Rare earth deposit mapping: Kriging vs machine learning 2026 false-positive drill targets

This guide compares kriging and machine learning for 2026 rare-earth prospectivity mapping and focuses on false-positive drill targets.

It sets a three-check promotion rule for cells before they receive drill budget.

Why ML targets fail: spatial leakage

Spatially leaked validation is the main reason an REE prospectivity model can look excellent in a notebook and still send a drill rig to barren ground. Scanminers, in "Validation and uncertainty in mineral prospectivity mapping" (Jul 2026), treats validation as a design problem rather than a single accuracy score: known occurrences are not a random sample, and untested ground is not necessarily barren. For REE work, that matters twice—carbonatite systems are clustered and structurally controlled, while ion-adsorption targets depend on weathering profiles that can be extensive but subtle—so a model that learns "near known deposits" can outrank genuinely new ground.

The first failure is spatial leakage: training and test samples that sit close together often share the same mineralizing event, so random cross-validation inflates AUC while the model has learned deposit proximity rather than depositivity. The check is spatially blocked cross-validation with blocks at least twice the deposit-scale correlation length, plus holdout geological domains, not random splits. In REE stacks this is not cosmetic: a carbonatite ring complex and its satellite intrusions, or adjacent ion-adsorption profiles on the same regolith surface, can leak information across a random fold even when no sample is duplicated.

The deeper REE-specific problem is the negative label. An untested carbonatite ring complex is not barren; it is censored. Treating "no recorded occurrence" as a hard negative teaches the classifier that absence of evidence is evidence of absence, which biases it toward high-confidence false positives in under-explored terrain. The practical rule is to treat untested cells as unknown or to use presence-only or spatially explicit designs, and to require field evidence before a negative label is allowed to influence a drill decision.

A target should therefore be promoted only after the model has been checked for all six mechanisms: blocked folds, domain holdouts, no post-hoc thresholding, no deposit-derived predictors, and a clear distinction between interpolation around known systems and prediction into untested ground. That is the difference between a ranked map and a drillable target.

How kriging and ML actually differ

Ordinary kriging is an interpolation estimator: it weights neighboring assays according to a fitted variogram and returns both an estimate and a kriging variance at every block. That variance is a built-in, per-location confidence measure. Machine-learning prospectivity, as described in MDPI's October 2025 review, integrates multi-source geological information—geophysics, geochemistry, and remote sensing—into a data-driven prospectivity model. The mechanism-level contrast is that kriging produces variance-bounded interpolation, whereas a classifier produces an unbounded score unless an external uncertainty layer is deliberately added.

The kriging variance is only as credible as the variogram behind it. Fit experimental variograms on REE assay data, then model the nugget effect from short-lag pairs so that near-zero separation behavior is not mistaken for spatial continuity. The resulting variance reflects both data configuration and the fitted spatial structure, not merely distance to the nearest sample.

Random forests and gradient boosting can rank cells by learned patterns, but they do not natively return a per-block variance comparable to kriging variance. Their scores can remain high in predictor combinations or spatial settings not represented in training, because the model has no explicit mechanism to state that a location is beyond supported range. Kriging reports its own uncertainty; the classifier needs a separate, deliberately designed uncertainty check.

A kriging variance check at the proposed collar therefore asks whether the interpolated REE estimate is locally supported by nearby data. It is a support check on the estimate, not a validation of the machine-learning ranking, and it does not by itself establish that the target should receive drill budget.

The practical reading is simple: kriging's uncertainty is part of the estimator, while ML's uncertainty must be constructed after the fact. That is why the two methods answer different questions—kriging asks how well a value is supported between samples, while ML asks how strongly a pattern resembles training examples.

The evidence: what 2026 sources show

The 2025–2026 sources converge on one point: in REE prospectivity work, the false-positive rate is governed more by validation design than by whether the model is kriging, machine learning, or an ensemble. A high score is therefore screening evidence until the validation protocol is visible, and the burden of proof sits with the validation, not the algorithm. The convergence is the evidence: three independent sources point to validation design as the controlling variable.

