Near-infrared spectroscopy unmixing for abundance estimation of multi-source composite adulterated wheat flour

Near-infrared spectroscopy (NIRS) has been gained extensive application in the field of flour adulteration detection due to its rapid and non-destructive analytical capabilities. Nevertheless, conventional methodologies merely offer simplistic simulations of adulteration scenarios, thereby failing to accurately represent the intricate conditions of multi-type and multi-source adulterants in actual production and circulation. To overcome this limitation, the present study constructed a multi-source composite adulteration data system: based on a handheld NIR spectrometer, multiple types and brands of wheat flour and common adulterants (cassava flour, gypsum powder, talcum powder) were systematically collected, and samples were selected for modeling and heterologous testing according to the weighted similarity (WS) criterion, preserving spectral diversity while controlling experimental costs. In conjunction with the proposed lightweight multi-task deep learning (DL) model Adulterated Flour Unmixing Net Multitask (AFUNet-MT), end-to-end joint analysis of adulteration types and component abundances was achieved. Compared with traditional machine learning (ML) methods and State-of-the-art (SOTA) DL models, AFUNet-MT exhibited superior comprehensive performance in 5-fold cross-validation: the classification accuracy reached 0.9816 ± 0.0036, and the overall R² value of component abundance prediction reached 0.9564 ± 0.0059. In heterologous testing composed of edge brands, AFUNet-MT still maintained a classification accuracy of 0.9560 ± 0.0021 and an abundance estimation R² value of 0.8992 ± 0.0050, with a residual predictive deviation (RPD) value of up to 5.7695, fully demonstrating its generalization ability and stability under cross-brand and cross-category adulteration. Regarding computational performance analysis, the proposed model exhibited low memory and high sample throughput per unit time, thereby furnishing practical technical support for embedded deployment with handheld NIR devices and rapid screening in agricultural field applications. • Multi-source composite adulteration with varied wheat flours and common adulterants. • Brand screening via weighted similarity for cost-effective sample preparation. • High-precision DL regression algorithms driven by classification results. • Homologous and heterologous testing for model performance evaluation.