Hyperspectral Mineral Exploration Reveals Rare Earth Mineralization Hidden Between Samples
Hyperspectral mineral exploration is increasingly moving beyond mineral identification toward something more ambitious: using dense spectral measurements to predict mineral composition and grade continuously across geological samples.
New research from the Weishan rare earth element (REE) deposit in China demonstrates just how significant that shift could become.
Key Takeaways: Hyperspectral Mineral Exploration and REE Detection
- Hyperspectral mineral exploration can increase measurement density along drill cores, providing spectral information between conventional geochemical sampling intervals.
- In the Weishan REE study, researchers collected 65,792 reflectance spectra at 5 cm intervals across five drill cores and combined the data with geochemical assays and deep learning.
- A hybrid 1D CNN-BiLSTM model achieved an R² of 0.89 on the independent test dataset for predicting total rare earth oxide concentrations.
- Model interpretation identified influential wavelengths from the visible through SWIR, including spectral regions associated with Nd³⁺ and Pr³⁺ electronic transitions in REE-bearing minerals.
- Most significantly, continuous spectral predictions identified five medium-grade REE anomalies in intervals previously classified as barren. Subsequent mineralogical analysis confirmed previously overlooked mineralization.
- The research demonstrates a complementary workflow in which spectroscopy increases measurement density, geochemical assays provide reference data, machine learning supports prediction, and mineralogical analysis verifies significant anomalies.
Researchers combined hyperspectral measurements of drill cores with geochemical assays and a hybrid deep learning model to predict total rare earth oxide (TRE₂O₃) concentrations. The approach achieved a test-set R² of 0.89 and, perhaps more importantly, identified five previously unrecognized medium-grade REE anomalies in intervals that had originally been classified as barren. Subsequent mineralogical analysis confirmed previously overlooked mineralization in those intervals.
The result highlights an important advantage of spectroscopy in mineral exploration. Laboratory assays remain essential for quantitative geochemical analysis, but practical constraints limit how densely drill cores can be sampled. Spectral measurements can fill some of the information space between those discrete samples.
When coupled with increasingly sophisticated analytical models, those spectra may reveal geological information that conventional sampling strategies never capture.
The Sampling Challenge in Mineral Exploration
Understanding elemental concentrations along a drill core is fundamental to delineating ore bodies and developing reliable geological models. Conventional methods including X-ray fluorescence (XRF), inductively coupled plasma mass spectrometry (ICP-MS), and inductively coupled plasma optical emission spectrometry (ICP-OES) provide precise quantitative measurements.
But precision comes at a cost.
Laboratory analysis requires sample preparation, time and expense. As the researchers note, those limitations make continuous, high-density geochemical analysis of drill cores difficult. The inevitable result is distance between sampling points and the possibility that narrow or unexpected mineralized intervals remain undetected.
Hyperspectral spectroscopy approaches the problem differently.
Instead of removing and chemically analyzing every section of a core, reflectance measurements can be acquired rapidly and non-destructively across hundreds of contiguous wavelengths. Minerals interact with electromagnetic radiation according to their composition and molecular structure, producing spectral features that can provide information about mineralogy, abundance and alteration.
For REE exploration, some of those signals are particularly useful because rare earth ions can produce distinctive absorption features.
The question becomes how to translate an enormous quantity of spectral information into reliable geological information.
Quick Answer
How can hyperspectral spectroscopy improve mineral exploration?
Hyperspectral spectroscopy can rapidly collect dense, non-destructive spectral measurements across geological samples such as drill cores. When those measurements are calibrated against geochemical data and analyzed with statistical or machine-learning models, they can support mineral identification, grade prediction and anomaly detection at a much higher spatial sampling density than conventional laboratory assays alone. The Weishan REE study demonstrates this potential by using hyperspectral data and deep learning to identify mineralized intervals that had previously been classified as barren.
Building a Denser Spectral Picture of the Drill Core
The Weishan study examined five drill cores from a carbonatite-type REE deposit in Shandong Province, China.
The cores included a range of lithologies, including granodiorite, quartz syenite, breccia and REE-rich carbonatite veins. Researchers collected hyperspectral measurements at 5 cm intervals, producing 65,792 reflectance profiles across the five cores.
That density is important.
A conventional assay provides detailed information about the material that was sampled. Dense spectral measurements provide another layer of information across the intervening core.
The researchers paired these spectra with geochemical measurements and developed a model capable of predicting TRE₂O₃ concentrations directly from the spectral data.