Scanminers (Jul 2026) lists six ways a high headline score can be spurious: spatial leakage, class imbalance, predictor contamination, same-domain testing, post-hoc thresholds, and interpolation mistaken for prediction. It also requires that the validation design be published alongside the result, so a reader can see what was tested rather than only the accuracy figure.

SRK (Dec 2025) states the ranking criterion in exploration terms: to the extent a prospectivity model prioritizes novel, weak targets over stronger, recognized ones, it diverts exploration capital toward poorer targets. That makes validation design an economic control, not a statistical afterthought.

A 2025 MDPI review describes machine-learning prospectivity as data-driven decision-making built from multi-source geological information. Read with the Scanminers and SRK warnings, that description implies the same caution: integrating many layers can improve a map, but it cannot substitute for a validation design that tests generalization.

Taken together, these sources do not say that one algorithm is universally better. They say that a target should be promoted only when its validation design is explicit, spatially honest, and tied to the next exploration decision.

Kriging vs ML: options compared

The comparison that matters for 2026 REE targeting is not which algorithm scores higher in a notebook; it is which output can be audited before a collar is funded. Kriging wins where a number must defend a location, while ML wins where many heterogeneous layers must be ranked quickly. The table below states the winner for each of the five criteria named in the thesis; multi-source integration and first-pass ranking sit outside those five and are noted after it.

On uncertainty reporting, the practical difference is calibration. Per-block kriging variance is already a decision-grade quantity, whereas ML probabilities are not calibrated unless a separate spatially blocked calibration step is performed. On extrapolation honesty, kriging variance blows up beyond data and visibly flags unsupported cells; ML can emit confident probabilities into geological domains it never saw, which is the Scanminers "testing inside the same domain" failure in reverse (Scanminers, Jul 26, 2026).

Data hunger and REE nugget handling are where the comparison becomes conditional. Kriging can be estimated from sparse assays, but its variance exposes the cost of that sparsity; ML is the more data-hungry route and needs enough labeled cells to avoid leakage. For REE nugget effects, neither method wins outright: kriging smooths high-nugget variability, while ML can chase noise or leakage in the same patches. Auditability favors kriging because variogram, weights, and variance are inspectable, whereas an ML rank is harder to trace to a physical control.

Outside the five criteria, ML still wins multi-source integration and first-pass ranking (MDPI, Oct 3, 2025). That strength is real, but it is a screening strength, not a collar justification. The defensible workflow is therefore sequential: use ML to integrate layers and rank cells, then require the kriging variance check at the proposed collar before any cell becomes a drill target.

Costs: what a false positive really costs

A false positive in REE prospectivity is best priced as a collar decision, not as a score. Drilling a cell whose machine-learning probability was inflated by spatial leakage spends a full collar program—mobilization, drilling, and assay—on ground that the model had no defensible reason to prioritize. That spend is unrecoverable, and the barren hole becomes a labeled non-occurrence that can bias the next training cycle if it is fed back without the same spatial discipline.

The opposite error is cheaper in a different way. A true positive vetoed by an over-conservative kriging variance threshold is not lost; it remains in the ranked inventory and can be re-tested when drill spacing tightens or new geophysics arrives. The cell keeps its place in the queue, and the decision can be revisited with better information rather than being erased by a single conservative screen.

This asymmetry is the reason to favor vetoing when evidence is thin. A drilled false positive consumes capital and contaminates future data; a skipped true positive defers a target that is still recoverable. In collar terms, one mistake is permanent and self-reinforcing, while the other is a scheduling problem.

As the promotion rule above already specifies, a cell earns a collar only when it survives both the spatially blocked cross-validation and the variance check at the proposed collar spacing; otherwise it stays in inventory as screening evidence.

Scanminers' validation guidance makes the same distinction between model validation and exploration validation: a high score is not the same as a drillable target. Treating that distinction as a cost control keeps the budget focused on cells that have passed both the spatial and uncertainty tests.

