How Blurry Labels Quietly Distort AI Maps of Hidden Mineral Wealth

Artificial intelligence has become an indispensable tool in the hunt for the critical minerals that power batteries, wind turbines, and electric vehicles. Across Canada and beyond, geoscientists now routinely deploy machine learning models over vast territories, asking algorithms to flag the most promising ground for deposits of graphite, nickel, copper, and other strategic commodities. But a new study published in Natural Resources Research by Steven E. Zhang and Mohammad Parsa of the Geological Survey of Canada reveals a hidden flaw lurking inside these continental-scale predictions, one that could change how exploration companies, governments, and researchers interpret every AI-generated mineral map they have ever seen.

The problem the researchers identified is called smoothing-induced weak annotation, or SiWA. In data-driven mineral prospectivity mapping, positive training samples are labeled locations where mineralization is known to occur. Ideally, the area or volume of each labeled sample would exactly match the footprint of the deposit it contains. In practice, at national and continental scales, this is nearly impossible. A single grid cell in a pan-Canadian study might cover more than five square kilometers, while the deposit inside it could occupy a tiny fraction of that space. The result is a positive sample that is overwhelmingly composed of barren ground, diluting the true geological signature of mineralization with noise from its surroundings.

Zhang and Parsa describe this with a simple analogy: a shoebox labeled as containing a tennis ball, when it actually holds a tennis ball buried among many other objects. The label is not wrong, but it is imprecise, and any learning system trained on it must guess which part of the box matters. In geostatistical terms, the annotated sample has a much larger support than the target it is meant to represent. When covariates are averaged over that larger area, the distinctive nonlinear fingerprints of mineralization are attenuated into the background, a process mathematically analogous to low-pass filtering a signal.

The theoretical consequences are striking. As the ratio of annotated area to deposit footprint grows, the averaging of many mostly negative subsamples invokes the central limit theorem, pushing the joint distribution of predictors and labels toward a multivariate normal distribution. The conditional expectation of the predictand then becomes a linear function of the predictors. In other words, smoothing artificially linearizes the relationship the machine learning model is trying to learn. The hypothesis the algorithm fits is no longer the true relationship between mineral-system evidence and mineralization, but a smoothed, linearized approximation of it. This mirrors a well-known effect in time-series forecasting, where averaging over windows longer than the autocorrelation timescale erases high-frequency, nonlinear patterns and erodes the advantage of nonlinear models.

To test these ideas empirically, the authors turned to two pan-Canadian datacubes, one targeting flake graphite and the other magmatic nickel, with or without copper, cobalt, and platinum-group elements. Both mineral systems host deposits believed to be small relative to the resolution of the data, making them naturally vulnerable to SiWA. The datasets contained 1,239 and 1,240 positive samples respectively, drawn from federal, provincial, and territorial geological surveys, ranging from minor occurrences to operating mines. The covariates included geological, geophysical, and geochronological layers such as crustal thickness, gravity and magnetic data, density models, and depth to the Mohorovicic discontinuity, all integrated on the H3 discrete global grid system at resolution 7, where each hexagonal cell averages 5.16 square kilometers.

The experimental design was deliberately rigorous. Rather than hand-picking a single algorithm, the researchers employed a deep ensemble workflow comprising 2,000 unique model configurations per target, spanning eight machine learning algorithms from logistic regression and naive Bayes to random forests, support vector machines, Gaussian processes, and neural networks, combined with 25 feature-space variants, five negative-labeling schemes, and two hyperparameter-tuning metrics. Smoothing was then applied exclusively to positive samples by expanding their labels into progressively larger neighborhoods of hexagonal cells, from the origin cell out to k-rings spanning an average centroid distance of 21.4 kilometers, roughly two orders of magnitude larger than the likely footprint of the deposits. Negative samples and covariates were left untouched, isolating annotation weakness as the sole experimental variable.

