Portable Near Infrared Spectroscopy Study for Turmeric Adulteration Detection

Turmeric powder is commonly adulterated with low-cost plant-based materials for economic gain; however, rapid on-site detection methods remain limited.

This study is based on the development of a rapid, non-destructive method that uses portable near infrared spectroscopy, chemometrics, and deep learning, enabling the detection of corn flour adulteration in turmeric powder.

A total of 330 adulterated samples (0-50% corn flour) were prepared, and corresponding NIR spectra were collected across the range of 900-1700 nm. A systematic comparison was carried out using eight pre-processing methods, three feature selection strategies (PCA, CARS, UVE), and multiple modeling algorithms. A dual-attention CNN-LSTM multi-task network was developed to simultaneously perform adulteration classification and content regression.

The optimization of the PLS model was performed through second derivative (D2) pre-processing and UVE feature selection (78.5% compression). The CNN-LSTM dual-attention model achieved a classification accuracy of 98.9% and a remarkable regression performance (R2 = 0.9927, RMSE = 1.3825, and RPD = 11.64), significantly outperforming traditional PLS models. Weight visualization revealed that the learned attention maps corresponded closely to the characteristic spectral features associated with the curcumin attenuation and starch enhancement.

The integration of portable NIR spectroscopy with a multi-task deep learning model offers a robust, efficient, and accurate strategy for on-site rapid screening of turmeric adulteration under controlled experimental conditions.