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: R²VC=0.891, R²SSC=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 R2 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.