Hyperspectral imaging with machine learning for non-destructive characterization of compacted bentonite

Compacted bentonite blocks are essential engineered barriers in high-level radioactive waste disposal. However, non-destructive characterization of their spatial heterogeneities remains a challenge. This study develops a hyperspectral imaging framework integrated with a multi-task autoencoder-inspired model (AE) that learns a compact spectral representation to predict water content and dry density. Performance is evaluated through specimen-level leave-one-out cross-validation (LOO-CV) and block-scale application. Under LOO-CV, the proposed AE achieved near-perfect bentonite type classification, predicted water content reliably, but predicted dry density less accurately due to weaker spectral sensitivity. Baseline models, which rely primarily on direct spectral relationships, performed comparably at this scale. At the block scale, however, baseline models exhibited large prediction errors and spatial instability due to surface heterogeneity, whereas the proposed AE maintained consistent spatial distributions of both properties. These results indicate that the proposed framework offers a robust and practically viable approach under realistic field-like conditions for non-destructive monitoring of engineered barriers in repository conditions.