查詢結果分析
來源資料
頁籤選單縮合
題名 | 近紅外線光譜分析中逐步回歸與主成分回歸的比較=The Comparison of Stepwise Regression and Principal Component Regression in Near-Infrared Spectroscopy |
---|---|
作者 | 蘇堃綺; 郭寶錚; 洪梅珠; Su, Kun-chi; Kuo, Bo-jein; Hong, Mei-chu; |
期刊 | 中華農學會報 |
出版日期 | 20010600 |
卷期 | 2:3 2001.06[民90.06] |
頁次 | 頁226-240 |
分類號 | 341.94 |
語文 | chi |
關鍵詞 | 近紅外線光譜; 多元共線性; 逐步回歸法; 主成分回歸; 直交成份; Near-infrared reflectance spectroscopy; Multicollinearity; Stepwise regression; Principal component regression; Orthogonal components; |
中文摘要 | 在國內,利用近紅外線光譜技術(near infrared reflectance spectroscopy, NIRS)分析稻米品質,九不建立本土化的食味計,是目前米質檢測的一大趨勢。而建立檢量線的統計方法,至今仍多採行逐步回歸法(stepwise regression),也就是一般所謂的複線性回歸(multiple linear regression, MLR),但利用此法所選出對化學值重要的波長間,常存在嚴重的多元共線性(multicollinearity)問題,因此易導致回歸係數估計上的不穩定。本報告中,除了利用MLR法建立稻米的直鏈澱粉含量、水分含量與煮飯液碘呈色度的檢量模型外,更進一步利用解決解釋變數間多元共線性問題的統計方法—主成分分析法(principal component regression, PCR)—分析上述資料。以MLR法分析直鏈澱粉含量、水分含量與煮飯液碘呈色度所得到的結果,校正組與驗證組的相關係數分別為0.921、0.979、0.841與0.896、0.963及0.754,校正標準機差(standard error of calibration, SEC)與預測標準機差(standard error of prediction, SEP)則分別為2.567%、0.212%、0.0380與2.984%、0.280%及0.0437;若以PCR法進行分析,校正組與驗證組的相關係數分別為0.931、0.980、0.827與0.910、0.962及0.741,標準機差與則分別為2.40.%、0.205%、0.0414與2.786%、0.386%及0.0447。兩種方法所建立的模型表現差異不大,但以MLR法建立的模型,都有逢機變異指數(index of random variation, IRV)值過大的現象,尤其是直鏈澱分含量與煮飯液碘呈色度;若利用PCR法後,各個模型的值都明顯降低,顯示PCR法已將波長間的多元共線性問題排除,因此在回歸係數的估計上更為穩定。 |
英文摘要 | Nowadays, analyzing rice quality by the near infrared reflectance spectroscopy to develop the domestic rice taste analyzer is a trend in Taiwan. However, it still takes stepwise regression called multiple linear regression (MLR) as the statistical method for building calibration lines. By this method, there is usually a serious multicollinearity problem leading the estimation of regression coefficients to be unstable. In our report, MLR was sued to build the calibration models of amylase content, moisture content, and iodine-starch value for rice, respectively. In addition, we also analyze these chemical values by the principal component regression (PCR) which can solve the problem of multicollinearity. The correlation coefficients of amylase content, moisture content, and iodine-starch value by MLR are 0.921, 0.979, and 0.841 for calibration sets, and 0.896, 0.963 and 0.754 for validation sets, respectively. The values of SEC are 2.567%, 0.212%, and 0.0380 and the values of SEP are 2.984%, 0.280%, and 0.0437, respectively. By PCR method, the correlation coefficients are 0.931, 0.980, and 0.827 for calibration sets and 0.910, 0.962 and 0.741 for validation sets, respectively. The values of SEC are2.403%, 0.205%, and 0.0414 and the values of SEP are 2.768%, 0.286%, and 0.0447, respectively. The difference between the models by the two different methods is not significant. But all the IRV values of MLR model seem to be too high, and they all decrease greatly after using PCR method. It implies that the problem of multicollinearity has been excluded. Hence it is more stable in the estimation of the coefficients. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。