This was not simply a matter of feeding raw spectra into an algorithm. The workflow included smoothing, noise reduction, continuum removal, normalization and baseline correction. The researchers also addressed an important problem common to mineral exploration datasets: most samples were relatively low-grade, producing a strongly imbalanced training dataset.
An oversampling technique called SMOGN was therefore used to improve representation of higher-grade samples during model training.
The resulting workflow illustrates an increasingly important principle in applied spectroscopy: the measurement is the foundation, not the finished analysis.
Why Combine CNN and BiLSTM Models?
Hyperspectral data present an unusual modeling challenge. Individual absorption features matter, but so do relationships extending across the spectral curve.
The researchers addressed both using a hybrid architecture combining a one-dimensional convolutional neural network (1D CNN) with a bidirectional long short-term memory network (BiLSTM).
The two components perform complementary functions.
The 1D CNN identifies localized patterns and absorption features within spectral curves. The BiLSTM evaluates longer-range relationships across the continuous spectral sequence. Combining them allows the model to consider both localized spectral characteristics and broader spectral structure when predicting REE grade.
The hybrid model achieved R² values of 0.99 for training, 0.92 for validation and 0.89 for the independent test dataset, outperforming the individual 1D CNN and BiLSTM models as well as MLP and XGBoost approaches evaluated by the researchers.
High predictive accuracy is useful, of course. But another part of the study may be more consequential for practical geological applications.
The researchers wanted to know why the model was making those predictions.
Connecting Deep Learning Back to Mineral Physics
One longstanding concern surrounding complex machine-learning models is interpretability. A model that predicts mineral grade accurately but provides little indication of what drives the prediction can be difficult to evaluate geologically.
The researchers therefore applied SHapley Additive exPlanations, or SHAP, to identify the wavelengths contributing most strongly to the model’s predictions.
The analysis identified influential spectral regions associated with electronic transitions of rare earth ions including Nd³⁺ and Pr³⁺ in REE-bearing minerals such as bastnäsite-(Ce) and monazite. The study identified important bands including 580, 624, 734, 795, 1550, 1908, 2207 and 2342 nm.
This matters because it connects statistical model performance back to physical spectral behavior.
Rather than functioning entirely as a black box, the model was assigning importance to spectral information with a plausible mineralogical basis.
For hyperspectral mineral exploration, that connection between predictive performance and spectral interpretation could become increasingly important as machine-learning models grow more complex.
The Most Interesting Result Was Found in the "Barren" Core
Ultimately, a mineral exploration method is useful because of what it helps geologists find.
After validating the model, the researchers applied it continuously along the drill cores. Five medium-grade anomalies appeared in intervals previously considered barren.
Those anomalies warranted another look.
Detailed petrographic observations and mineralogical analysis subsequently identified two types of previously overlooked REE mineralization. In alkaline granite, calcite and bastnäsite-(Ce) occurred between alkali feldspar grains. Elsewhere, fine carbonatite veins within biotite-plagioclase gneiss contained REE minerals including bastnäsite-(Ce) and monazite.
That result illustrates the potential value of increasing measurement density.
The spectroscopy did not eliminate the need for conventional mineralogical or geochemical analysis. In fact, those methods were essential to training the model and verifying its discoveries.
Instead, hyperspectral measurements provided a way to interrogate much more of the drill core and identify locations deserving closer investigation.
This suggests a complementary workflow:
Geochemical analysis provides high-confidence reference measurements. Hyperspectral spectroscopy increases measurement density. Predictive modeling connects the two. Targeted mineralogical analysis verifies significant anomalies.
The value lies in the combination.
From Spectral Identification to Predictive Exploration
Hyperspectral spectroscopy has a well-established role in mineral identification and alteration mapping. Studies like this one point toward a broader future for the technology.
As spectral datasets become larger and modeling methods become more capable, spectroscopy can increasingly serve as an input to predictive geological workflows.
That creates possibilities extending beyond REE grade estimation. Dense spectral measurements could potentially help prioritize samples for laboratory analysis, characterize alteration continuously along cores, identify subtle transitions between geological units and flag anomalous intervals that warrant additional investigation.
The Weishan research also points toward the integration of multiple data sources. The authors suggest future work combining spectral information with other remote sensing data, including thermal infrared and LiDAR, as well as investigating attention mechanisms, residual networks and Transformer-based architectures.
The trajectory is clear: the analytical capabilities surrounding spectral data are becoming considerably more sophisticated.
But increasingly sophisticated models also increase the importance of the data going into them.
High-Quality Spectral Data Remains the Foundation
Machine learning can identify relationships within data, but it cannot manufacture spectral information that was never captured.
