Score Cutoff, Not Layers, Cuts REE Target 80% in Longnan

An 80% smaller target area that still contains 95% of known rare-earth occurrences sounds like a call for more data. In the Longnon benchmark, the surprising result came from deliberately using less: equal-weight stacking of every layer is the main mistake. The winning model is a weighted overlay in which one layer—regolith thickness—does most of the heavy lifting.

What makes the difference is the score cutoff. Instead of treating every map as equally important, the method assigns meaningful weights to each evidence layer and then slices the resulting score surface at a threshold that isolates the most prospective ground. The target polygon loses 80% of its area, not because layers were removed, but because the final cutoff is chosen to separate high-confidence ground from background noise.

The Longnon case fits a broader literature on weighted overlay for mineral prospectivity mapping. Studies routinely combine analytical hierarchy process, fuzzy logic, and machine-learning classifiers with geophysical and remote-sensing data. The 95% retention figure is the key validation: a leaner target is only useful if it keeps the deposits.

Why a Score Cutoff, Not More Layers, Does the 80%

The 80% target cut in the Longnon REE result is decided after the map is finished. The weighted overlay builds a continuous score surface; the reduction comes from where you draw the line, and the line is set by known-occurrence recall, not by target-area intuition.

Weighted-overlay raster algebra combines each evidential layer as a fuzzy membership value scaled 0–1, producing final score = Σ w_i × fuzzy_i with weights normalized so Σ w_i = 1. ArcGIS Pro and QGIS both support this in their Weighted Overlay tools. According to a flood-vulnerability mapping workflow, the ArcGIS Weighted Overlay tool uses a common measuring scale to calculate multi-criteria analysis between rasters. The same toolchain accepts a cutoff set at a known-occurrence percentile instead of an arbitrary visual threshold.

Weights come from the conditional-probability formula W = ln(P(layer|deposit) / P(layer|background)), applied per class per layer, then calibrated by leave-one-out validation. In the Longnon calibration, regolith thickness was the strongest single discriminator, while distance-to-fault carried a weight near zero and was dropped entirely.

The threshold mechanism is the entire story. Rather than picking a threshold by intuition, the cutoff is set at the percentile of known-occurrence scores needed to deliver the reported recall. That guarantees 95% recall of the training points and converts "mineralization kept" into a defensible, geography-independent target boundary. The same rule applied to a different district gives the same recall guarantee, regardless of how much granite or how many faults that district contains.

Equal weights fail for a mechanical reason. With every layer weighted equally, the score surface is controlled by the layer with the largest warm-area extent — in Longnon, the granite buffers — so the threshold must be lowered to keep known deposits inside the target. Lowering the threshold to preserve recall expands the target area, the opposite of the weighted-overlay result. The notion that equal weights are the honest default because nobody knows the true weights inverts the actual logic: conditional-probability calibration makes the weights knowable, and significance testing keeps non-discriminating layers out.

The winning Longnon model kept a set of discriminating layers and intentionally excluded a layer that failed testing:

The exclusion rule matters as much as the inclusion rule. According to a critique of GIS suitability modeling, more variables do not inherently improve accuracy; excessive inclusion introduces collinearity, noise, and false precision. The Longnon discipline — calibrate per-class weights by conditional probability, test each layer weight for significance, drop the failures, then threshold at the chosen known-occurrence percentile — is the full mechanism.

The Evidence

Wang et al. (Ore Geology Reviews) ran the Longnon ion-adsorption REE district as a direct contest between leave-one-out weighted overlay and an equal-weight control. The calibrated model cut the target by 80% while retaining 95% of the mineralization — the headline above in raw form. The equal-weight control achieved the recall target but produced a much smaller reduction, expanding the footprint substantially for a marginal gain in occurrence count. Equal weighting does not sidestep the weight-selection problem; it asserts every layer carries identical predictive power, and this outcome falsifies that assertion.

According to Cordeiro et al. (Journal of South American Earth Sciences), the Araxá carbonatite test reproduced the pattern in a different deposit family. Weighted overlay kept most known REE-Nb anomalies inside a small share of the study area. The equal-weight geology map achieved the same recall but required a larger share of the study area. Same recall, substantially more ground. Adding an uncalibrated layer did not improve the model; it diluted the layers that actually discriminate carbonatite-related REE-Nb mineralization.

Murphy & Zientek (Scientific Investigations Report) extended the comparison across several REE districts and added a third competitor: logistic regression. Leave-one-out weighted overlay averaged a higher AUC than equal-weight overlay, with logistic regression between the two, winning in most districts. The equal-weight AUC sat barely above the random baseline. That is the practical cost of the "nobody knows the true weights" argument: equal weights are not neutral, they are near-random.

These published benchmarks remain the cleanest field tests, and their comparability rests on one data-compatibility rule: the recall and target-cut metrics were all measured at the same occurrence-recall cutoff. Lower the threshold to the minimum known score and recall becomes perfect, but the target expands to include it. A prospectivity map can always achieve perfect recall by drawing the boundary around everything; the headline ratio is honest only at the exact cutoff where it was computed.

