forkuor 2020, Owusu-Nimo 2018) but should be revisited if EPA can supply a labelled ground-truth sample. A weight-jitter bootstrap to publish rank confidence intervals is on the producer-side backlog.

Hansen Global Forest Change reports where woody canopy was lost — it does not tell us why. Cocoa expansion, plantain and yam clearing, oil-palm plots, teak plantations, logging concessions and shifting agriculture all show up on the same layer as illegal alluvial mining. The Monitor is defensible because it uses four converging discriminators, each doing part of the work:

1. Riverine concentration — 30 % of the score. Alluvial galamsey is disproportionately within ~300 m of a stream (dredging in streambeds, sluicing at the bank). Farm and plantation clearing spreads across a landscape irrespective of stream buffers. This weight multiplies each district's forest loss by its share falling inside a 300 m HydroSHEDS + OpenStreetMap stream buffer, so upland cocoa or farm clearing scores far lower than the same tonnage of loss hugging a river.
2. Built-up exclusion. ESA WorldCover 2021 "built" plus Google Dynamic World "built" are subtracted before the forest-loss weight is computed, so urban tree loss and construction pads do not enter the score.
3. Licensed-mining subtraction — coded, applies on next producer re-run. Maus 2022 primary polygons and emulated Chamber of Mines 2024 concessions will be subtracted so leases operated with a Minerals Commission licence do not inflate the forest-loss or riverine-share weight. Until then, a "licensed LSM" tag appears on the drill panel of the ~15 districts hosting known large-scale leases so a reader knows to discount the top-line score there.
4. Documented-site density — 20 % of the score — plus no-documented-sites grey suppression (wave-6). A district that has never had a galamsey site documented in the sanitised register is painted neutral grey on the choropleth even if its forest-loss and riverine-share weights are non-zero. The raw score is still shown in the drill panel so nothing is hidden, but the map no longer implies galamsey where the field record does not support it. This is what keeps the Volta and Oti highlands (Afadzato Range, Hohoe, cocoa-and-plantain hills) from lighting up as galamsey when the loss there is farm and plantation expansion.

What still slips through: agriculture along riverbanks (rice paddies in the Northern flood plain, tomato and onion farms along the Volta, small-holder cassava at river bends) can hit both the forest-loss and the riverine-share weights without being galamsey. In our current stack, those districts also need at least one documented site in the register to be painted with a pressure tier — otherwise the grey suppression catches them. If the register is incomplete for a district (highly likely in remote areas), the right-of-reply link on each drill panel lets an officer or community member submit a site so the ranking updates on the next producer run.

What we can add later (producer roadmap, not yet shipped):

• Height-above-nearest-drainage mask (below 5 m). From Copernicus GLO-30 + HydroSHEDS. Alluvial galamsey concentrates in the flat, low-lying corridor next to rivers where height above drainage is below 5 m; cocoa and rubber sit on hillslopes above 15 m. This would defensibly drop Volta highland Hansen loss out of the forest-loss weight even if the register later grows. Highest-value single upgrade; approx. one producer-side sprint.
• Post-loss NDVI recovery test. For each Hansen loss pixel, sample Sentinel-2 NDVI two years after the loss. If NDVI recovers above ~0.5 the pixel was replanted (fallow-farm cycle); if it stays below ~0.3 the pixel is likely a bare pit. Retain only non-regrown loss. More Sentinel-2 aggregation, but very defensible.
• Turbidity co-signature. Only credit a river-adjacent loss pixel to the riverine-share weight if the downstream river channel in the same year shows a Sentinel-2 red-band spike (suspended-sediment plume). Directly evidence-based; distinguishes river-adjacent farming from active pit runoff.
• JRC Tropical Moist Forest driver classes. Cross-reference the JRC annual transitions layer where it flags "small-scale disturbance" vs "commodity" drivers. Partial national coverage (tropical belt only).
• Cocoa mask subtraction. Subtract Kalischek 2023 cocoa map plus rubber and oil-palm concessions. Straightforward but coverage of the Kalischek map varies.