For geological applications, spectral range, resolution, signal quality, measurement geometry, calibration and consistent acquisition protocols all influence what information is available for subsequent analysis.
The Weishan study is particularly illustrative because diagnostically important wavelengths were distributed across the visible, near-infrared and shortwave-infrared regions. Its SHAP analysis identified influential wavelengths ranging from approximately 580 nm through 2342 nm.
For field spectroscopy and geological research, broad VNIR-SWIR measurements therefore provide more than a conventional spectral fingerprint. They can create information-rich datasets capable of supporting statistical analysis, machine learning and emerging AI-driven workflows.
That changes the role of spectroscopy.
The objective is no longer necessarily to look at a spectrum and identify a mineral from a handful of absorption features. Increasingly, the spectrum becomes a quantitative dataset that can be integrated with geochemistry, mineralogy, spatial information and computational models.
A New Layer of Information for Mineral Exploration
The Weishan study does not suggest that hyperspectral spectroscopy will replace conventional geochemical assays. Nor should it.
Instead, it demonstrates something potentially more useful: spectroscopy can provide a dense, non-destructive layer of information between conventional measurements.
In this case, that additional information helped a deep learning model predict REE grades continuously along drill cores and draw attention to mineralized intervals that previous exploration had overlooked.
For researchers working in mineral exploration, that is an important distinction. The future may not be a choice between spectroscopy, laboratory analysis and machine learning.
It may be the ability to combine all three.
As hyperspectral datasets become larger and analytical tools more sophisticated, high-quality spectral measurements can provide the bridge between physical samples and predictive geological models, helping researchers extract more information from every meter of core.
Go Deeper: Read the Full Research
Frequently Asked Questions
What is hyperspectral mineral exploration?
Hyperspectral mineral exploration uses reflectance measurements across many contiguous wavelengths to characterize geological materials. Minerals interact with electromagnetic radiation according to their composition and structure, producing spectral features that can be analyzed for mineral identification, alteration mapping and, when combined with appropriate reference data and modeling, quantitative prediction.
Can hyperspectral spectroscopy detect rare earth elements?
Hyperspectral spectroscopy can detect spectral responses associated with some rare earth elements and REE-bearing minerals under appropriate conditions. In the Weishan study, model interpretation identified influential wavelengths associated with electronic transitions of Nd³⁺ and Pr³⁺ in minerals including bastnäsite-(Ce) and monazite. However, spectroscopy should not be interpreted as a universal direct measurement of elemental concentration.
Can hyperspectral spectroscopy predict rare earth ore grades?
Research indicates that hyperspectral data can support REE grade prediction when combined with suitable calibration data and predictive modeling. In the Weishan study, a hybrid 1D CNN-BiLSTM model predicted total rare earth oxide concentrations with an R² of 0.89 on the independent test dataset.
Why use hyperspectral spectroscopy on drill cores?
Hyperspectral spectroscopy allows drill cores to be measured rapidly, densely and non-destructively. This can provide spectral information between conventional geochemical sampling points, potentially revealing mineralogical changes or anomalous intervals that warrant additional laboratory investigation.
Can hyperspectral spectroscopy replace geochemical assays?
No. The Weishan research demonstrates a complementary relationship rather than a replacement. Geochemical assays provided the quantitative reference measurements needed to train and evaluate the predictive model, while hyperspectral measurements provided much denser coverage of the drill cores. Mineralogical analysis was then used to verify newly identified anomalies.
How is machine learning used with hyperspectral mineral data?
Machine-learning models can analyze relationships across hundreds or thousands of spectral variables and connect those patterns with known mineralogical or geochemical properties. In the Weishan study, researchers combined a 1D convolutional neural network, which extracts localized spectral features, with a bidirectional long short-term memory network designed to capture broader sequential relationships across the spectrum.
Why is spectral range important in mineral exploration?
Different minerals and chemical constituents exhibit diagnostic behavior at different wavelengths. In the Weishan study, wavelengths identified as important to REE prediction extended from approximately 580 to 2342 nm, spanning the visible, near-infrared and shortwave-infrared regions. Broad spectral coverage can therefore capture multiple sources of mineralogical information within the same measurement.
What is the advantage of combining hyperspectral spectroscopy with AI?
The primary advantage is the ability to extract predictive information from large, information-rich spectral datasets. In mineral exploration, this can potentially transform dense spectral measurements into continuous estimates or classifications that help researchers prioritize areas for further investigation. The Weishan study provides a particularly strong example: the model identified five medium-grade REE anomalies in intervals previously considered barren, which were subsequently confirmed through mineralogical analysis.
