光譜標準化模式應用於水果糖度檢測之研究

Abstract

近年來，近紅外光（Near Infrared, NIR）光譜分析技術已經廣泛地應用於不同領域之中，然而每台分光光度計間儀器的構造差異卻在使用上造成很大之限制，即使進行相同樣本之量測，獲得之光譜響應也不盡相同，造成由主儀器（Master）所建立之光譜資料庫和預測模式，並無法直接應用於其他副儀器（Slave）作預測。為了解決此問題，光譜標準化之技術便因應而生。 現今之光譜標準化方法大略可分為以下幾種策略，包含光譜前處理、模式調整、光譜映射、以及強健檢量線之建立。有鑑於生物材料成份之複雜性，本研究將光譜前處理以及光譜映射兩種標準化策略結合，實際應用於水果糖度檢測之標準化。在此使用了標準常態變數（Standard Normal Variate, SNV）、多元性散射修正（Multiplicative Scatter Correction, MSC）以及一次微分三種光譜前處理方法，搭配片段直接標準化（Piecewise Direct Standardization, PDS）以及支援向量標準化（Support Vector Standardization, SVS）兩種線性和非線性之光譜映射方法進行水果光譜之標準化，並使用修正型部份最小平方迴歸（Modified Partial Least Square Regression, MPLSR）進行水果糖度檢量線之建立，探討不同型號、配件儀器標準化後的光譜預測能力。 在兩部相同型號之FOSS NIRSystem 6500，分別搭配快速成份分析儀（Rapid Content Analyzer, RCA）以及光纖（Fiber Optic Probe, FOP）兩種不同配件的儀器光譜標準化方面，由一次微分搭配PDS有著最佳之標準化結果，相對於主儀器預測組之最佳SEP為1.05 °Brix、RSEP為7.33 %而言，副儀器光譜標準化前之SEP為3.01 °Brix、RSEP為21.11 %，標準化後預測能力提升至SEP為1.27 °Brix、RSEP為9.07 %；而FOSS NIRSystem 6500搭配RCA與FOSS NIRSystem On-line 6500，兩部相同光感測器不同構造之儀器方面，同樣由一次微分搭配PDS有著最佳之標準化能力，相對於主儀器預測組之SEP為0.67 °Brix、RSEP為4.91 %而言，光譜標準化前後副儀器之SEP原先的15.95 °Brix下降至0.95 °Brix，而RSEP也從112.94 %下降至6.73 %，已有著和主儀器相近之預測能力，並且可以直接使用主儀器之資料庫進行水果糖度之預測。而在FOSS NIRSystem 6500搭配RCA與水果品質非破壞性線上檢測系統，兩部構造完全相異儀器之標準化中，由SVS有著最佳之標準化能力，能將標準化後之SEP從6.92 °Brix下降至1.38 °Brix，RSEP從48.52 %下降至9.47 %。

Near-infrared (NIR) has been widely used in many fields; however, there are still some constraints in application due to the constructional differences among spectrometers. Even if the measurements conducted on the same sample, the spectral responses of instruments are different. Therefore, the databases and calibration models built on Master instrument can’t be used for the spectra measured by Slave instrument. In order to solve this problem, the method of Spectral Standardization has been developed. In this research, Spectral Standardization Models were practically applied to the Sugar Content Inspection for Fruits. In this study, Standard Normal Variate (SNV)、Multiplicative Scatter Correction (MSC) and 1st Derivatives Spectral Pretreatment method were combined with the Spectral Mapping methods, such as Piecewise Direct Standardization, (PDS) and Support Vector Standardization (SVS) , in order to find the best Standardization strategy among different instruments. The Standardization ability will be evaluated by the Spectral prediction ability after standardization. The best result for same type spectrometer but different associated apparatus was using PDS with 1st Derivatives. After Standardization, the SEP of slave instrument could be reduced from 3.01 °Brix to 1.27 °Brix, the RSEP also reduced from 21.11 % to 9.07 %. The best result for same type detectors but different optical design was using 1st Derivatives and PDS together, the SEP and RSEP of slave instrument were 15.95 °Brix and 112.94 % before Spectral Standardization; the SEP and RSEP were 0.95 °Brix and 6.73 % after Spectral Standardization, while the SEP and RSEP of master instrument were 0.67 °Brix and 4.91%. The Slave spectra after Standardization almost have the same prediction ability of sugar content as Master spectra, and can be directly applied to the databases built by master instrument.