【摘要】 为构建基于高光谱数据和域适应算法的不同批次北柴胡种子跨年份精准鉴定模型，采集10个批次北柴胡（Bupleurum chinense DC.）及5种近缘物种400～1 000 nm波段的高光谱数据，利用随机森林（RF）、支持向量机（SVM）、多层感知器（MLP）、偏最小二乘判别分析（PLS-DA）及集成学习算法建立北柴胡种子跨批次鉴定模型，通过竞争性自适应权重（CARS）结合协方差对齐（CORAL）算法探索提高模型对不同批次种子的跨年份数据预测能力的可行性，最后使用SHAP算法对模型进行可解释性分析。结果表明：SNV-PLS-DA模型的预测准确率为85.6%。采用CARS算法筛选出97个关键特征波段，构建的SNV-CARS-PLS-DA模型预测准确率为91.4%。2024和2025年采集同一批次（2024年10月）收获的万荣柴胡的跨年份高光谱数据间存在以线性偏移为主导的混合偏移（集中于830～900 nm）。经CORAL算法校准后，模型预测2024年采集的光谱数据准确率从79.7%提升到94.3%。SHAP算法表明673.5～746.6、810.6～842.7和907.8～984.7 nm是模型预测的关键波段。综上，CORAL算法适用于校正北柴胡不同批次跨年份光谱数据的偏移，能够提高模型预测的鲁棒性，可用于不同批次北柴胡种子跨年份精准鉴定。更多还原

【Abstract】 To construct a cross-year accurate identification model for different batches of Bupleurum chinense DC. seeds based on hyperspectral data and domain adaptation algorithms, hyperspectral data in the 400 to 1 000 nm band were collected from 10 batches of B. chinense DC. and 5 related species. Machine learning models for cross-batch discrimination were established using random forest(RF), support vector machine(SVM), multi-layer perceptron(MLP), partial least squares discriminant analysis(PLS-DA), and ensemble learning algorithms. The feasibility of improving the model’s prediction ability on the cross-year data of different seed batches was explored by combining the competitive adaptive reweighted sampling(CARS) with CORrelation ALignment( CORAL) algorithm. Finally, the SHAP algorithm was used to conduct interpretability analysis on the model. The results showed that: The prediction accuracy of the SNV-PLS-DA model reached 85. 6%. Using the CARS algorithm, a total of 97 key characteristic bands were identified, and the prediction accuracy of the constructed SNV-CARS-PLS-DA model was 91. 4%. There existed a mixed shift dominated by linear shift(concentrated within 830 to 900 nm) between the hyperspectral data collected in 2024 and 2025 for the same batch of Wanrong Bupleurum harvested in October 2024. After calibration by the CORAL algorithm, the accuracy of the model in predicting the spectral data in 2024 increased from 79. 7% to 94. 3%. The SHAP algorithm indicates that the bands of 673. 5-746. 6, 810. 6-842. 7, and 907. 8-984. 7 nm were the key bands for the model prediction. In conclusion, the CORAL algorithm is suitable for correcting the shift of spectral data of cross-year for different batches of Bupleurum chinense DC. seeds, and can improve the robustness of model prediction.更多还原