Mineral prospecting on a shoestring

A registered gold claim sits in the Mberengwa (Belingwe) greenstone belt near Zvishavane, Zimbabwe. The holder has walked the ground, has the Geological Survey sheet, and has zero budget for trenching, geophysics, or commercial remote sensing. The question is not "can satellites find gold" (they cannot) but a narrower, answerable one: can free satellite and DEM data narrow ~300 hectares down to the few hundred metres worth sampling next?

This report works that question end to end on free, cloud-native data, with every figure produced by a re-runnable folia.fetch recipe. It is organized as the holder's three questions, plus two later arms: a multi-decadal disturbance time series, and the one result here with measured skill, an AlphaEarth embedding analog validated against the belt's known gold.

The strongest result first. Most of this report is honest prioritization with no ground truth: anomaly layers we cannot yet grade. The exception is the embedding analog. Trained on 85 confirmed gold occurrences and read directly from open AlphaEarth COGs on Source Coop, it shows real, held-out skill: an unseen gold site still lands at the 99th percentile of background, AUC 0.96 (0.5 is chance). Applied to the claim it reads intermediate, above country rock but below the gold band, which is a resemblance diagnostic, not evidence that gold is present. It is the first validated, skillful result in the whole study.
True-colour Sentinel-2 over the claim
Sentinel-2 true colour over the claim, 2025-06-21 (0% cloud). The greenstone belt reads as the darker vegetated ground; granite surrounds it.

The location, and a correction worth seeing

The holder supplied corner coordinates labelled UTM Zone 36S. Taken at face value, the claim centroid lands at ~36°E on the Mozambique coast, about 615 km from Zvishavane. The eastings near 30°E are a Zone 35S signature: read as EPSG:32735 the claim sits ~19 km north of Zvishavane, in the greenstone belt, which the supplied geology sheet confirms. The Sentinel-2 tile that covers it is T35KRT, MGRS zone 35, an independent check.

One wrong zone label silently relocates the whole analysis 615 km. The fix is to state the CRS explicitly and check it against a known landmark. The corrected AOI is built reproducibly in research/mineral-prospecting/aoi/build_aoi.py.

Data sources

All free. Tier-1 was already in the Folia catalog; Tier-2 was registered for this report. Heavy reads use windowed folia.fetch (only the COG blocks touching the AOI), so the whole study pulled about 15 MB, not gigabytes.

Source Folia id Role
Sentinel-2 L2A @esa/sentinel-2/l2a iron oxide, broad alteration, vegetation + cloud masks
Landsat C2-L2 @usgs/landsat/c2-l2 cross-check; long historical record
Copernicus DEM GLO-30 @esa/copernicus-dem/glo-30 structure: hillshade, slope, curvature, lineaments
Sentinel-1 RTC @esa/sentinel-1/rtc radar texture, cloud-proof structural read
ASTER L1T @nasa/aster/l1t pre-2008 SWIR for real alteration-mineral ratios
EMIT L2A @nasa/emit/l2a-reflectance hyperspectral ceiling (coverage check)
NISAR GCOV @nasa/nisar/gcov L-band, vegetation-penetrating (coverage check)
AlphaEarth Embedding Fields v1 @tge-labs/aef (Source Coop COGs) 64-D learned embedding; the validated analog

Block and claim assessments

Subsequent assessments apply the validated arms to individual claim blocks. Each page is a where-to-walk verdict for that block only.

Block Size Fused P(gold) Short read
Kwekwe block 157 ha 0.129 — 8th pct No peg-worthy signal; block-scale surface reads as ordinary country rock, best corner ~0.2
Kwekwe block 2 ~358 ha 0.254 — 45th pct Average block overall; NW corner is the outlier at P(gold) 0.65–0.70, the highest spot in the set
Belingwe-West chain 8 adj. blocks 0.394 — highest of all sites Above-background across the chain; strongest fused result in this study
Kwekwe block 3 ~93 ha 0.285 — 36th pct Terrain arm at 98th percentile (strongest single-arm terrain read in the set); NE corner is the walk-first ground
Kwekwe partners block ~150 ha 0.229 — 43rd pct Terrain positive (82nd pct) but iron-oxide below background on fully bare ground; north interior is the walk-first corridor
Kwekwe reef prospect ~195 ha window 0.228 — 40th pct Old-map "red oxide reef", pegging pending; P(gold) climbs NE→SW along the reef trace to a 0.55–0.59 cluster at its SW terminus — walk-first ground. No ferric signal on the trace (13th pct)
Silobela point sites 2 × 1 km² windows 0.358 — 71st pct (site 2) Two supplied coordinates (location pending confirmation vs the reef prospect); site 2's exact point at the 77th pctile with a 0.74 spot 300 m NNW — the strongest point reading outside Belingwe-West
Silobela block 3 29 ha 0.346 — 69th pct The only free ground next to site 2; the site 2 analysis's best spot (P 0.73) falls inside this block. Ground-truthed 2026-07-30: five rock grabs 31 m from that spot assayed 2.33, 1.52, 1.34, 0.16, 0.02 g/t Au — the first grade-bearing test of a model pick, and it backs the pick. Holder has since registered a 25 ha block here
Kwekwe block 4 ~1,561 ha 0.238 — 34th pct Exploration-scale block 10 km north of block 2; background overall but all six top spots (0.62–0.71) cluster in one ~1.5 km west-central pocket — recon that pocket before committing
Mvuma cluster 4 × 36 ha windows 0.325 — 82nd pct (block 3) Four owned blocks in the Mvuma district (Falcon group 11 km E), first site east of 30°E. Eastern pair carries it: block 3 the strongest whole-block ground, block 2 the hottest supplied point (0.382, 90th pct); block 4 background. Blocks 1–3 read in the same band as the assayed Silobela ground

