Answers are grounded in this paper and cite their sources.
Smectite-bearing rocks may slake, swell, and weaken, posing risks in tunnel engineering. This study examined whether a one-dimensional convolutional neural network (CNN) could classify powdered rock specimens into X-ray diffraction (XRD)-estimated smectite-content classes (< 10 or ≥ 10 wt.%) using visible-to-near-infrared (VNIR; 450–950 nm) hyperspectral data. Specimens comprised hydrothermally altered andesite and tuff-siltstone mainly affected by diagenesis, and a separate model was developed for each domain. Three levels of evaluation are distinguished: (i) spectral-level five-fold cross-validation, with folds split by specimen and repeated with ten random seeds; (ii) within-domain tests; and (iii) cross-domain tests. Mean cross-validation accuracy and Macro-F1 (mean ± SD) were 74.45 ± 5.32% and 74.09 ± 5.75% for andesite, and 72.28 ± 2.93% and 71.44 ± 3.25% for tuff-siltstone. Within-domain test accuracy was 100% ( n = 8) and 88.89% ( n = 9), respectively, but fell to 53.33% and 57.58% in cross-domain tests. Occlusion analysis identified different wavelength regions; because lithology and geological history were confounded, these regions do not demonstrate direct detection of smectite. A portable darkroom enabled spectral acquisition at a construction site under controlled lighting, but field classification has not been validated. Validation across additional geological domains, specimen conditions, and moisture levels could establish whether this approach can support rapid screening, prioritize confirmatory tests and geological investigations, and improve timely characterization of smectite-bearing ground in tunnel construction.
No takes yet. Share an insight, caveat, or question.
Suzuki et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: