SkipSymNet: Skip-Connected Symmetric Network for Robust Small-Sample NIR Spectroscopy

Accurate quantification of biochemical components, such as protein content in agricultural products, via near-infrared (NIR) spectroscopy is often hindered by the scarcity of labeled samples. Under small-sample conditions, models confront three intertwined bottlenecks: aliasing of key absorption features caused by strong spectral coupling between components, prediction instability introduced by instrument noise and sample state variations, and the consequent decline in model discriminability and generalization capability. To systematically address these challenges, this paper proposes a Skip-Connected Symmetric Network (SkipSymNet), whose core innovation lies in the synergistic integration of physics-guided feature engineering and a skip-connected symmetric neural architecture. First, we introduce a feature engineering pipeline based on prior knowledge of spectral-protein correlations, extracting rich statistical and morphological features from predefined physicochemical absorption regions to explicitly enhance input representation and interpretability. Second, we design a symmetric encoder-decoder network as the core predictor. Equipped with bidirectional information flow and hierarchical skip connections, it simultaneously preserves low-level spectral details and high-level contextual features, thereby improving training stability under small-sample conditions. Experiments conducted on a public wheat protein NIR dataset demonstrate that SkipSymNet consistently outperforms classical chemometric methods and representative deep learning baselines. Compared with the best-performing baseline method OMP-BVS, SkipSymNet increased R2from 0.9438 to 0.9536 and reduced RMSE from 0.4125 to 0.3750 on an independent test set, achieving relative improvements of 1.03% in R2and 9.1% in RMSE reduction. These quantitative results validate that, by integrating domain-specific priors with a robust network structure, SkipSymNet offers a high-precision and highly robust solution for small-sample NIR spectroscopic analysis.