What the arms resolved

Arm Short answer Detail
Embedding analog (validated) The strongest result, and the only one with measured skill. An AlphaEarth 64-D embedding analog, trained on 85 confirmed gold occurrences and read from @tge-labs/aef COGs on Source Coop via HTTPS range reads, recognizes held-out gold at leave-one-out AUC 0.96 (0.5 is chance); an unseen gold site lands at the 99th percentile of background. Applied to the claim it reads intermediate (71st percentile, above country rock and below the gold band), a resemblance diagnostic, not evidence of gold. Embedding analog →
1. Gossans / iron oxide Yes, as a prioritization layer. S2 B04/B02, B11/B08, Crosta PCA. Laterite is the main false positive. Iron oxide & gossans →
2a. Alteration without a spectral library An anomaly, yes; a mineral ID, no, not from Sentinel-2. ASTER (5 pre-2008 scenes here) and EMIT reach the spectral ceiling, but EMIT found no alteration mineralogy in either the wet (2024-02-23) or the dry (2024-11-04) season: the apparent wet-scene Al-OH signal in the NW corner was shown by the dry-season check to be a bright-soil and dry-litter artifact, not white mica. Structure (DEM) is the robust free layer. Alteration mapping →
2b. Gold-bearing structures The strongest free-data structural layer. DEM lineaments resolve below the 1:100k map; dominant NW-SE grain. The layer now uses the FABDEM bare-earth DEM (canopy stripped), which confirmed the lineaments are real structure and left the targets unchanged. The promised upgrades (PALSAR-2 L-band, NICFI) were tested too: PALSAR-2's free annual mosaic gives texture only, not sub-canopy structure, and NICFI was license-excluded. Neither changed the conclusion. Structure & lineaments →
Multi-decadal disturbance A 1987-2024 Landsat bare-patch time series, the temporal signature of old workings. No persistent bare patch falls inside the claim (the interior is stable vegetation); the 99 candidates cluster in the surrounding belt and are most likely roads, outcrops, and settlement, not workings. Prioritization, never detection. research/mineral-prospecting/disturbance/
3. Free tools for a beginner QGIS, Copernicus Browser, GEE, STAC + Python, SNAP, and the tradeoffs. Tools & platforms →

A shared gold occurrence reference set underpins the validated arm: 85 confirmed gold occurrences in the belt (Bartholomew/Blenkinsop ZGS national inventory, with production and grade), assembled and de-duplicated against MRDS, mindat, and OSM. The nearest confirmed gold is 11 km east of the claim, where the greenstone belt proper begins. These are the positives that trained the embedding analog, and they are plotted as a toggleable gold layer on the interactive map. Recipe: research/mineral-prospecting/occurrences/.

Honest framing

Prioritization, not discovery. Every layer here is a where-to-walk-first surface, never a mineral map and never "dig here". The greenstone gold here is orogenic (sericite-chlorite-carbonate, structurally controlled), not the epithermal alunite-kaolinite style most online tutorials assume. Tropical laterite and dense canopy are the dominant confounds, so indices are relative and vegetation-masked.
First ground truth arrived 2026-07-30. For most sites these anomaly maps remain unsupervised. The exception: five assayed rock grabs now exist at the Silobela block 3 walk-first spot (2.33 to 0.02 g/t Au, three of five at or above 1.34 g/t, taken 31 m from the model's top-ranked point). That is one point, sampled because the model ranked it, so it calibrates targeting, not grade. The calibration hook in the recipes now has its first labelled points (in-situ/samples-2026-07-30-hashu.csv); low-scoring control samples are the next ask of every holder.

Reproducibility

The recipes live in research/mineral-prospecting/ in the Folia repo, one directory per arm, each a working Python script over folia.fetch. Re-running any arm is a cache hit (windows are keyed by AOI), so it costs no new download. One implementation note surfaced along the way: folia.fetch currently returns only the first band of a multiband COG, which matters for ASTER (its SWIR ships as one six-band file); the alteration recipe documents the workaround.