劉小丹,馮旭萍,劉 飛,何 勇
?
基于近紅外高光譜成像技術(shù)鑒別雜交稻品系
劉小丹,馮旭萍,劉 飛,何 勇※
(浙江大學(xué)生物系統(tǒng)工程與食品科學(xué)學(xué)院,杭州 310058)
種子的篩選和鑒別是農(nóng)業(yè)育種過(guò)程中的關(guān)鍵環(huán)節(jié)。該文基于近紅外高光譜成像技術(shù)(874~1 734 nm)結(jié)合化學(xué)計(jì)量學(xué)方法以及圖像處理技術(shù)實(shí)現(xiàn)雜交稻種的品系鑒別及可視化預(yù)測(cè)。采集了3類(lèi)不同品系共2 700粒雜交水稻的高光譜圖像,用SPXY算法,按照2∶1的比例劃分建模集和預(yù)測(cè)集?;谒緲颖镜墓庾V特征,采用主成分分析(PCA)方法初步探究3類(lèi)樣本的可分性。采用連續(xù)投影算法(SPA),提取出7個(gè)特征波長(zhǎng):985.08、1 106、1 203.55、1 399.04、1 463.19、1 601.81、1 645.82 nm。基于特征波長(zhǎng)和全波段光譜,建立了偏最小二乘判別分析(PLS-DA)和支持向量機(jī)(SVM)模型。試驗(yàn)結(jié)果表明,所建模型判別效果較好,識(shí)別正確率均達(dá)到了90%以上,其中,SVM 模型的判別效果優(yōu)于PLS-DA 模型,基于全譜的判別分析模型結(jié)果優(yōu)于基于特征波長(zhǎng)的判別模型。結(jié)合SPA-SVM校正模型和圖像處理技術(shù),生成樣本預(yù)測(cè)偽彩圖,可以直觀的鑒別不同品系的水稻種子。結(jié)果表明,近紅外高光譜成像技術(shù)可以實(shí)現(xiàn)雜交稻的品系識(shí)別及可視化預(yù)測(cè),為農(nóng)業(yè)育種過(guò)程中種子的快速篩選及鑒定提供了新思路。
圖像處理;光譜分析;無(wú)損檢測(cè);高光譜成像;水稻種子;連續(xù)投影算法
中國(guó)對(duì)雜交水稻的研究始于1964年,隨后三系雜交水稻和兩系雜交水稻配套成功并被迅速示范推廣,雜交水稻的出現(xiàn)解決了全國(guó)乃至全球的糧食安全問(wèn)題,取得了巨大的社會(huì)和經(jīng)濟(jì)效益。根據(jù)國(guó)家水稻品種區(qū)試、審定及推廣情況統(tǒng)計(jì)資料,兩系雜交水稻在產(chǎn)量和米質(zhì)方面均優(yōu)于三系、抗病蟲(chóng)性與三系雜交水稻大致相當(dāng)[1],但目前市場(chǎng)上水稻種子種類(lèi)繁多,種類(lèi)系別間相似性較大,兩系和三系雜交水稻種子難以區(qū)分,魚(yú)目混珠的事件時(shí)有發(fā)生,嚴(yán)重影響作物的產(chǎn)量及農(nóng)民的利益。傳統(tǒng)的種子鑒別方法主要有形態(tài)鑒別和化學(xué)鑒別。這些方法主觀性強(qiáng)、耗時(shí)費(fèi)力、對(duì)種子具有一定的破壞性。因此,急需一種快速無(wú)損檢測(cè)技術(shù)實(shí)現(xiàn)水稻種子的品系鑒別。
高光譜圖像技術(shù)是一種結(jié)合機(jī)器視覺(jué)和光譜技術(shù)的新興快速無(wú)損檢測(cè)技術(shù),能夠同時(shí)獲取圖像信息和光譜信息,近年來(lái)被廣泛的用于種子的品質(zhì)及類(lèi)別鑒定[2-6]。Williams等[7]利用近紅外高光譜技術(shù)實(shí)現(xiàn)了玉米不同硬度的可視化分析,直觀的區(qū)分不同硬度的玉米;Zhang等[8]利用高光譜技術(shù)結(jié)合判別分析模型對(duì)6類(lèi)玉米種子進(jìn)行鑒別;Kong等[9]將多種化學(xué)計(jì)量學(xué)分析方法應(yīng)用到基于高光譜技術(shù)的雜交稻品種鑒別中,判別效果較好。但上述研究中,樣本量都較少,而且在對(duì)水稻種子的研究中并未涉及對(duì)水稻系別的區(qū)分。在實(shí)際農(nóng)業(yè)育種及生產(chǎn)中,需要因地制宜在大批量的水稻種子中篩選和鑒定合適的水稻品系,因此,探究一種水稻品系的快速篩選鑒定技術(shù)顯得尤為重要。
