Belingwe-West chain: a first read

A third case study, and the newest. A holder handed over corner pegs for eight adjacent claim blocks in the Midlands greenstone belt of Zimbabwe, walked on the ground the day before, and asked the same question the Zvishavane and Kwekwe holders did: how does it look on satellite data?

The pegs arrived as bare eastings and northings with no zone label. Read as UTM 36S they land in Mozambique; read as UTM 35S (EPSG:32735) they form a tight 3.0 × 3.2 km chain at ~28.82° E, −19.78° in central Zimbabwe. 35S is the only consistent read, the same label-free correction the other two sites needed. That single interpretation, shown in the open, is the difference between walking the right ground and being 600 km off.

Bottom line: the most gold-like of the three claims we've studied, though still only a prioritization, not a discovery. Four methods plus a regional prospectivity model all read the chain as above-background: mildly iron-stained (ferric 65th percentile), structurally complex (lineament density 85th, NW quadrant 91st), on mine-like terrain (69th, terrain model AUC 0.87), and — the strongest signal — the highest absolute regional P(gold) of the three sites (0.394, vs Zvishavane 0.282 and Kwekwe 0.129). The regional model, trained on 333 gold deposits across the central greenstone belts and spatially-cross-validated to 15 km (AUC 0.93), ranks this chain essentially level with Zvishavane and far above Kwekwe's background. The honest caveats still bind: every method reads at 10–80 m pixel scale, so this is evidence about block-scale ground, not a narrow vein or small outcrop; ferric staining is not a local gold vector here (nearby mines sit at background ferric); and P(gold) reflects greenstone-belt surface geology, which is necessary but not sufficient for gold. The spots below are best-case corners to walk first, never "dig here." One real ground sample from the chain would re-weight every layer.

The ferric spots, as a list (UTM 35S, the holder's grid):

What the methods say

Four analyses plus the regional model, each scored over the claim vs its local background. The purple points on the map are the regional model's best-case ground; toggle on the ferric spots to compare.

Method	Signal on the chain	Read
Red-oxide ferric (S2)	65th percentile	mildly stained, but nearby mines are background-ferric → not a local gold vector
Structure / lineaments (DEM)	85th pct (NW quadrant 91st)	structurally complex ground — the real control on orogenic gold
DEM terrain match	69th pct (model AUC 0.87, spatial-CV holds)	mine-like terrain type
Regional P(gold) (fused)	0.394 — highest of the 3 sites	above-background; strongest signal

The regional model

The four arms are local reads. The regional prospectivity model is the synthesis: a RandomForest trained on 333 known gold deposits across the central Zimbabwe greenstone belts (the province encompassing all three sites we've studied), scored on terrain + structure + AlphaEarth-embedding features. It is spatially cross-validated to 15 km (AUC 0.93) — held-out mines with no training mine within 15 km still score correctly, so the signal is regional surface geology, not memorised locations.

Scored against that one model, the three claims rank:

Site	Regional P(gold)	Verdict
Belingwe-West (this chain)	0.394	most gold-like ground of the three
Zvishavane	0.282	strong (the original "better parcel")
Kwekwe	0.129	background — not gold-like

An honest note on the model: ~89% of its skill is the AlphaEarth embedding, which reads surface appearance. A Landsat 1985→2022 change test found the known mines show no elevated disturbance vs background, and the embedding signal generalises across space (holds at 15 km) — two independent checks that it is reading geology, not mining scars. The conservative terrain-only floor (AUC 0.81) and the embedding ceiling (0.93) bracket the confidence. Recipes: research/mineral-prospecting/model/.

The eight blocks

The pegs form a chain, not a cluster: ordered as supplied they march north-west to south-east across ~3.2 km, the signature of pegging along a structure (a reef or shear strike) rather than gridding an area. Total ~26 ha over eight blocks, each triangle or quad 0.5 to 5.7 ha. The ferric best-case spots concentrate in the south-east half of the chain (blocks 4 to 7), where it bends SE; the northern blocks (1 to 3) produce no top spots.

What the map shows (iron-oxide arm)

Ferric ratio (Sentinel-2 B04/B02) is the holder's red-oxide thesis, computed on a 0%-cloud dry-season scene with vegetation and cloud masked. The ground is fully exposed (0% vegetation-masked), so the optical read is clean. Over the chain it is modestly elevated, 65th percentile of the surrounding terrain.
The mine calibration is the honest test. The ferric ratio is sampled at the 38 known gold mines within 25 km on the same scene. They sit at −0.08σ, indistinguishable from background, so on this belt a bright ferric patch is weak evidence for gold. A bright chain does not inherit meaning it hasn't earned.
Gold mines sized by production give the district context.
Spots are the ferric arm's best-case ground inside the pegs, coordinates in the table. None is a gold indication; all are walk-first-with-low-expectations points.

Data and methods

Sentinel-2 L2A (10 m) for the ferric ratio; Copernicus GLO-30 (30 m) for the structure and terrain analyses; AlphaEarth embeddings (v1 annual, ~80 m) for the regional model. Known-mine references are the Blenkinsop Zimbabwe gold-deposit compilation. Code and reproducible outputs are in research/mineral-prospecting/ (arms/, sites/, model/).

The single biggest lever from here is one real ground observation from the chain: a confirmed sample or a strike/dip reading re-weights every layer against real data, instead of the model reasoning from surface signature alone.