Feature-optimized ResNet–BiLSTM model with hyperspectral imaging for non-destructive quality prediction of salted kimchi cabbage processed by various salting methods

Short-wave infrared hyperspectral imaging (HSI) is a rapid, non-destructive approach for tracking physicochemical variations in food materials. Here, we present a novel framework integrating HSI with hybrid deep learning and spectral feature selection to evaluate the salt content and firmness of salted kimchi cabbage (SKC) prepared using three distinct salting methods (rotary agitation, vacuum, and combined rotary agitation–vacuum). To our knowledge, this is the first study to systematically compare the effects of multiple salting processes on SKC quality using a data-driven approach. The proposed residual neural network-bidirectional long short-term memory (ResNet–BiLSTM) architecture effectively captured spatial–spectral and sequential correlations, while uninformative variable elimination (UVE) and the successive projection algorithm (SPA) minimized data redundancy. The hybrid UVE–SPA–ResNet–BiLSTM achieved superior predictive performance (R p 2 > 0.98) with an 87% reduction in spectral variables, outperforming conventional regression models. Microstructural and elemental analyses further validated that ion redistribution and tissue deformation influenced reflectance patterns. Overall, the proposed framework provides a scalable and interpretable tool for real-time quality control in salted vegetable processing, advancing the practical application of deep learning–based HSI in food manufacturing. • Salting methods altered physicochemical traits, microstructure, and elemental contents. • ResNet-BiLSTM with UVE–SPA achieved the best performance (R p 2 > 0.98, RPD > 7.9). • The hybrid model reduced data volume by over 87% compared with full-band models. • Feature-selected hybrid deep learning outperformed regression models. • HSI enabled robust quality prediction across diverse salting treatments in SKC.