Deep Learning for Oil Prediction in Camellia Seeds Using HSI

Yuan, W., Zhou, H., Zhang, C., Zhou, Y., Jiang, X., & Jiang, H. (2024). Prediction of oil content in Camellia oleifera seeds based on deep learning and hyperspectral imaging. Industrial Crops and Products, 222(Part 2), 119662. https://doi.org/10.1016/j.indcrop.2024.119662

Camellia oil is valued both nutritionally and commercially. Efficient and non-destructive methods for assessing oil content in Camellia oleifera seeds are essential for optimizing production. This study explores the integration of hyperspectral imaging (HSI) with deep learning (DL) techniques to predict oil content rapidly and accurately.

HSI data in the 400–1000 nm range were captured from Camellia oleifera seeds. Regions of interest were extracted using a threshold segmentation method. A traditional Partial Least Squares Regression (PLSR) model was first applied to evaluate the effects of spectral preprocessing, showing a 7.4% improvement after Standard Normal Variate (SNV) correction. Deep learning models, including a convolutional neural network regression (CNNR) and its enhanced version with an attention mechanism (ACNNR), were then tested using raw spectra. Additionally, dimensionality reduction was performed using algorithms such as Successive Projections Algorithm (SPA), Genetic Algorithms (GA), CNN, and ACNN.

The ACNNR model demonstrated strong prediction capability with R²P = 0.816, RMSEP = 2.552, and RPD = 2.348. When combined with ACNN-based feature extraction, the PLSR model further improved performance, achieving R²P = 0.829, RMSEP = 2.462, and RPD = 2.425. These results confirm the superiority of deep learning approaches in both prediction accuracy and spectral dimensionality reduction. The optimized model also enabled spatial visualization of oil content within the seeds, highlighting the potential of HSI and DL as robust tools for non-invasive oil quality assessment.