Detection of blueberry based on hyperspectral imaging and deep learning

Blueberries with different sugar levels have different flavors and market values. This study utilized fractional-order derivative (FOD) and the improved laplacian eigenmap (ILE) to preprocess and select bands for blueberries hyperspectral data of different cultivars and maturity levels in complex scenarios. The ability of hyperspectral imaging technology combined with deep learning models to detect the sugar content of blueberries of different cultivars and maturity was explored, which can achieve rapid detection of different blueberry sugar contents and meet the requirements of blueberry mass production. In this paper, three blueberry cultivars, F6, L11, and L25, were used to classify blueberry ripeness into three categories (ripe, semi-ripe, and unripe). Firstly, the hyperspectral data of blueberries is preprocessed using the MSC + FOD combination to enhance the spectral information of different blueberry samples and highlight the subtle features. Then, the ILE based on distance function and kernel function is used to extract characteristic bands, so that it can better adapt to the data processed by FOD and efficiently select characteristic bands.Finally, a customized shallow convolutional neural network (CNN) model is built for training, and this method is compared with other traditional preprocessing, band screening, and modeling methods to obtain the optimal prediction model. The results show that after MSC and FOD processing, the CNN model at the 1.10 order has the best overall effect（Rp² = 0.8597,RMSEP = 0.8552%,RPDP = 2.6694）, which is used ILE to extract the characteristic spectrum and has good robustness and accuracy. At the same time, this also fully verifies the effectiveness and applicability of FOD combined with ILE in processing blueberry hyperspectral data in complex scenarios, and the combination of lightweight CNN model can significantly improve the accuracy and performance of predicting blueberry sugar while simplifying the model structure. The above results show that the model proposed in this study has great potential application value for the detection of blueberry sugar levels of different varieties and ripeness levels in complex scenarios, which can provide a reference for subsequent related studies.