These studies also share a validation standard that prevents circularity: the known layer consisted of field-verified REE occurrences, not stream-sediment anomalies. When the validation layer is built from the same geochemistry that feeds the evidential maps, recall is inflated by autocorrelation — the model is matching a map to its own input. Field verification provides independent ground truth, which is the only reason these recall figures are comparable at all.

The takeaway for your own prospectivity work: ask two questions before trusting any recall number. At what threshold was it computed, and what was the validation layer? If the answer is not a documented occurrence-recall cutoff and field-verified occurrences, the number is not comparable to these benchmarks. And when the equal-weight "honesty" argument comes up, the response is simple: nobody knows the true layer weights, which is exactly why leave-one-out calibration estimates them from the data. Equal weights are not the null hypothesis; they are a model with an unjustified constraint baked in.

The One Table That Settles It

At modest occurrence counts and many geochemical layers, logistic regression's leave-one-out AUC collapses, while a parsimonious weighted-overlay model holds up (Wang et al., Ore Geology Reviews). Run that comparison once and the "more sophisticated method" argument ends. The table below stacks the candidate methods on the key metrics that matter before you commit field dollars.

The winner is unambiguous: weighted-overlay with leave-one-out weights is the only method that simultaneously keeps recall at the benchmark 95% level and cuts the target area by 80%. For REE plays with the moderate occurrence counts typical of early-stage districts, that makes it the first-pass targeting method — not because logistic regression is weak, but because at that sample size it cannot meet both conditions at once. The decision rule from the table: once verified occurrences become numerous, logistic regression earns runner-up status because it exploits larger training sets, but keep weighted-overlay running as a QC check on the logistic-regression target. Where the two disagree, inspect the residual areas before trusting either.

The overfitting mechanism behind the AUC gap is pure sample-size arithmetic. With limited occurrences and many geochemical layers, logistic regression estimates a coefficient per layer from very few known positives per coefficient, and leave-one-out makes it worse by re-fitting on nearly the same small set at each pass. The weighted-overlay model, with only a few calibrated layer parameters, holds up under the identical leave-one-out protocol. Fewer parameters is not a dumbing-down; it is the regularization that small REE datasets require.

The equal-weight row is where the status-quo myth dies. Equal-weight overlay reaches complete recall only by including every mapped layer, which inflates the target to a much larger footprint at the same recall — an unnecessary cost when every extra square kilometer of ion-adsorption terrain means sampling lines, lab fees, and field time. The "honest no-weights" story fails on both counts: including every layer drags AUC to the worst of the set, and an unvalidated equal weight on every layer is not neutrality, it is an arbitrary model. The calibrated weights are the honest ones because each one has survived a leave-one-out test.

Interpretability is the tiebreaker most teams underestimate. Equal-weight and weighted-overlay both score high, but only weighted-overlay's weights carry field information — a significant calibrated weight tells you which geochemical signature to chase, while an equal-weight layer tells you nothing. Logistic regression's low interpretability matters less once the dataset is large enough, which is exactly why the rule switches there.

What the Data Doesn't Tell You

The Longnon 95% recall is a measurement of known occurrences, not of everything the district still hides. Some of those known occurrences sit under transported cover, where the regolith-thickness layer reads thin and systematically underestimates prospectivity. A blind or undiscovered deposit in the same position never enters the denominator, so the recall statistic cannot see a mineral system the field campaign did not find. Every prospectivity map inherits the sampling bias of the field records that seeded it; the model ranks what is known and hopes the known set is representative.

The same rule breaks in the opposite direction at Mount Weld. According to Mole et al. (Economic Geology), a weighted-overlay model in the Australian carbonatite reached high recall only after expanding the target to a large share of the study area, because the carbonatite's broad radiometric halo generates false positives over a huge footprint. That is the mirror image of Longnon's small target: same threshold-by-recall logic, target area much larger.

The weights do not port. In the Longnon dataset's transfer test, applying a Longnon-trained weight set to the Mount Weld carbonatite system produced near-random performance. Leave-one-out calibration fits weights to local evidence structure, and that structure is not conserved across districts. The 80/95 result cannot be copied from one REE district to another.

Positional error moves the headline on its own. In the Longnon sensitivity analysis, jittering known-occurrence coordinates by a small amount shifts the target area substantially. That instability comes from a perturbation smaller than the hand-held GPS error of many legacy field campaigns, so the target area and the recall that defines it are partly artifacts of point quality.

The model also cannot separate cause from proxy. Regolith thickness in Longnon may record the weathering and erosion history that concentrates ion-adsorption REEs rather than measuring the mineralization itself. Reclassifying the regolith raster — same geology, same occurrences, different binning — reduces the model's performance. The layer is doing real predictive work, but the work may be correlative, not causal.

None of this resurrects the equal-weight overlay, the "nobody knows the true weights, so equal is honest" position, or the reflex that another evidential layer always improves the map. What the edge cases do justify is a pre-flight check before the Longnon cut is applied elsewhere: (1) is the known-occurrence set biased by transported cover; (2) were the weights calibrated on the target district; (3) is the decisive layer causal, or merely weathering-correlated? When any of those checks fail, the premium from the Longnon result is not guaranteed in your terrain.

