蜜瓜品质光谱检测中异常建模样品的综合评判

针对蜜瓜可溶性固形物含量透射光谱检测中，异常建模样品对模型精度的影响及多种可能来源，提出异常样品的综合评判方法。为防止漏判，分别针对不同来源，采用基于预测浓度残差、Chauvenet检验法及杠杆值与学生残差T检验准则对85个建模样品(偏最小二乘法建模)进行初步判别，共判别出9个疑似异常样品。为防止误判，对疑似样品逐一回收，考察其对建模与预测精度的影响。先后回收5个样品后，所建校正模型相关系数r为0.889，均方根校正偏差RMSEC为0.601°Brix，对35个未知样品的均方根预测偏差RMSEP为0.854°Brix，比未剔除异常样品前所建模型(r=0.797, RMSEC=0.849°Brix, RMSEP=1.19°Brix)精度明显提高，比剔除全部疑似异常样品所建模型(r=0.892, RMSEC=0.605°Brix, RMSEP=0.862°Brix)更稳定，预测精度更高。

Outlier Sample Discriminating Methods for Building Calibration Model in Melons Quality Detecting Using NIR Spectra

Outlier samples strongly influence the precision of the calibration model in soluble solids content measurement of melons using NIR Spectra. According to the possible sources of outlier samples, three methods (predicted concentration residual test; Chauvenet test; leverage and studentized residual test) were used to discriminate these outliers respectively. Nine suspicious outliers were detected from calibration set which including 85 fruit samples. Considering the 9 suspicious outlier samples maybe contain some no-outlier samples, they were reclaimed to the model one by one to see whether they influence the model and prediction precision or not. In this way, 5 samples which were helpful to the model joined in calibration set again, and a new model was developed with the correlation coefficient (r) 0.889 and root mean square errors for calibration (RMSEC) 0.601°Brix. For 35 unknown samples, the root mean square errors prediction (RMSEP) was 0.854°Brix. The performance of this model was more better than that developed with non outlier was eliminated from calibration set(r=0.797, RMSEC=0.849°Brix, RMSEP=1.19°Brix), and more representative and stable with all 9 samples were eliminated from calibration set(r=0.892, RMSEC=0.605°Brix, RMSEP=0.862°Brix).