本研究將近紅外高光譜成像技術(shù)與化學(xué)計(jì)量學(xué)方法相結(jié)合,建立分類(lèi)模型鑒別雜交稻品系,并結(jié)合圖像處理技術(shù),基于判別效果較好的數(shù)學(xué)模型生成水稻種子的可視化預(yù)測(cè)偽彩圖,進(jìn)而直觀的對(duì)3類(lèi)不同品系的水稻種子進(jìn)行鑒別。
試驗(yàn)所用3類(lèi)水稻種子由江蘇省明天種業(yè)科技有限公司提供,分別為兩系雜交水稻種子深兩優(yōu)862,和兩優(yōu)713,三系雜交水稻種子內(nèi)2優(yōu)6號(hào)。所有種子均為正常品質(zhì),外觀沒(méi)有明顯瑕疵。實(shí)驗(yàn)共用水稻種子2 700粒,每類(lèi)各900粒。采用SPXY[10]算法,按照2∶1的比例劃分建模集和預(yù)測(cè)集,得到1 800粒雜交稻種作為建模集,900粒雜交稻種作為預(yù)測(cè)集。
試驗(yàn)采用高光譜成像系統(tǒng)(圖1)獲取水稻種子的圖譜信息。該系統(tǒng)主要由成像光譜儀、鏡頭、CCD相機(jī)線光源、電控移位平臺(tái)、計(jì)算機(jī)、暗箱,樣品臺(tái)等組成。
圖1 高光譜成像系統(tǒng)
系統(tǒng)的光譜分辨率為5 nm,光譜范圍為874~1 734 nm,近紅外高光譜圖像分辨率為320×256像素。為得到清晰可用的光譜圖像,在光譜采集前,需要對(duì)試驗(yàn)系統(tǒng)的相關(guān)參數(shù)進(jìn)行設(shè)置。經(jīng)過(guò)反復(fù)對(duì)比調(diào)試,得到最適的系統(tǒng)參數(shù):物鏡的高度為29.5 cm,曝光時(shí)間3.2 ms,平臺(tái)移動(dòng)速度為23 mm/s。為了消除各波段下光強(qiáng)度分布不均以及鏡頭中存在的暗電流所產(chǎn)生的噪聲,采集標(biāo)準(zhǔn)白色校正板(反射率接近100%)的高光譜圖像作為白色標(biāo)定圖,用黑板遮擋鏡頭,采集高光譜圖像(反射率接近0)作為黑色標(biāo)定圖,利用黑白標(biāo)定圖對(duì)樣本的高光譜圖像進(jìn)行矯正[11-12],公式如下
式中0代表黑白矯正后的水稻種子高光譜圖像,為水稻種子的原始高光譜圖像,為黑色標(biāo)定圖,為白色標(biāo)定圖。
采集圖像時(shí),CCD相機(jī)中的線列探測(cè)器在光學(xué)焦面的垂直方向做橫向掃描,同時(shí)在樣品臺(tái)移動(dòng)方向上做縱向掃描,以此獲得樣本在整個(gè)平面的光譜圖像。為了提高采集效率,將多粒水稻種子規(guī)則的排列于樣品臺(tái)上,同時(shí)獲取多個(gè)樣本的光譜信息。文中采用ENVI4.6軟件處理黑白校正后的高光譜圖像,基于MATLAB 2014a軟件提取與高光譜圖像對(duì)應(yīng)的光譜數(shù)據(jù)。
高光譜圖像包含圖像和光譜信息,全波段光譜不僅數(shù)據(jù)量大,而且存在大量的冗余和共線信息,影響模型的準(zhǔn)確性和計(jì)算速度[13-14]。因此采用連續(xù)投影算法(successive projections algorithm,SPA),從水稻種子的平均光譜中提取特征波長(zhǎng),以簡(jiǎn)化模型,提高模型的可靠性。SPA是一種前向特征變量選擇方法,其選擇的是含有最少冗余信息及最小共線性的變量組合,因此在光譜特征波長(zhǎng)選擇中有廣泛的應(yīng)用[15-16]。本文用MATLAB 2014a運(yùn)行連續(xù)投影算法,從波長(zhǎng)為975~1 646 nm的光譜中提取特征波長(zhǎng),輸入建模集和預(yù)測(cè)集樣本的光譜數(shù)據(jù)和類(lèi)別序號(hào),設(shè)定特征波長(zhǎng)選擇數(shù)量范圍為5~20,共得到7個(gè)特征波長(zhǎng)用于后續(xù)建模分析。