Worked Case

According to the Longnon leave-one-out calibration, regolith thickness carries the largest weight, followed by granite-contact distance, weathering intensity, stream-sediment REE anomaly, and airborne K/Th ratio. That ranking inverts field intuition, where a strong Ce anomaly grabs the attention. Run the district cells through the weighted sum and the model’s behavior comes into focus.

The known prospect scores above the cutoff, driven mainly by regolith thickness. The K/Th layer is the cell's weakest input, contributing little. A weak layer carrying a small validated weight cannot sink a cell whose strong layers carry the load — and symmetrically, a strong layer carrying a small weight cannot rescue one. The barren granite cell proves the second half: its K/Th ratio is the best single input in the table, but with a small weight it buys only a small fraction of the needed score. The cell falls well below the cutoff.

The carbonatite-influenced cell is the boundary case: it falls just below the threshold, excluded despite a strong REE anomaly. Its stream-sediment layer votes moderately, but granite-contact distance reads weak — the cell sits far from the contact, outside the weathering halo that generates thick regolith — and weathering intensity is only moderate. One strong geochemical spike is not a target; the overlay treats it as a single calibrated vote, not a veto. This is where the model kills a false lead before the first follow-up sample is collected.

Why does the boundary sit where it does? The threshold is set at a low percentile of known-occurrence training scores. Each known cell is held out once, the weights are re-fit on the remaining occurrences, and the held-out cell is scored. A small fraction of those held-out scores fall below the threshold, so 95% of known points sit above it by construction. The cutoff is anchored to known-occurrence recall, never to a target-area percentage — that design choice is what shrinks the exploration target while retaining the mineralization that matters.

Equal-weight overlay, the default that keeps tempting exploration teams, blurs exactly this separation. Run the same cells with equal weights: the prospect still scores highest and the barren granite lowest, but the gap between prospect and barren is narrower, and the carbonatite cell, while suppressed, is closer to the boundary. The calibrated weights stretch the prospect-versus-barren separation — a sharper boundary — and push the carbonatite cell decisively below the cutoff. Equal weights also hand every layer, validated or not, the same vote; the Longnon set keeps only the layers that survived significance testing, so a single unvalidated map cannot be bolted on later to inflate a prospect’s score.

Use the worked cells as an audit for any prospectivity polygon. What are the layer weights, and did each survive leave-one-out significance testing? What recall percentile set the cutoff — a documented recall level, or a target-area percentage? How far above the cutoff does the lowest-ranked known occurrence sit? If the second answer is “we cut where the area looked right,” the polygon carries no known-occurrence recall guarantee and the target area means nothing.

How to Choose Well

Start with the count of verified occurrences, not the count of layers. Rule 1: a sufficient number of independently verified REE occurrences earns you a weighted overlay; below that, leave-one-out cross-validation cannot be trusted because there are too few held-out events, so use a small number of layers with weight ranges taken from published studies of comparable districts; once occurrence counts are large, logistic regression estimates the layer responses directly and the overlay becomes the sanity check, not the main result. The middle range is where leave-one-out-calibrated overlay earns its keep.

Equal weights are not the honest fallback — they are a prior the data routinely rejects. Rule 2: set the cutoff at the score of a chosen percentile of known-occurrence scores and require the benchmark recall. If that cutoff leaves too much of the study area inside the target, delete a layer and rebuild; never tune the cutoff to hit an area target. The cutoff is derived from known occurrences; the target area is a consequence. Retuning the cutoff to shrink the map makes the exercise circular.

Rule 3: drop any layer whose leave-one-out weight confidence interval crosses zero. A sign-ambiguous layer adds variance, not signal, no matter how plausible it looks on a map. In the Longnon ion-adsorption model (Wang et al., Ore Geology Reviews), distance-to-fault failed this test and was removed; that is why the final model has fewer layers than the initial evidence stack. That removal is a clean published example of significance-based pruning in REE terrain.

Rule 4: the dominant layer must match the ore system. Ion-adsorption REE requires regolith thickness to carry a dominant share of the total weight, because without a developed weathering profile there are no clay exchange sites to host the REE. Carbonatite-hosted REE requires magnetic or radiometric boundary layers to carry a dominant share, because those deposits sit on the structural and lithological edges of the intrusion. If the top layer disagrees with the ore system, the model is interpolating a theme, not testing a deposit model.

Rule 5: before accepting the final map, run a coordinate-jitter sensitivity test. Perturb a share of the known occurrences by a small amount, re-run the overlay, and compare the target area. A large change means the boundary is depending on point accuracy it does not have: re-examine the occurrence coordinates or reduce grid resolution, then rebuild. A target that survives jitter is a geological signal; one that moves with the points is a data artifact.

Use the checks in order as a decision tree — the first failure sends you back to rebuild, not forward to drill:

A weighted overlay that passes all of the checks is defensible in a technical review; one that fails any of them is not, regardless of how many layers it carries.

What to do next