The results were counterintuitive and, in some ways, alarming. As annotation resolution decreased, apparent model performance actually improved, with both the area under the receiver operating characteristic curve and the weighted F1-score rising in an asymptotic fashion toward an elbow near k equals 3. At the same time, the prospective area shrank dramatically, with high-probability zones vanishing first, particularly in regions far from known positive samples. The sensitivity of the mapping outcome to annotation resolution proved to be on the same order as the sensitivity to the choice of machine learning algorithm itself, and an order of magnitude greater than the sensitivity to negative samples, tuning metrics, or feature dimensionality. In short, an aspect of study design often treated as innocuous turned out to rival the most heavily engineered component of the entire workflow.

The mechanism behind these trends became clear when the authors stratified their ensembles by performance. The lowest-performing quartile of workflows was dominated by simple, almost exclusively linear models such as logistic regression and naive Bayes configured with few features, and these models were the least affected by smoothing, retaining target areas most similar to the unsmoothed baseline. The highest-performing, most expressive nonlinear models, by contrast, lost nearly all of their advantage once smoothing was applied, contributing essentially nothing to the merged ensembles beyond the baseline. This is the signature of linearization: when the learnable relationship has been flattened by averaging, a linear model can capture it as well as a deep network, and the extra capacity of nonlinear methods is wasted. Meanwhile, workflow-induced uncertainty, measured as the dissent among the 2,000 equiprobable models, spiked precisely where positive samples were located, indicating that models shifted from broad agreement near known mineralization to fundamental disagreement as annotation weakened.

The implications ripple outward in several directions. First, the probabilistic meaning of a prospectivity map depends directly on annotation strength. If cells are coarser than deposit footprints, the chance of actually finding a deposit within any flagged cell is the model’s posterior probability multiplied by an annotation probability that can be vanishingly small; for the Canadian datasets, the authors estimate odds as poor as one in nearly 38,000 for the smallest targets. Field validation campaigns must therefore acquire far more samples to reach statistical significance when annotation is weak. Second, performance comparisons between studies conducted at different resolutions, on different targets, or in different regions are fundamentally incomparable unless annotation strength is explicitly controlled, a caution that strikes at the heart of the benchmarking and review culture in machine learning for geoscience. Third, and perhaps most practically, model complexity should be matched to annotation strength rather than maximized, because expressive models trained on heavily smoothed data learn relationships that are artifacts of the labeling strategy rather than properties of the Earth.

Solutions are not straightforward. Approaches developed in medical image segmentation and other weak-supervision domains, such as blending strong and weak labels, refining annotations, or rejecting noisy samples, assume that strongly annotated data exist somewhere in the pipeline, an assumption that fails at continental scales where deposit footprints are unknown and covariate resolution is costly to improve. The authors argue that the most reliable remedy, at least for now, is the one already familiar from time-series forecasting: choose simpler, more linear models when annotation strength is low, and never let model expressivity exceed what the data can support. Over the longer term, improving the resolution of primary and secondary datasets and profiling the geostatistical footprints of positive samples could reduce the severity of SiWA at its source. What is certain is that annotation quality, long a quiet afterthought in mineral prospectivity mapping, must now be recognized as a first-order control on what these powerful AI maps actually mean, and on whether the critical minerals they promise can truly be found.

Subject of Research: The effect of smoothing-induced weak annotation on data-driven mineral prospectivity mapping models

Article Title: The Effects of Smoothing-Induced Weak Annotation in Mineral Prospectivity Mapping

Article References: Zhang, S. E., & Parsa, M. (2026). The Effects of Smoothing-Induced Weak Annotation in Mineral Prospectivity Mapping. Natural Resources Research. https://doi.org/10.1007/s11053-026-10760-6

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10760-6

Keywords: mineral prospectivity mapping, machine learning, weak annotation, geodata science, data quality, smoothing, critical minerals, graphite, magmatic nickel, uncertainty, geostatistics, mineral exploration