A Comprehensive Framework for Acquisition and Preprocessing of Hyperspectral Images of Fresh Turmeric Rhizomes for Curcumin Prediction

Abstract

Hyperspectral imaging (HSI) integrated with deep learning has proven to be a promising non-destructive approach for quantifying biochemical compounds in agricultural products. This study presents a comprehensive pipeline for predicting curcumin concentration in fresh turmeric rhizomes using HSI and a convolutional neural network (CNN). We describe the complete process—from image acquisition and spectral calibration to reflectance correction, noise reduction, and spectral normalization. Comparative evaluation of preprocessing techniques, including Savitzky-Golay smoothing and wavelet denoising, shows that denoising slightly improves prediction accuracy for low-concentration samples, while raw data and the proposed preprocessing method achieve the best overall performance. The pipeline maintains a consistent signal-to-noise ratio (SNR) of 24.8 dB and achieves a mean absolute error (MAE) of 111.8 ppm across the dataset. On a smaller set of representative test samples, MAE values as low as 0.13 ppm were observed.