Multimodal deep learning for oil content prediction in Camellia oleifera fruits using image, morphometric, and categorical features

Abstract

Accurate determination of oil content is essential for food composition analysis and quality control in the Camellia oil industry, yet conventional chemical analyses are destructive and difficult to implement at large scale. In this study, a multimodal oil content prediction model (MPCM-OC) was developed as an indirect, non-destructive approach to support oil content assessment in Camellia oleifera fruits based on reference chemical measurements. The proposed framework integrates fruit images, morphometric traits (transverse diameter, longitudinal diameter, and fruit shape index), and categorical information (cultivar, maturity stage, acquisition date, and sampling location), using separate feature extraction networks and an adaptive fusion module. Seed oil content values obtained using standardized chemical analysis served as reference data. The MPCM-OC model achieved an overall coefficient of determination (R²) of 0.8353, with a mean absolute percentage error of 13.38 %, a mean absolute error of 4.52, and a root mean squared error of 6.30. Ablation and comparative analyses showed that incorporating morphometric and categorical features with image data consistently improved prediction accuracy over image-only models. The proposed framework serves as a rapid, low-cost complementary tool for preliminary screening and batch-level quality evaluation, enhancing efficiency in food composition analysis and quality control of Camellia oleifera.