AI System Boosts Robust Insect Authentication for Food Supply

Researchers have developed a novel AI framework called the Batch-Invariant Spectral Network (BISN) to enhance the reliability of insect species authentication. This system is particularly relevant for the burgeoning edible insect industry, where accurate identification is crucial for safety, quality control, and regulatory compliance. The core challenge BISN addresses is the performance drop of near-infrared spectroscopy (NIRS) when applied to new production batches, which often exhibit spectral variations unseen during initial training. BISN tackles this by integrating a learnable preprocessing module, initialized with Savitzky-Golay filtering, and an entropy-regularized adversarial objective. This unique combination allows BISN to suppress batch-specific spectral noise *before* species-specific features are extracted, a key differentiator from other domain adaptation methods. Evaluated on a dataset of 2,700 spectra from three common edible insect species across independent production batches, BISN achieved a mean leave-one-batch-out accuracy of 0.93, outperforming existing baselines. Furthermore, explainable AI techniques confirmed that BISN's decisions consistently relied on biochemically relevant lipid and protein absorption regions, ensuring both robustness and interpretability under realistic industrial conditions. The source code and dataset are publicly available.