本研究采用主成分分析(PCA)初步探究3類(lèi)水稻的可分性,運(yùn)用偏最小二乘分析(PLS-DA)和支持向量機(jī)(SVM)算法建立基于全譜和特征波長(zhǎng)的水稻品系鑒別模型。PCA是一種常用有效的數(shù)據(jù)降維壓縮算法,其基本原理是通過(guò)線性變換將原始數(shù)據(jù)中的多個(gè)相關(guān)變量轉(zhuǎn)換成新的綜合變量(主成分),新的變量中前幾個(gè)貢獻(xiàn)率大的主成分涵蓋了原始數(shù)據(jù)的主要信息[17-18]。因此,本文保留前3個(gè)主成分進(jìn)行分析,通過(guò)比較樣本在3個(gè)主成分上的空間分布定性的區(qū)分3類(lèi)水稻。
PLS-DA是1種有監(jiān)督的模式識(shí)別方法,被廣泛的用于光譜數(shù)據(jù)的分類(lèi)分析[19-22]。本文將樣本的光譜數(shù)據(jù)作為自變量,類(lèi)別序號(hào)作為因變量,在Unscrambler 10.1中采用留一法交互驗(yàn)證建立PLS-DA模型,并基于此分類(lèi)模型對(duì)預(yù)測(cè)集樣本進(jìn)行預(yù)測(cè)。根據(jù)樣本的實(shí)際類(lèi)別序號(hào)和模型的預(yù)測(cè)值之差的絕對(duì)值()計(jì)算建模集和預(yù)測(cè)集的判別正確率,由于值帶有小數(shù),在實(shí)際計(jì)算時(shí),設(shè)定閾值為0.5[23],即小于0.5則判別正確,否則判別錯(cuò)誤。模型的參數(shù)即隱含變量(latent variables,LVs)的數(shù)量通過(guò)預(yù)測(cè)殘差平方和確定。
SVM是基于統(tǒng)計(jì)學(xué)習(xí)VC維理論和結(jié)構(gòu)風(fēng)險(xiǎn)最小原理的機(jī)器學(xué)習(xí)算法,可用于數(shù)據(jù)的定性及定量分析[24-26]。SVM通過(guò)核函數(shù),將輸入空間映射到高維空間,構(gòu)建最優(yōu)分類(lèi)面準(zhǔn)確無(wú)誤的將2類(lèi)分開(kāi),并引入懲罰系數(shù)和松弛系數(shù)(,)進(jìn)行修正,使2類(lèi)的分類(lèi)間隔最大從而保證風(fēng)險(xiǎn)最小,在數(shù)據(jù)的分類(lèi)分析中應(yīng)用廣泛[10]。本文在MATLAB 2014a輸入樣本的建模集和預(yù)測(cè)集,運(yùn)行鑒別雜交稻種品系的SVM程序,采用徑向基函數(shù)(RBF)作為SVM模型的核函數(shù),采用網(wǎng)格搜索法在2-8到28尋優(yōu)范圍內(nèi)確定最優(yōu)的(,)參數(shù)組合[27],輸出模型的識(shí)別正確率。
種子形態(tài)特征是育種過(guò)程中種子篩選和鑒別的重要參考依據(jù),傳統(tǒng)的種子形態(tài)特征判別方法是肉眼觀察或工具測(cè)量等方式,費(fèi)時(shí)費(fèi)力且誤差較大。本文通過(guò)高光譜圖像,提取單粒水稻種子的面積周長(zhǎng)等特征,從形態(tài)學(xué)方面鑒別不同品系雜交稻種,為育種篩選提供新的方法和思路。
在MATLAB 2014a中,讀取水稻種子的高光譜圖像,根據(jù)種子及背景的波段比率進(jìn)行閾值分割,得到二值圖像作為掩模圖像去除原始光譜圖中的背景信息,分離出水稻種子。用每粒種子輪廓像素內(nèi)的像素總數(shù)代表種子的面積,輪廓像素?cái)?shù)代表周長(zhǎng),長(zhǎng)軸和短軸長(zhǎng)度的像素?cái)?shù)代表長(zhǎng)軸和短軸的長(zhǎng)度。對(duì)獲取的種子形態(tài)學(xué)特征利用Duncan方法進(jìn)行方差分析,結(jié)果如表1所示。由表可知,三系水稻的4個(gè)形態(tài)學(xué)參數(shù)都顯著大于兩系水稻(<0.01),而兩系水稻的形態(tài)學(xué)參數(shù)較為接近。但由于形態(tài)學(xué)參數(shù)受環(huán)境影響較大,因此需進(jìn)一步結(jié)合水稻種子的光譜信息對(duì)水稻系別進(jìn)行區(qū)分。
