本研究應用多重線性迴歸(Multiple Linear Regression, MLR)、部份最小平方迴歸(Partial Least Square Regression, PLSR)及類神經網路(Artificial Neural Network, ANN)三種不同檢測模式來探討不同光譜處理(原始光譜、一次微分光譜、二次微分光譜)之近紅外線光譜對果汁糖度檢測的影響。以196個水蜜桃果汁樣本作為各模式之分析與驗証的範例,近紅外線光譜量測範圍為1000~2500nm。在MLR模式中,三種光譜處理對糖度檢測的影響,以二次微分光譜所得的結果最好,原始光譜所得到的結果最差。MLR模式之最大優點在於應用少數相關波長之線性組合即可作為糖度之檢測,在二次微分光譜中2273、1689及2080nm所組成的三波長校正方程式,有最佳的預測結果(r(下标 p^2)=0.974,SEP=0.271,bias=-0.0343)。PLSR模式乃是將整段波長範圍內每個光譜均納入計算,因此在原始光譜及一次微分光譜之校正與預測的結果均較MLR之結果有明顯改善。但是由於運算較為複雜,所以應用時最好擷取有效波長範圍作為輸入;其中原始光譜在部份波長範圍(2050~2350nm)的預測結果最好(r(下标 p^2)=0.987,SEP=0.192,bias=-0.004)。ANN模式之應用,乃是以非線性的關係來描述光譜形狀與糖度的關係,對三種光譜處理的差異性並不明顯,一次微分光譜在2050~2350nm有極佳的預測結果(r2(下标 p^2)=0.986,SEP=0.201,bias=-0.003),但是在實際應用時因為原始光譜可以有效地簡化光譜處理的步驟,而且仍檢測準確,具有較高的應用價值。
Three kinds of models including Multiple Linear Regression (MLR), Partial Least Square Regression (PLSR), and Artificial Neural Network( ANN) were evaluated for the determination of sugar content in peach juice. Three different kinds of mathematical treatments(original, first derivative and second derivative)of spectra in the range of 1000~2500nm were discussed. There were 196 peach juice samples used for analysis in this study. In MLR analysis, the second derivative spectra showed the best results in both calibration and prediction while the original spectra were the worst. The linear combination of effective wavelengths in second derivative mode could prove very good in predictions, the triple-wavelength calibration equation consisted of 2273,1689 and 2080 nm had a high coefficient of determination(0.974)and low SEP(0.271).All the wavelengths in the specified range were included in PLSR analysis, which made the model complex. Therefore, it was necessary to investigate the most effective wavelength range as an input to PLSR. The performance of the original spectra and the first derivative of spectral data was improved compared to those in MLR analysis. The original spectra achieved the best result (r(subscript p^2)=0.987,SEP=0.192,bias=-0.004) if the range of 2050~2350nm was used as an input to PLSR. The ANN model established the nonlinear relationship between the spectrum and the sugar content. The difference in the performance of mathematical treatments of the spectra was little when using ANN analysis. The first derivative of the spectral data in the range of 2050~2350 nm gave the best prediction result(r2(subscript p^2)=0.986,SEP=0.201,bias=-0.003),however, the original spectra, which required no derivative transformation, yielded a comparable result. The ANN model with the original spectra was suggested for applications due to its simplicity and accuracy.