Deep Spectral Signature Analysis and Feature Selection for Non-Destructive Curcumin Quantification in Turmeric Rhizomes from Hyperspectral Data

Abstract

This study presents a spectral signature analysis and band selection framework for the non-destructive quantification of curcumin in fresh turmeric rhizomes using hyperspectral imaging (HSI). The analysis involves acquiring reflectance spectra across 200+ contiguous bands ranging from 388.9 nm to 1026.0 nm and identifying the most informative wavelengths associated with curcumin concentration. Mathematical feature selection techniques, including mutual information (MI), successive projections algorithm (SPA), and a hybrid MI-correlation-based ranking method, were applied to reduce spectral dimensionality.

Among these, the hybrid method demonstrated the best trade-off between band relevance and diversity. It consistently selected the chemically significant 425 nm band, as validated by HPLC, while enhancing deep learning model performance with lower RMSE. These findings make hybrid band selection a key enabler for efficient and interpretable 3D-CNN-based curcumin prediction.