Development of Field-Deployable Detection Models for Asbestos Host Rocks Using Close-Range Hyperspectral Imaging and Diverse Lithological Training Data

Abstract

This study presents a field-applicable detection models for host rocks of naturally occurring asbestos (NOA) through hyperspectral imaging and machine learning. We developed detection models using the most comprehensive dataset to date: 104 rock samples spanning 27 distinct lithologies, including 5 NOA-associated and 22 non-NOA rock types representing prevalent lithospheric formations. Sample characterization integrated visual inspection, X-ray fluorescence, X-ray diffraction, and spectral analyses, revealing distinctive signatures in NOA-associated rocks. High MgO content emerged as a crucial indicator for asbestos formation. NOA-carbonate spectral characteristics were primarily controlled by calcite, dolomite, and tremolite asbestos, while NOA-ultramafic spectral responses were governed by chrysotile, lizardite, olivine, and pyroxene. Traditional absorption-based methods (SAM, MSRFF) demonstrated poor performance due to their reliance on endmember spectra and inability to manage complex spectral variability. Machine and deep learning approaches showed superior performance by effectively leveraging key spectral bands: MgOH and CO32- for NOA-carbonate detection and MgOH and FeOH for NOA-ultramafic identification. The 3-D CNN model achieved superior performance (93.6% accuracy, Kappa coefficient: 0.88), excelling in discriminating subtle mineralogical variations through complex spatial-spectral pattern recognition. Field validation using 2 NOA and 3 non-NOA outcrops confirmed the model's reliability (89.2%-99.6% accuracy). Given the fact that the model utilizes mineral absorptions while excluding atmospheric bands and has been field-tested, the model can be applied to NOA surveys in various contexts, including construction projects and underground development.