Identification and quantification of chicken adulteration in lamb using multi-task deep learning based on near-infrared spectroscopy.

To address lamb adulteration using portable near-infrared (NIR) spectroscopy with small samples, this study proposed a unified deep learning framework (Dual-path Multi-scale Task-Former, DMT-Former) for simultaneous classification and regression of chicken adulteration. Systematic experiments, including ablation analysis, performance comparison and interpretability evaluation, were conducted on one-dimensional (1D) spectral sequences. The model achieved robust performance on the prediction set, with binary classification accuracy of 98.12% and four-class classification accuracy of 90.08%, and regression performance of R2 = 0.9362. Comparative results indicated that performance in small-sample NIR analysis was primarily determined by architecture and data compatibility rather than model complexity, featuring structures explicitly designed for 1D spectral sequences, thereby offering superior robustness. Integrated Gradients (IG) analysis further identified chemically meaningful wavelength regions (1375-1550 and 2000-2134 nm) which dominated model decisions, supporting the interpretability and reliability of DMT-Former. This work offers an interpretable multi-task deep learning solution for small-sample lamb adulteration analysis.