MARTr-KiwiNet: Accurate classification of kiwifruit sugar content integrating multi-scale attention residual convolution and transformer

The sugar content (Soluble Solids Content, SSC) of kiwifruit is a key indicator of its intrinsic quality and market value. The traditional detection and classification methods relying on chemometric sampling are destructive and inefficient, failing to meet the demands of modern fruit industries for high-efficiency and non-destructive grading. The rapid development of Near-Infrared Spectroscopy (NIRS) and artificial intelligence technologies offers a new approach for nondestructive detection and grading. However, existing models exhibit insufficient capability in capturing complex kiwifruit spectral features. This study proposes a novel hybrid Deep Learning (DL) model named MARTr-KiwiNet, which integrates Multi-scale Attention Residual convolution and Transformer. The Multi-scale Convolution extracts local spectral features with different receptive fields, combined with the Channel Attention mechanism to enhance fine-grained feature information, and ensures training stability through Residual Connections. The Transformer effectively captures the global correlations and long-range dependencies in the spectral sequences. After outlier removal, SSC classification, and spectral preprocessing, the model was trained and optimized using the modeling set. The model's performance was systematically evaluated through ablation experiments, single validation, five-fold cross-validation (5-CV), and independent test set validation. The results show that MARTr-KiwiNet achieved a single validation accuracy of 97.16% on the modeling set, with a 5-CV average accuracy of 97.01%. On an independent cross-year test set, the accuracy of a single test was 93.67%, and the average accuracy of 5-CV external validation was 92.89%, indicating its excellent classification performance and generalization ability.