表1 種子的形態(tài)學(xué)特征
注:不同小寫(xiě)字母表示在<0.01水平上品種間存在顯著差異。下同。
Note: Different lowercase letters indicate significant difference among the varieties at<0.01. The same as below.
本試驗(yàn)采集水稻種子在874~1 734 nm波長(zhǎng)范圍的近紅外光譜,但受儀器及周?chē)h(huán)境影響,光譜前后端噪聲明顯。因此,去除噪聲明顯的波段,只對(duì)975~1 646 nm(波段31~波段230)間的光譜數(shù)據(jù)進(jìn)行分析,得到3類(lèi)水稻種子的平均光譜如圖2所示。由圖2可知,3類(lèi)水稻種子的光譜曲線趨勢(shì)一致,波峰、波谷的位置相同,但是反射率有所不同,其中兩系雜交稻的反射率接近,三系雜交稻的反射率較高。這可能是由于不同品系水稻種子的化學(xué)成分及分子結(jié)構(gòu)存在差異,為后續(xù)的化學(xué)計(jì)量學(xué)分析提供了依據(jù)。
圖2 雜交稻樣本的平均反射光譜
對(duì)3類(lèi)不同品系水稻種子的光譜進(jìn)行主成分分析(PCA),結(jié)果如圖3所示。前3個(gè)主成分累計(jì)貢獻(xiàn)率為99.93%(PC1,PC2,PC3的貢獻(xiàn)率分別為86.4%,12.8%,0.73%),解釋了絕大部分變量。由圖3可知,3類(lèi)水稻種子成簇分布,僅有少部分重疊區(qū)域,表明3類(lèi)水稻種子在PCA三維得分圖中存在分類(lèi)趨勢(shì),可進(jìn)一步建立分類(lèi)模型對(duì)不同品系的水稻種子進(jìn)行鑒定。
圖3 3類(lèi)不同品系雜交稻種的PCA 3D得分圖
全波段光譜數(shù)據(jù)量大,建模復(fù)雜,因此,本文采用SPA算法提取特征波長(zhǎng)以減少建模變量,提高建模速度。將選擇的特征波長(zhǎng)數(shù)量范圍設(shè)置為5~20,共得到7個(gè)特征波長(zhǎng),分別為985.08、1 106、1 203.55、1 399.04、1 463.19、1 601.81、1 645.82 nm。近紅外光譜由分子內(nèi)部振動(dòng)光譜的倍頻與合頻產(chǎn)生,包含多數(shù)有機(jī)物的分子結(jié)構(gòu)和組成信息,能夠反映組成分子的含氫基團(tuán)X-H(X為N、O、C、S等)的振動(dòng)[28]。本文提取的特征波長(zhǎng)與分子官能團(tuán)中的N-H基團(tuán)(1 000 nm及1 400~1 800 nm附近[29],)、C-H基團(tuán)(1 050~1 200 nm附近[29];1 300~1 500 nm附近[29])及NH3+基團(tuán)(1 400 nm附近[30])的振動(dòng)較為接近,表明所選特征波長(zhǎng)具有一定的代表性,可用于建立有效可靠的判別分析模型。
基于全波段的光譜及特征波長(zhǎng)建立PLS-DA和SVM判別分析模型,并以識(shí)別正確率作為模型性能的評(píng)價(jià)指標(biāo),模型的判別結(jié)果如表2所示。
表2 基于全譜及特征波長(zhǎng)的PLS-DA和SVM判別結(jié)果
注:PLS-DA模型的參變量是隱含變量(LVs)個(gè)數(shù);SVM模型的參變量是懲罰系數(shù)和松弛系數(shù),表示為(,)。
Note: PLS-DA model’s parameter means the optimal number of LVs; SVM model’s parameter means different penalty parameters () and kernel function parameters (),shown as (,).
