Prediction of oil content in Camellia oleifera seeds based on deep learning and hyperspectral imaging

Camellia oil has high commercial and nutritional value. This study combined hyperspectral imaging (HSI) technique with deep learning (DL) to realize rapid and accurate prediction of oil content for Camellia oleifera seeds. First, spectral images of Camellia oleifera seeds from the 400–1000 nm rang were captured, and spectral data from the regions of interest were extracted based on a threshold segmentation method. Then, the partial least squares regression (PLSR) model was used to examine the influence of various preprocessing techniques. It was found that the model performance improved by 7.4 % after standard normal variate (SNV) preprocessing. Meanwhile, an attention mechanism (AM) was introduced into the convolutional neural network regression (CNNR) model, known as ACNNR. The prediction performance of the established models based on full spectra using the traditional (PLSR) and DL methods (CNNR and ACNNR) were compared. Specifically, only raw spectra were taken as inputs for the DL models. The study demonstrated that the model constructed using ACNNR achieved satisfactory results, with R²_P, RMSEP, and RPD values of 0. 816, 2.552, and 2.348 in the prediction set, respectively. Moreover, to reduce data dimensionality, the full spectra were downscaled using the successive projections algorithm (SPA), genetic algorithms (GA), CNN, and ACNN. Compared to traditional modelling and dimensionality reduction methods, DL showed excellent performance. Finally, the experimental results indicated that the PLSR model developed using spectral features extracted by the ACNN method achieved the optimal performance, with R²_P, RMSEP, and RPD values of 0.829, 2.462, and 2.425 in the prediction set, respectively. The optimal simplified model was utilized to visualize the spatial distribution of oil content in Camellia oleifera seeds. Generally, the HSI technique combined with DL provides a reliable and effective method for achieving non-destructive detection and visualization of oil content in Camellia oleifera seeds.