Deep learning-enabled hyperspectral imaging for high-accuracy non-destructive quantification of nutritional components in multi-variety apples

Conventional methods for quantifying soluble solids content (SSC), vitamin C (VC), and soluble protein (SP) levels in apples are destructive and unsuitable for large-scale postharvest quality monitoring. This study aimed to develop a convolutional neural network-bidirectional gated recurrent unit-attention (CNN-BiGRU-Attention) model based on hyperspectral imaging (HSI) to achieve high-precision non-destructive quantification of VC, SSC, and SP in apples. The model was established using six apple varieties from diverse geographical origins, leveraging hyperspectral data spanning 400–1000 nm with 512 spectral bands. The model framework demonstrated superior performance with raw hyperspectral cube inputs. Optimal predictions for VC and SSC were achieved using full-spectrum modeling (test set: R² VC =0.891, R² SSC =0.807, RPD VC =3.117, RPD SSC =2.337). For SP quantification, feature wavelength selection (403, 430, 551, 617, and 846 nm) via successive projections algorithm (SPA) yielded R²=0.848, RPD=2.642, which aligned with the N-H/C-H vibrational overtones and aromatic amino acid absorption bands. Cross-year validation of 2024 hyperspectral dataset confirmed the robustness of the model, with R 2 values of 0.829, 0.779, and 0.835 (RPD>2.000) for VC, SSC, and SP, respectively. Taken together, this study resolves high-dimensional data redundancy through hybrid architectures and offers a deployable solution for multi-variety fruit quality monitoring.