由表2可知,基于光譜數(shù)據(jù)的判別分析模型識(shí)別效果較好,其中基于全波段光譜的SVM模型識(shí)別效果最佳,建模集和預(yù)測(cè)集的識(shí)別率達(dá)到了99.67%和97.11%。對(duì)2種判別分析方法進(jìn)行比較發(fā)現(xiàn),SVM的判別效果優(yōu)于PLS-DA,這可能是因?yàn)镾VM模型采用徑向基函數(shù)(RBF)作為核函數(shù),并在尋優(yōu)范圍內(nèi)進(jìn)行網(wǎng)格搜索,能獲取全局最優(yōu)(,)參數(shù)組合且泛化能力強(qiáng)。對(duì)比分析基于全譜和特征波長(zhǎng)的判別模型可知,在采用特征波長(zhǎng)建立判別模型后,模型的判別效果有所下降。但基于特征波長(zhǎng)的分類(lèi)模型識(shí)別率均在90%以上,說(shuō)明選擇的特征波長(zhǎng)有效可靠,這為水稻品系的在線檢測(cè)提供了參考依據(jù)。
結(jié)果表明,采用近紅外高光譜技術(shù)結(jié)合化學(xué)計(jì)量學(xué)方法可以快速有效的識(shí)別不同品系的水稻種子,尤其SVM模型識(shí)別效果較好。
高光譜圖像能夠同時(shí)提供樣本的光譜信息和空間信息,且二者具有一定的對(duì)應(yīng)關(guān)系,因此,基于樣本平均光譜及對(duì)應(yīng)類(lèi)別值建立校正模型,可以用于待測(cè)樣本的類(lèi)別預(yù)測(cè)。而將此模型與圖像處理技術(shù)相結(jié)合,能生成樣本類(lèi)別預(yù)測(cè)偽彩圖,用不同顏色區(qū)分不同的樣本,實(shí)現(xiàn)類(lèi)別判定的可視化。由于全波段的光譜數(shù)據(jù)量大,計(jì)算復(fù)雜,不利于樣本的快速預(yù)測(cè)。因此,本文選擇基于SPA算法提取的特征波長(zhǎng)建立的SVM模型作為校正模型,將每粒水稻種子的平均光譜作為輸入,選取3類(lèi)不同品系的共502粒水稻進(jìn)行可視化預(yù)測(cè),結(jié)果如圖4所示。
圖4 不同品系雜交稻種的可視化預(yù)測(cè)圖
對(duì)比分析上圖可知,深兩優(yōu)862被判別為藍(lán)色,內(nèi)2優(yōu)6號(hào)被判別為黃色,和兩優(yōu)713被判別為紅色, 雖然有部分錯(cuò)判,但總體判別正確率為96.8%,判別結(jié)果較好,表明高光譜可視化偽彩圖可以直觀準(zhǔn)確的鑒別3類(lèi)不同品系的水稻種子。受高光譜圖像分割算法以及圖像分辨率的影響,可視化圖中水稻種子發(fā)生一定的形變,但整體的外形特征大致維持原樣,不影響鑒別分析??偟膩?lái)說(shuō),高光譜可視化圖可以實(shí)現(xiàn)單粒水稻種子的鑒別和定位,為農(nóng)業(yè)育種中種子的快速精確篩選提供了新的方法。
本研究采用近紅外高光譜成像技術(shù)鑒別3類(lèi)不同品系的水稻種子。選取1 800粒水稻種子作為建模集,900粒種子作為預(yù)測(cè)集,采用SPA算法選取7個(gè)特征波長(zhǎng),建立基于全波段以及特征波長(zhǎng)的PLS-DA和SVM分類(lèi)模型,其中基于全波段的SVM模型判別效果最佳,建模集和預(yù)測(cè)集的識(shí)別率分別為99.67%和97.11%。利用SPA-SVM模型結(jié)合圖像處理技術(shù)生成類(lèi)別預(yù)測(cè)偽彩圖,鑒別不同品系的水稻種子。結(jié)果表明,近紅外高光譜成像技術(shù)結(jié)合化學(xué)計(jì)量學(xué)方法以及圖像處理技術(shù)可以實(shí)現(xiàn)不同品系水稻種子的可視化預(yù)測(cè),為農(nóng)業(yè)育種中種子的快速篩選提供思路和幫助。
[1] 程本義,夏俊輝,沈偉峰. 長(zhǎng)江中下游稻區(qū)秈型兩系與三系雜交水稻組合的比較分析[J]. 雜交水稻,2009,24(5):61-65.
