Abstract Variables selection methods have been proven successfully in the field of visible-near infrared spectroscopy (vis-NIRS) to optimize the predictive performance of regression models. However, because only selected spectral variables have been used, and discarding of residual spectral variables result in loss of spectral information. In this work, soluble solids content (SSC) of Chinese bayberry was non-destructively measured by a portable vis-NIRS equipment in the interactance spectral acquiring mode, and combined with a consensus modeling approach. The first member model was developed with the full spectra by applying a competitive adaptive reweighting algorithm (CARS), and the remainder developed successively with the residual spectral variables until the performance of the CARS- partial least square (PLS) model was not improved over that of the residual-spectral-based PLS model. A series of consensus models were developed with different number of top member-models in a fusing strategy of distributing the weightings. Results showed the residual spectral wavelengths after variables selection still reserved some useful information. In total, five CARS-PLS member models were developed. All consensus models performed better than any univocal member model, and the second consensus model F2 that fused the top two- member models performed best. Compared to the full-spectral-based PLS model, the F2 model promoted its performance with RMSECV of 0.80 by 11.3 % in the calibration set, and an RMSEP of 0.85 by 9.1 % in prediction set. The fusing strategy combined with member models that were successively developed with the discarded spectral variables utilized more useful information and improve the predictive performance.

中文翻译：

---

连续CARS-PLS融合模型用于vis-NIRS技术测定杨梅可溶性固形物含量

摘要 变量选择方法已在可见-近红外光谱 (vis-NIRS) 领域得到成功证明，以优化回归模型的预测性能。然而，因为只使用了选定的光谱变量，丢弃剩余的光谱变量会导致光谱信息的丢失。在这项工作中，杨梅的可溶性固形物含量（SSC）在交互光谱采集模式下通过便携式 vis-NIRS 设备无损测量，并结合一致建模方法。第一个成员模型是通过应用竞争性自适应重新加权算法（CARS）开发的全光谱，其余的随着残差谱变量相继发展，直到 CARS 偏最小二乘 (PLS) 模型的性能没有比基于残差谱的 PLS 模型的性能有所提高。在分布权重的融合策略中，使用不同数量的顶级成员模型开发了一系列共识模型。结果表明，变量选择后的剩余光谱波长仍然保留了一些有用的信息。总共开发了五个CARS-PLS成员模型。所有共识模型的表现都比任何单一成员模型都要好，融合了顶级双成员模型的第二个共识模型 F2 表现最好。与基于全光谱的 PLS 模型相比，F2 模型在校准集中以 0.80 的 RMSECV 提升了 11.3% 的性能，以及 0.85 的 RMSEP 提升了 9。1% 在预测集中。融合策略与成员模型相结合，利用丢弃的光谱变量连续开发出更多有用的信息，提高了预测性能。