Improving accuracy and generalization in single kernel oil characteristics prediction in maize using NIR-HSI and a knowledge-injected spectral tabtransformer

Near-infrared spectroscopy hyperspectral imaging (NIR-HSI) is widely used for seed component prediction due to its non-destructive and rapid nature. However, existing models often suffer from limited generalization, particularly when trained on small datasets, and there is a lack of effective deep learning (DL) models for spectral data analysis. To address these challenges, we propose the Knowledge-Injected Spectral TabTransformer (KIT-Spectral TabTransformer), an innovative adaptation of the traditional TabTransformer specifically designed for maize seeds. By integrating domain-specific knowledge, this approach enhances model training efficiency and predictive accuracy while reducing reliance on large datasets. The generalization capability of the model was rigorously validated through ten-fold cross-validation (10-CV). Compared to traditional machine learning methods, the attention-based CNN (ACNNR), and the Oil Characteristics Predictor Transformer (OCP-Transformer), the KIT-Spectral TabTransformer demonstrated superior performance in oil mass prediction, achieving R p 2 = 0.9238 ± 0.0346, RMSE p = 0.1746 ± 0.0401. For oil content, R p 2 = 0.9602 ± 0.0180 and RMSE p = 0.5301 ± 0.1446 on a dataset with oil content ranging from 0.81 % to 13.07 %. On the independent validation set, our model achieved R 2 values of 0.7820 and 0.7586, along with RPD values of 2.1420 and 2.0355 in the two tasks, highlighting its strong prediction capability and potential for real-world application. These findings offer a potential method and direction for single seed oil prediction and related crop component analysis. • The KIT-Spectral TabTransformer model achieved accurate predictions of oil mass and content rate for individual maize seeds. • It outperforms ACNNR and OCP-transformer in accuracy on datasets with oil content rates ranging from 0.81 % to 13.07 %. • Our model demonstrates excellent generalization ability and convergence speed. • The end-to-end design eliminates cumbersome processes.