Cheng Benyi, Xia Junhui, Shen Weifeng. Comparison analysis between two-line and three-line indica hybrid rice combinations in the middle and down area of the Yangtse Valley of China[J]. Hybrid Rice, 2009, 24(5): 61-65. (in Chinese with English abstract)
[2] 吳翔,張衛(wèi)正,陸江鋒,等. 基于高光譜技術(shù)的玉米種子可視化鑒別研究[J]. 光譜學(xué)與光譜分析,2016,36(2):511-514.
Wu Xiang, Zhang Weizheng, Lu Jiangfeng, et al. Study on visual identification of corn seeds based on hyperspectral imaging technology[J]. Spectroscopy and Spectral Analysis, 2016, 36(2): 511-514. (in Chinese with English abstract)
[3] 柴玉華,畢文佳,譚克竹,等. 基于高光譜圖像技術(shù)的大豆品種無(wú)損鑒別[J]. 東北農(nóng)業(yè)大學(xué)學(xué)報(bào),2016,47(3):86-93.
Chai Yuhua, Bi Wenjia, Tan Kezhu, et al. Nondestructive identification of soybean seed varieties based on hyperspectral image technology[J]. Journal of Northeast Agricultural University, 2016, 47(3): 86-93. (in Chinese with English abstract)
[4] Shrestha S, Knapi? M, ?ibrat U, et al. Single seed near-infrared hyperspectral imaging in determining tomato (L) seed quality in association with multivariate data analysis[J]. Sensors & Actuators B Chemical, 2016, 237: 1027-1034.
[5] Huang M, Tang J, Yang B, et al. Classification of maize seeds of different years based on hyperspectral imaging and model updating[J]. Computers & Electronics in Agriculture, 2016, 122: 139-145.
[6] Yang S, Zhu Q B, Huang M, et al. Hyperspectral image-based variety discrimination of maize seeds by using a multi-model strategy coupled with unsupervised joint skewness-based wavelength selection algorithm[J]. Food Analytical Methods, 2017, 10(2): 424-433.
[7] Williams P J, Kucheryavskiy S. Classification of maize kernels using nir hyperspectral imaging[J]. Food Chemistry, 2016, 209: 131-138.
[8] Zhang X, Fei L, Yong H, et al. Application of hyperspectral imaging and chemometric calibrations for variety discrimination of maize seeds[J]. Sensors, 2012, 12(12): 17234-17246.
[9] Kong W, Zhang C, Liu F, et al. Rice seed cultivar identification using near-infrared hyperspectral imaging and multivariate data analysis[J]. Sensors, 2013, 13(7): 8916-8927.
[10] 何勇. 光譜及成像技術(shù)在農(nóng)業(yè)中的應(yīng)用[M]. 北京:科學(xué)出版社,2015.
[11] 張保華,李江波,樊書(shū)祥,等. 高光譜成像技術(shù)在果蔬品質(zhì)與安全無(wú)損檢測(cè)中的原理及應(yīng)用[J]. 光譜學(xué)與光譜分析,2014,34(10):2743-2751.
