Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/37180
標題: 數據前處理對近紅外光光譜預測水稻品質之改善
The Pretreatment of Near-Infrared Spectrum to Improve the Predictivity of Rice Quality
作者: 閻菁珠
Yen, Ching-Chu
關鍵字: 近紅外光光譜儀;nir
出版社: 農藝學系
摘要: 
中文摘要
近紅外光分析儀運用於稻米成分分析工作時,其測定時間短,測定後樣本粉末可再重複利用於其它特性分析,但分析原理與方法有異於一般的分析工具,其需要校正線。物體的近紅外光光譜之構成因子,除了各組成份所引起的吸收外,還有干擾吸收(interference absorption),與基線飄移(baseline shift)等因子,導致預測效能降低,可利用各種數據處理方法降低干擾。600nm碘呈色度可被用來預測稻米的食味,但它是一利用近紅外光光譜預測表現不佳的性狀。蛋白質含量亦可被用來預測稻米的食味,而糙米粒蛋白質含量之預測效能不如糙米粉,但糙米粒因可減少研磨之時間進而降低成本。故本研究利用各種數據處理方法對稻米樣本之光譜值先進行處理,再利用淨最小平方法進行校正,以期能達到增加解析度及減少干擾的功能,進而得到較佳的建模效能及預測效能。
在600nm碘呈色度使用之稻米材料,包括秈稻155品種、稉稻120品種及糯稻61品種共計336品種;在蛋白質含量使用之稻米材料,共計932品種糙米粒和671品種糙米粉,將樣品進行NIR光譜掃瞄及化學測定後,對原光譜值再進行各種數據前處理,再利用淨最小平方法對此樣本作校正,求得校正統計指標值,再將驗證組樣本作驗證,求得驗證統計指標值。
在600nm碘呈色度所有的各種數據前處理中,以一次差分(segment=1,gap=0)加上正規化處理之預測效能表現最佳,選擇14個成分來建構最適模型,其預測效果最佳,模型的 、 、 、 與 等效能指標分別為0.897、0.0331、0.807、0.0420與1.718, 為26.895。一次差分(segment=1,gap=0)加上正規化處理後表現較原光譜12個成分所建構之最適模型表現佳,其預測效能提昇。其 值亦較原光譜小,表示逢機干擾降低。
在糙米粒進行各種數據前處理的預測效能表現中,以平滑化及未將反射光譜值作 轉換只作正規化處理表現較佳,此二種數據前處理方法是糙米粒蛋白質含量預測表現最佳的,推測可能是因其可降低雜波干擾。此外,一次差分較二次差分及三次差分表現佳,差分時segment及gap設定值大於1的表現亦較佳。但各種數據處理對糙米粒預測效能之改善皆不如糙米粉表現佳,或許是因這些數據處理方法,無法將粒徑差異所導致的光譜散射問題有效改善,所以將來可以嘗試將光譜作MSC處理後,再進行差分處理,以達有效改善糙米粒預測效能之目的。

ABSTRACT
Near-infrared reflectance spectroscopy (NIRS) play an important role on analysising components of rice. It spend less time in predicting the quality of rice samples and repeat using them. But it is necessary to develop the calibration equation. Near-infrared reflectance spectra consist of components reflectance spectra, interference absorption and baseline shift which decrease ability of prediction. Pretreatments could decrease noise to increase ability of prediction. 600 nm Starch-iodine blue value of residual liquid (BV) could be used to predict the palatability of rice. But the performance of model-building and prediction demonstrated that the calibration for BV by NIRS is less accurate. Protein content could also be used to predict the palatability of rice. But the performance of model-building and prediction demonstrated that the calibration for protein content in brown rice is less accurate than that in brown rice flour. However, brown rice samples could spend less time and money. The objective of this study was to improve the performance of model-building and prediction in developing the calibration equation using partial least squares regression (PLSR) by derivative, smooth and normalization pretreatments which could increase resolution and decrease noise, respectively.
A total of 336 rice samples, including 155 indica, 120 japonica, and 61waxy rices, were employed in analyzing the BV. A total of 932 brown rice samples and 671 brown rice flour samples, were employed in analyzing the protein content of rice. The absorbance of the BV was measured with a colorimeter at the wavelength of 600nm. A Bran + Luebbe InfraAlyzer 500 was used to collect spectrum measurement for each sample. The performance of model-building and prediction was evaluated using PLSR by derivative, smooth and normalization pretreatments.
The PLSR model of BV with 14- components by first derivative (segment =1, gap=0) + normalization pretreatment gave the highest correlation coefficient and the lowest standard error of prediction. The , , , , and of the model were 0.897, 0.0331, 0.807, 0.0420, 1.718 and 26.895, respectively. The performance of model-building and prediction was better than that in PLSR with 12 components of raw data. The IRV value of the pretreatment model was smaller than that of raw data model, suggesting that less random noise was involved in the pretreatment model.
The PLSR model of brown rice by smooth and only normalization pretreatments which decrease noise gave the highest correlation coefficient and the lowest standard error of prediction. The PLSR model by first derivative was better than that by second and third derivative, and segment and gap which over 1 of derivative was better. The performance of model-building and prediction in brown rice flour was better than that in brown rice with pretreatments. The derivative, smooth and normalization pretreatments couldn't decrease the effect of multiplicative scatter led by particle size. So it is suggested that using multiplicative scatter correction (MSC) which delete multiplicative scatter and then derivative which increase resolution to get better performance of model-building and prediction in the future.
URI: http://hdl.handle.net/11455/37180
Appears in Collections:農藝學系

Show full item record
 

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.