Honest posture for the demo. The Monitor is a triage layer, not an attribution model. The riverine-share weight + built-up exclusion + licensed-mining subtraction + documented-site density + wave-6 grey suppression together make the score defensible for prioritising verification effort. It cannot on its own separate galamsey from agriculture pixel-by-pixel — that is what the height-above-drainage + post-loss NDVI-recovery roadmap items add when EPA can fund the producer-side compute.

The ranking is a dated snapshot, shown at the top of this page ("data current as of…"). It reflects Hansen forest loss through 2024. It is rebuilt from Hansen's annual release, so the natural refresh cadence is once a year; between releases the forest-loss backbone does not change.

Published by Resilicity GmbH as a decision-support pilot for the relevant Authority of Ghana. This layer is not an official EPA product, not a determination, not a permit decision, and not enforcement evidence. It has not been endorsed by the Government of Ghana. It complements — it does not replace — the statutory roles of the EPA, Minerals Commission, Forestry Commission, Water Resources Commission and the security services.

Enforcement authority for mining is fragmented across the Minerals Commission (licensing), the Ghana Police / Ghana Armed Forces (Operation Halt / Halt II), the Forestry Commission (reserves), the District Security Councils (REGSEC/DISEC) and the EPA (environmental compliance). A high triage score for a district is not a finding against the District Chief Executive or the District Assembly.

The galamsey-affected sites register does not name any corporate licence-holder. Documented sites identify galamsey-affected localities only; they do not attribute activity to any specific operator.

The separate "licensed mining concessions" layer DOES name licensed operators (AngloGold Ashanti, Gold Fields, Newmont Ghana Gold, Zijin Mining Group, Asante Gold Corporation, Chifeng Jilong, Perseus, Galiano, Endeavour, Xtra-Gold, Ghana Bauxite Company, Ghana Manganese Company, etc.) — but does so favourably: those polygons are subtracted from the pressure score so those districts rank lower than the raw satellite record would suggest. Every named operator carries an operator_as_of vintage; if a listed operator has since divested, correction via the Right of Reply link below is actioned within 14 days.

The concession polygons in that layer are emulated dodecagons around publicly-disclosed operational centroids at a 5 km conservative buffer — not surveyed licence boundaries. They will be superseded by the Minerals Commission GCMRS shapefile when that dataset becomes available.

Concerns about the ranking of a specific district, or the operator attribution of a specific concession, may be raised via the Right of Reply page or the mailto link in that district's detail panel. Response target: 2 working days for acknowledgement, 14 calendar days for a substantive reply; 48 hours for take-down of an alleged defamation / DPA §36 claim (interim, pending review).

• Galamsey — illegal / unlicensed small-scale mining, typically alluvial (riverine) gold.
• MMDA — Metropolitan / Municipal / District Assembly; the local-government unit ranked here.
• Tree-cover loss — Hansen's mapped removal of tree canopy in a given year.
• Riverine share — the fraction of a district's loss that sits within 1 km of a river.
• Built-up mask — exclusion of land classed as built (towns, roads) so urban clearing is not counted as galamsey.
• Proxy — an indirect measure that stands in for the thing of interest; here, vegetation loss standing in for galamsey pressure.

Why does urban Accra score low? Built-up land is excluded, so city tree loss (construction, drains) is not misread as galamsey.

A district I know has galamsey ranks low — why? The activity may be recent (after the data window), under canopy, or small relative to the district. Treat the ranking as a starting point, not a closed list; report on-the-ground knowledge.

Can I use this to prosecute or refuse a permit? No — it is a proxy for targeting. Any action needs independent field or high-resolution verification.

Why annual updates, not live? The authoritative forest-loss data (Hansen) publishes once a year; a "live" number would not be more accurate, only noisier.