Zhang Baohua, Li Jiangbo, Fan Shuxiang, et a1. Principles and applications of hyperspectral imaging technique in quality and safety inspection of fruits and vegetables[J]. Spectroscopy and Spectral Analysis, 2014, 34(10): 2743-2751. (in Chinese with English abstract)
[12] 劉民法,張令標(biāo),何建國(guó),等. 基于高光譜成像技術(shù)的長(zhǎng)棗表面農(nóng)藥殘留無(wú)損檢測(cè)[J]. 食品與機(jī)械,2014,30(5):87-92.
Liu Minfa, Zhang Lingbiao, He Jianguo, et a1. Study on non-destructive detection of pesticide residues on Lingwu long jujubes’ surface using hyperspectral imaging[J]. Food & Machinery, 2014, 30(5): 87-92. (in Chinese with English abstract).
[13] Wu D, Sun D W. Advanced applications of hyperspectral imaging technology for food quality and safety analysis and assessment: A review — Part II: Applications[J]. Innovative Food Science & Emerging Technologies, 2013, 19(1): 15-28.
[14] 于雷,洪永勝,周勇,等. 高光譜估算土壤有機(jī)質(zhì)含量的波長(zhǎng)變量篩選方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(13):95-102.
Yu Lei, Hong Yongsheng, Zhou Yong, et al. Wavelength variable selection methods for estimation of soil organic matter content using hyperspectral technique[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(13): 95-102. (in Chinese with English abstract).
[15] 馮雷,陳雙雙,馮斌,等. 基于光譜技術(shù)的大豆豆莢炭疽病早期鑒別方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(1):139-144.
Feng Lei, Chen Shuangshuang, Feng Bin, et al. Early detection of soybean pod anthracnose based on spectrum technology[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(1): 139-144. (in Chinese with English abstract).
[16] Fan S, Huang W, Guo Z, et al. Prediction of soluble solids content and firmness of pears using hyperspectral reflectance imaging[J]. Food Analytical Methods, 2015, 8(8): 1936-1946.
[17] Skrobot V L, Castro E V R, Pereira R C C, et al. Use of principal component analysis (PCA) and linear discriminant analysis (LDA) in gas chromatographic (GC) data in the investigation of gasoline adulteration[J]. Energy & Fuels, 2016, 21(6): 5-19.
[18] Jung Y M. Principal component analysis based two-dimensional (PCA-2D) correlation spectroscopy: PCA denoising for 2D correlation spectroscopy[J]. Bulletin of the Korean Chemical Society, 2003, 24(9): 1345-1350.
[19] Ballabio D, Consonni V. Classification tools in chemistry. Part 1: linear models. PLS-DA[J]. Analytical Methods, 2013, 5(16): 3790-3798.
[20] Silva V A G D, Talhavini M, Peixoto I C F, et al. Non-destructive identification of different types and brands of blue pen inks in cursive handwriting by visible spectroscopy and PLS-DA for forensic analysis[J]. Microchemical Journal, 2014, 116: 235-243.
[21] Grasel F S, Ferr?o M F. A rapid and non-invasive method for the classification of natural tannin extracts by near-infrared spectroscopy and PLS-DA[J]. Analytical Methods, 2016, 8(3): 644-649.
[22] Hobro A J, Kuligowski J, D?ll M, et al. Differentiation of walnut wood species and steam treatment using ATR-FTIR and partial least squares discriminant analysis (PLS-DA) [J]. Analytical & Bioanalytical Chemistry, 2010, 398(6): 2713-2722.
[23] 張初,劉飛,孔汶汶,等. 利用近紅外高光譜圖像技術(shù)快速鑒別西瓜種子品種[J]. 農(nóng)業(yè)工程學(xué)報(bào),2013,29(20):270-277.
Zhang Chu, Liu Fei, Kong Wenwen, et al. Fast identification of watermelon seed variety using near infrared hyperspectral imaging technology[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2013, 29(20): 270-277. (in Chinese with English abstract)
[24] 程術(shù)希,孔汶汶,張初,等. 高光譜與機(jī)器學(xué)習(xí)相結(jié)合的大白菜種子品種鑒別研究[J]. 光譜學(xué)與光譜分析,2014,34(9):2519-2522.