What the evidence does not establish

The 3-of-3 rule is a screening gate with explicit limits, not a universal promotion rule. No source in the available sources gives a REE-specific false-positive rate for either kriging or machine learning, so any percentage quoted for “ML false positives in REE exploration” is not from those sources; the defensible substitute is to quantify leakage on your own data by comparing spatially blocked cross-validation with random splits and treating the gap as a local warning rather than a published rate.

In data-scarce greenfields, the Nature-published ensemble work targets prospectivity under data scarcity, but sparse data can break variogram fitting. Kriging variance at the proposed collar then becomes unreliable when there are too few sample pairs in a lag bin to fit a stable model, so the check is to inspect pair counts per lag bin before trusting the variance surface; if those counts are inadequate, the kriging variance check should not be used as a promotion gate even if the blocked cross-validation passes.

Ion-adsorption deposits are a separate limit because their weathering profiles can make one global deposit-scale correlation length unstable. In that case the fixed block size and a single variance threshold may not describe the target, so the blocked cross-validation and kriging variance check should be re-estimated locally for the profile rather than applied once across the whole cell.

Where the blocked-versus-random gap is large, the rule still wins as a stop rule: reject or retrain before spending drill budget, and do not convert the gap into an REE-specific false-positive percentage. Where pair counts and variogram stability are adequate, the rule still wins as a promotion rule only when all three checks pass.

Worked Example: Run the Numbers

Illustration only: a 2026 REE carbonatite screening exercise uses a 20 km by 20 km grid (400 km²) divided into 1 km² cells, so 400 cells. The deposit-scale correlation length is taken as 2 km for this example, making the minimum spatial block 4 km; the team uses 5 km blocks, giving 16 blocks (4 by 4). A model flags 20 cells, or 5% of 400, as candidates.

Step 1, spatially blocked cross-validation: the 20 candidates are distributed across 16 blocks; 12 lie in four held-out blocks. Of those 12, 9 match the illustrative training-era labels, so the held-out candidate precision is 9/12 = 0.75. Step 2, kriging variance: at the proposed collar, 7 of the 12 candidates have kriging variance at or below the illustrative threshold of 0.25 (units²); the other 5 are above it. Step 3, the third check from the reader rule is applied to the 7 survivors; 5 pass, leaving 5 cells.

Do not let a strong blocked rank rescue a high kriging variance, and do not let a clean variance check rescue a cell that is only being predicted because it sits inside the same mapped domain as your training set. Force the pass/fail calls in writing for each proposed collar, and only cells that clear all three earn the step from prospectivity to a drill target.

Frequently Asked Questions

Which validation failure is identified as the main source of false-positive REE drill targets?

Spatially leaked ML validation, not kriging smoothing.

What two checks must a target pass before it earns drill budget?

Spatially blocked cross-validation and a kriging variance check at the proposed collar spacing.

How many of the three promotion checks must a prospectivity cell pass to become a drill target?

It must pass 3 of 3 checks, because a partial pass is insufficient.

What minimum block size is required for blocked cross-validation?

Blocks must be at least 2× the deposit-scale correlation length, and random splits are disallowed.

Why can an REE prospectivity model look excellent in a notebook yet send a drill rig to barren ground?

Spatially leaked validation is the main reason.

How does Scanminers frame validation in mineral prospectivity mapping?

It treats validation as a design problem rather than a single accuracy score, noting known occurrences are not a random sample.

Quick answers

Also worth reading: Verify 50 m drill spacing and kriging variance in carbonatite REE deposits: Verify 50 m drill spacing · Mineral prospectivity mapping: Sustainability peer-review checks, CHF 2,400—Verify or Skip: Mineral prospectivity mapping: Sustainability peer-review · Mountain Pass REE: Kriging Uncertainty and 2026 Resource Model: Mountain Pass REE: Kriging Uncertainty