NorBlueNet: Hyperspectral imaging-based hybrid CNN-transformer model for non-destructive SSC analysis in Norwegian wild blueberries

Abstract

Soluble solids content (SSC) is a vital parameter in blueberries, reflecting the concentration of dissolved sugars (primarily fructose and glucose) and directly influencing the fruit’s sweetness, flavour, and ripeness. As part of this study, Norwegian wild blueberries were carefully hand-picked from a forest in Norway and subsequently imaged using a hyperspectral camera to capture their detailed spectral characteristics. This study introduces NorBlueNet, a hybrid CNN-transformer architecture, for accurately predicting SSC in wild blueberries through hyperspectral imaging and deep learning. This hybrid architecture combines CNN layers for local feature extraction and spatial hierarchy representation, followed by transformer layers that capture global relationships and long-range dependencies. The hybrid approach combines the computational advantages of CNNs with the advanced attention mechanisms of transformers, achieving enhanced accuracy while maintaining computational efficiency. A comprehensive evaluation is conducted by comparing the proposed model with two additional deep learning models on the custom dataset. The results indicate that the NorBlueNet achieves the highest prediction accuracy, with an $R^2$ = 0.98, RMSE = 0.0136, and RPD = 9.3759 thereby demonstrating its superior performance. To foster community engagement, collaboration and facilitate re-implementation of our work, we have made our code available at:https://github.com/NorBlueNet.