SSTLNetwork: a self-supervised spectral reconstruction network with hybrid attention for near-infrared spectral transfer learning

Near-infrared (NIR) spectroscopy is widely used for non-destructive chemical analysis, yet models trained on one instrument or process condition often degrade when applied to data from different sources—a challenge known as domain shift. Transfer learning methods that rely on labelled target-domain data are not always practical in industrial or clinical settings. This work proposes SSTLNetwork, a self-supervised transfer learning approach for one-dimensional spectral data. The model employs a masked reconstruction pre-training objective with 75% patch masking and investigates a hybrid attention mechanism that pairs External Attention with multi-head self-attention. Pre-trained on a single melamine resin NIR dataset (3032 samples), SSTLNetwork is evaluated on held-out melamine recipe datasets, a pharmaceutical tablet dataset from a different spectrometer, and a virus classification task. After brief unsupervised fine-tuning (10 epochs), reconstruction on the tablet dataset achieves a mean absolute percentage error (MAPE) of 3.19 ± 1.04% and tolerance-based reconstruction accuracy (TA@0.01) of 86.55 ± 7.86%. Multi-seed ablation studies indicate that all architectural variants produce comparable reconstruction quality, suggesting that the masked reconstruction framework itself, rather than the specific attention configuration, is the primary driver of transfer learning performance. A preliminary downstream classification experiment on viral detection is also reported, though these results require further validation. These results, validated across five independent training runs, suggest that self-supervised masked reconstruction can yield transferable spectral representations, though further validation with additional datasets is warranted. • A self-supervised masked reconstruction framework adapted for one-dimensional NIR spectra. • A hybrid attention block combining External Attention and multi-head self-attention is investigated for capturing global and local spectral features. • Unsupervised fine-tuning on unlabelled target data enables cross-domain spectral reconstruction without reference values. • Multi-seed ablation studies reveal that the masked reconstruction framework drives transfer performance, with individual attention variants performing comparably. • A preliminary downstream classification experiment provides initial, suggestive evidence regarding representation transferability.