U-Mamba-Spectra: A novel generative and explainable framework for camel milk adulteration detection using
Ao Bao1, Jiangping Liu2, Binfeng Bian3
1College of Computer and Information Engineering, Inner Mongolia Agricultural University, Hohhot 010011, China.
None:
Camel milk is vulnerable to adulteration due to its high value and limited supply. This study proposes an interpretable framework, U-Mamba-Spectra, for qualitative and quantitative detection of camel milk adulteration using near-infrared (NIR) spectroscopy. A total of 460 samples (0-100% adulteration) were analyzed within 780-1830 nm. The proposed U-Mamba classifier, integrating U-Net and Mamba modules to capture local and global spectral dependencies, achieved superior classification performance compared with conventional machine learning and deep learning models. SHAP analysis identified chemically meaningful bands associated with water, protein, and fat. For quantitative analysis, a two-stage MCR-ALS-RF model enabled accurate and chemically interpretable concentration prediction, with validation according to ICH Q2 (R1) demonstrating recoveries of 96.8-99.5%, RSD < 5%, and detection limits of 0.025-0.060 g/100 g. To improve robustness under limited sample conditions, a hybrid generative model (CVAE-CWGAN-GP-SAM) was introduced for spectral augmentation, increasing classification accuracy from 95.21% to 98.60%.
