Non-destructive concrete strength detection using a hyperspectral vision-based modeling framework.

Abstract

Nondestructive testing offers non-invasive, real-time, and efficient advantages over traditional damage detection. Hyperspectral imaging is widely used for damage identification and material degradation assessment through non-destructive detection, yet research on mechanical property inversion is limited. This study applies hyperspectral data (400-1000 nm) to concrete compressive strength evaluation. By calibrating spectral features with destructive test results, a machine learning model is developed to quantitatively correlate spectra with strength, extending the technology's role in structural performance assessment. Experimental results indicate that incorporating Kubelka-Munk theory into spectral preprocessing substantially reduces prediction error. The root mean square errors of the backpropagation, genetic algorithm-optimized backpropagation, support vector machine, and random forest models decreased by 9.18%, 4.33%, 2.26%, and 5.64%, respectively. Compared with the RMSE of 4.71MPa reported in the literature, the optimal model in this study achieved an RMSE of 3.54MPa, representing a relative improvement of. The coefficient of determination (R2) reached 0.91, corresponding to a relative increase of 0.02. This study validates the effectiveness of hyperspectral vision technology as a reliable approach for non-destructive concrete strength assessment.