Source-linked AI summary
Minerals in the Wild: A Hyperspectral-XRF Dataset for Elemental Composition Estimation
Eleftheria Tetoula-Tsonga, George Arvanitakis, Theodoros Giannakas
TL;DR
Mineral identification needs efficient hyperspectral methods and reliable paired ground-truth data, but publicly available datasets remain scarce. The paper releases a paired HSI–XRF dataset, defines an elemental-composition task, and uses USGS-informed convex optimization with pruning; the dataset exhibits cross-modal structure and varied spectral and elemental composition. Its pipeline treats USGS compounds consistently with compounds in the XRF vectors.
Problem
Publicly available mineral HSI datasets with reliable elemental ground truth are scarce, limiting development and evaluation of HSI-based identification methods.
Method
The paper pairs HSI input with USGS support data and XRF targets, then estimates specimen composition using convex optimization with a customized pruning filter.
Results
The dataset shows nontrivial HSI–XRF clustering structure, relatively small within-specimen spectral variation, and broad elemental-composition variability.
Takeaways & Limitations
Minerals in the Wild supports studying cross-modal relationships between spectral and elemental information and evaluating HSI-to-XRF composition estimation.
Takeaways & Limitations
USGS compounds are treated exactly as the compounds appearing in the XRF-exported vectors.
Abstract
from arXiv · showhide
Rapid mineral characterization is essential for applications ranging from mineral exploration to industrial ore processing. To this end, Hyperspectral Imaging (HSI) has emerged as a promising sensing modality thanks to its fine spectral resolution, enabling mineral discrimination in both close-range and remote sensing settings. However, the scarcity of publicly available datasets with reliable ground-truth labels hinders the development and evaluation of HSI-based mineral identification methods. We release Minerals in the Wild, a multi-purpose dataset comprising 1,132 rock specimens collected across Europe. For each specimen, we provide an HSI acquisition together with an elemental characterization obtained via an XRF sensor. We define the task of elemental characterization on our dataset and propose a pruning mechanism that removes distant signatures from the USGS dictionary prior to a convex optimization approach for matching HSI pixels with USGS spectral signatures. Finally, we empirically show that our approach outperforms simpler baselines.
1. INTRODUCTION
Mineral characterization needs efficient image-based sensing and algorithms that connect hyperspectral imagery with mineralogical or elemental measurements. Minerals in the Wild addresses this need with paired HSI–XRF data from intact natural rock specimens and a USGS-informed composition-estimation case study.
- Motivation: Critical raw materials support AI, advanced manufacturing, electrification, and clean energy, while exploration and mining increasingly require efficient characterization.Manual high-precision equipment is impractical for workflows with latency requirements often measured in seconds.
- Motivation: HSI captures fine spectral features, including mineral absorption bands at wavelengths unavailable to conventional RGB cameras.The sensing modality still requires an identification algorithm to infer minerals from imagery.
- Motivation: Reliable mineral identification associates hyperspectral images with matched mineralogical or elemental characterization across modalities.The paper frames HSI and matched characterization as complementary inputs for developing identification algorithms.
- Contributions: Minerals in the Wild contains 1,132 intact natural rock specimens, pairing HSI cubes with XRF-derived elemental compositions for specimen-level prediction.The dataset differs from controlled powder-mixture benchmarks by using naturally occurring, solid specimens.
- Contributions: The paper defines an HSI-to-elemental-composition case study using the USGS library as support data and validates convex-optimization estimates against XRF ground truth.A customized regularizing filter is paired with the convex optimization algorithm.
2. A HYPERSPECTRAL-XRF DATASET
Minerals in the Wild contains 1,132 European hand-sized rock specimens with paired HSI cubes and XRF-derived elemental compositions. The dataset exhibits broad elemental diversity, relatively small within-specimen spectral variation, and nontrivial cross-modal structure.
- Dataset construction: The dataset comprises 1,132 hand-sized specimens gathered from mines and exploration areas across Europe.Each specimen is represented by HSI and XRF measurements.
- Hyperspectral modality: HSI acquisition uses a push-broom camera recording 384-pixel spatial lines and 273 spectral bands from 1000–2500 nm.Specimen scans are calibrated with dark- and white-references and manually masked to remove background pixels.
- XRF modality: XRF characterization averages five 8 mm point measurements and harmonizes mixed element-and-compound labels into a 51-dimensional element-only composition.Compound coordinates are decomposed using stoichiometric mass fractions, light elements are removed, and the result is renormalized.
- Elemental diversity: Around 30% of specimens have Si as their top element, while the distribution of non-zero elemental fractions shows substantial variability.The absence of a single dominant element supports the dataset's use as a challenging benchmark for hyperspectral unmixing.
- Spectral variability: Within-specimen spectral variation is relatively small overall, although spectra fluctuate more around the mean as variation increases.MWAD summarizes angular differences between individual spectra and each specimen's mean spectrum.
- Cross-modal structure: HSI cubes show some, but nontrivial, clustering aligned with five groups formed from XRF vectors, and higher XRF entropy is associated with increased spectral variability.The entropy–MWAD trend is positive but not definitive.
3. CASE STUDY: CHARACTERIZING SPECIMENS FROM HYPERSPECTRAL IMAGERY
The case study estimates specimen-level elemental composition from HSI by matching pixels to a filtered USGS spectral basis, aggregating mineral weights, and harmonizing them with XRF elements. The proposed method outperforms simpler baselines in accuracy and inference speed.
- 3.1. Task Formulation: The task maps each specimen HSI cube to its elemental composition, evaluated against XRF-derived ground truth using L2 and cosine-similarity metrics.L2 measures compositional error in direction and magnitude, while Cos measures agreement in direction.
- 3.2. Proposed Pipeline: USGS signatures are resampled to the camera bands and preprocessed with continuum removal before matching HSI pixels to library spectra.The original signatures have finer spectral resolution than the camera.
- 3.2. Proposed Pipeline: The pipeline estimates pixel-level compositions under the USGS spectral basis, spatially averages them, then converts mineral composition into the XRF elemental basis.The final vector discards elements outside the XRF basis and is renormalized.
- 3.2. Proposed Pipeline: The candidate filter retains signatures whose first- and second-order derivatives match each pixel, reducing the library before convex optimization.Excluded signatures receive zero weight when the reduced solution is embedded back into the full USGS basis.
- 3.3. Experimental Evaluation: The threshold controls candidate-selection strictness, creating a trade-off between retaining signatures and achieving specimen coverage.Coverage is the percentage of specimens with at least one candidate USGS signature.
- 3.3. Experimental Evaluation: 0.773 median Cos and 0.150 median L2 are achieved by the proposed method, compared with 0.577 and 0.221 for MaxCorr.The method achieves the best performance across both reported metrics and is 6× faster than vanilla SLS.
4. CONCLUSIONS
The paper contributes a paired HSI-XRF dataset of naturally occurring mineral specimens and a reproducible USGS-informed pipeline for elemental characterization. Its convex optimization method uses a customized pruning filter and is evaluated against related benchmarks.
- 4. CONCLUSIONS: The dataset pairs HSI acquisitions with XRF measurements for naturally occurring mineral specimens, supporting cross-modal study of spectral and elemental information.The collection contains 1,132 specimens gathered across Europe.
- 4. CONCLUSIONS: The proposed pipeline uses HSI as input, the USGS library as support data, and XRF as the target for elemental characterization.Its convex optimization approach includes a customized pruning filter and is compared with other USGS-informed benchmarks.