Short answer

Critical-minerals screening in GIS ranks large areas of ground by their likelihood of hosting a target commodity — lithium, rare earths, copper, cobalt, nickel, graphite and the rest of the strategic list — so exploration spend goes where the geology supports it. The method is mineral prospectivity mapping: translate a mineral-system model into mappable evidence layers, standardise them onto a common grid, combine them with a defensible weighting scheme, and validate the result against known occurrences.

The output is a continuous prospectivity surface, not a treasure map. Its value is in ranking and de-risking — telling you which 5% of a licence to drill or sample first, and explaining why, in a way a technical committee can audit.

Start from a mineral-system model, not the data

The cardinal rule: define the mineral system before you touch a layer. Every deposit type forms through a sequence of processes — a source of metal, a fluid pathway, a trap, and a preservation history. For each process you ask: what is its mappable footprint? That question turns geology into evidence layers.

Worked example for hard-rock lithium (LCT pegmatites):

  • Source/fertility → proximity to evolved, fractionated S-type granites.
  • Pathway → structural corridors; pegmatites intrude along faults and fold hinges.
  • Trap/host → competent host rocks (e.g. metasediments, amphibolites) near the granite cupola.
  • Indicators → geochemical Li, Cs, Ta, Sn anomalies; sometimes a radiometric K signature.

For sediment-hosted copper or REE-bearing carbonatites, the layers differ entirely. There is no universal stack; the model dictates the layers.

The evidence layers

  • Geological map — host and source lithologies, extracted as polygon classes (see digitising geology into a proper schema).
  • Structure — faults, shear zones, fold axes as line features; the screen usually uses distance to structure.
  • Geochemistry — stream-sediment or soil multi-element data, interpolated and thresholded into anomaly surfaces.
  • Geophysics — magnetics (reduced to pole), radiometrics, gravity; used both directly and as derivatives.
  • Remote sensing — alteration indices from ASTER or Sentinel-2 band ratios where lithology drives a spectral response.
  • Known occurrences — point data used to derive weights (in data-driven methods) and to validate (always held back, in part).

Standardise every layer to a common frame

Before combining anything, all layers must share a grid, CRS, extent, and cell size. Pick a projected CRS appropriate to the area (e.g. a UTM zone, EPSG:32733 for UTM 33S) and a cell size matched to the coarsest meaningful input — there is no point in 10 m cells if the geochemistry samples are kilometres apart. Convert vector evidence to raster proxies:

# Distance to faults, as a raster proxy for structural control
gdal_rasterize -burn 1 -tr 50 50 -a_nodata 0 faults.gpkg faults.tif
gdal_proximity.py faults.tif fault_dist.tif -distunits GEO

In QGIS the equivalents are Rasterize (vector to raster) and Proximity (raster distance). Each layer is then reclassed onto a common evidence scale.

Two ways to combine — knowledge-driven vs data-driven

Weighted overlay (knowledge-driven)

Reclassify each evidence raster to a common integer scale (say 1–5, where 5 is most favourable), assign each layer a weight reflecting its importance in the mineral model (weights sum to 1), and compute the weighted sum. In QGIS Raster Calculator or ArcGIS Weighted Overlay:

prospectivity = 0.30*host_rock + 0.25*fault_dist + 0.25*geochem + 0.20*magnetics

Simple, transparent, and easy to defend — but the class breaks and weights are subjective.

Fuzzy logic (knowledge-driven, continuous)

Instead of hard classes, map each layer to a fuzzy membership between 0 and 1 (e.g. a fuzzy-large function so high Li values approach 1 smoothly). Combine memberships with operators: fuzzy OR (max — any strong evidence raises the score), fuzzy AND (min — demands all evidence), fuzzy SUM/PRODUCT, or the fuzzy gamma operator, which balances the increasing tendency of SUM against the decreasing tendency of PRODUCT. Fuzzy logic avoids the artificial edges of reclassification and models partial evidence well.

Weights of evidence (data-driven)

If you have enough known occurrences, weights of evidence computes positive and negative weights (W+, W−) for each evidence class from the spatial association between that class and the occurrences, producing a posterior probability. It is objective but assumes conditional independence between layers — an assumption that magnetics and lithology often violate, inflating the result if unchecked.

Worked screen, end to end

  1. Model: LCT pegmatite lithium. Layers: granite proximity, fault distance, Li-Cs-Ta geochem anomaly, competent host rock, radiometric K.
  2. Standardise to a 100 m grid, EPSG:32733, clipped to the licence.
  3. Fuzzy-membership each layer (e.g. fuzzy-small on distance-to-granite so near approaches 1).
  4. Combine with fuzzy gamma (γ ≈ 0.85) so multiple weak signals still register.
  5. Threshold the top decile as priority targets.
  6. Validate against the known prospects held out in step 1.

Validation — the step people skip

A prospectivity map you cannot test is an opinion. Hold back a portion of known occurrences and measure how well the high-prospectivity zones capture them. A useful summary is the prediction-rate / success-rate curve: plot cumulative occurrences captured against cumulative area classified prospective. A model that captures 80% of known deposits in 20% of the area is performing; one that needs 70% of the area to find them is no better than chance. The area under that curve is a single, comparable score for tuning weights.

Common pitfalls and why they happen

  • Stacking layers without a mineral model. You get a generic "interesting ground" map that captures topography and data density, not ore controls — because the layers were never tied to ore-forming processes.
  • Mismatched cell sizes and extents. Map algebra silently aligns on the wrong grid and produces shifted results; always resample to one master grid first.
  • Conditional-independence violation in weights of evidence. Correlated layers (magnetics and mafic lithology) double-count the same signal and overstate prospectivity.
  • Using all known occurrences for both weighting and validation. The model then "validates" against the data that trained it, which proves nothing. Split the points.
  • Sampling-density bias in geochemistry. Densely sampled areas look anomalous simply because they were sampled more; normalise or note it.

QA and validation summary

Confirm every layer shares CRS, extent, and cell size; document the mineral-system rationale for each layer and weight; keep an independent validation set of occurrences; report a capture-vs-area curve; and flag where data coverage is too sparse for the screen to be meaningful, so the map's blank confidence does not mislead a non-specialist.

Bathyl perspective

We build prospectivity screens as auditable systems: the mineral-system logic, the standardisation grid, the weights, and the validation curve all travel with the output. The map is a ranking tool that focuses fieldwork, and it states plainly where the data is too thin to trust — because an honest "we don't know here" protects an exploration budget better than a false high.

Related reading

Sources