Cheng Shuxi, Kong Wenwen, Zhang Chu, et al. Variety recognition of Chinese cabbage seeds by hyperspectral imaging combined with machine learning[J]. Spectroscopy and Spectral Analysis, 2014, 34(9): 2519-2522. (in Chinese with English abstract).
[25] Liu C, Liu W, Lu X, et al. Nondestructive determination of transgenicrice seeds (L.) using multispectral imaging and chemometric methods[J]. Food Chemistry, 2014, 153(12): 87-93.
[26] Tao D, Xiao Z, Zhang F, et al. Cloth defect classification method based on SVM[J]. International Journal of Digital Content Technology & Its Applications, 2013, 7(3): 614-622.
[27] Feng X, Zhao Y, Zhang C, et al. Discrimination of transgenic maize kernel using NIR hyperspectral imaging and multivariate data analysis[J]. Sensors, 2017, 17(8): 1894.
[28] Fearn T. Standardisation and calibration transfer for near infrared instruments: A review[J]. Journal of Near Infrared Spectroscopy, 2001, 9(1): 229-244.
[29] 陶琳麗,黃偉,楊秀娟,等. 20種氨基酸近紅外光譜及其分子結(jié)構(gòu)的相關(guān)性[J]. 光譜學(xué)與光譜分析,2016,36(9):2766-2773.
Tao Linli, Huang Wei, Yang Xiujuan, et al. Correlations between near infrared spectra and molecular structures of 20 standard amino acids[J]. Spectroscopy and Spectral Analysis, 2016, 36(9): 2766-2773. (in Chinese with English abstract)
[30] Selvaraju K, Kirubavathi K, Kumararaman S. Growth and characterization of a new semi-organic nonlinear optical crystal: Thiosemicarbazide cadmium acetate[J]. Journal of Minerals & Materials Characterization & Engineering, 2012, 11(3): 303-310.
劉小丹,馮旭萍,劉 飛,何 勇.基于近紅外高光譜成像技術(shù)鑒別雜交稻品系[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017,33(22):189-194. doi:10.11975/j.issn.1002-6819.2017.22.024 http://www.tcsae.org
Liu Xiaodan, Feng Xuping, Liu Fei, He Yong. Identification of hybrid rice strain based on near-infrared hyperspectral imaging technology[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(22): 189-194. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2017.22.024 http://www.tcsae.org
Identification of hybrid rice strain based on near-infrared hyperspectral imaging technology
Liu Xiaodan, Feng Xuping, Liu Fei, He Yong※
(,,310058,)
The selection and identification of seeds are a key link in the process of agricultural breeding. In this study, near infrared (874-1 734 nm) hyperspectral imaging technology combined with chemometrics and image processing technology was successfully performed to identify and visualize strains of hybrid rice seeds. A total of 2 700 samples of 3 different strains of rice seeds were collected, and all samples were divided into the calibration set and the prediction set according to the ratio of 2:1 using the SPXY algorithm. PCA (principle component analysis) was applied to explore the separability of different rice seeds based on the spectral characteristics of rice samples, and the preliminary results demonstrated that hybrid rice seeds of 3 different strains showed a trend of classification. The full spectrum has a large data volume, and contains a large amount of redundant and collinear information, which would affect the accuracy and calculation speed of the model. Since the optimal wavelength selection can help to extract important information from the whole data to improve the performance of the model while simplifying it, we adopted SPA (successive projections algorithm) to select sensitive wavelengths. Seven sensitive wavelengths (985.08, 1 106, 1 203.55, 1 399.04, 1 463.19, 1 601.81, 1 645.82 nm) were determined from the range of 975-1 646 nm, and these wavelengths were related to functional groups in molecules (N-H, C-H, NH3+), which indicated the reliability of the selected wavelength for modeling. Partial least squares-discriminant analysis (PLS-DA) and support vector machine (SVM) were applied to build the classification models based on the full spectra and optimal wavelengths, and an excellent classification was achieved, with the classification accuracy of over 90% for all models. The SVM model performed better than PLS-DA, and especially the full spectrum-based SVM model achieved outstanding identification results, with 99.67% classification accuracy for calibration set and 97.11% for prediction set. Compared with full spectrum-based models, optimal wavelengths-based models performed relatively worse, but still offered correct discrimination rates of over 90.22%. This results revealed that the selected wavelength is effective and reliable, which can provide a reference for on-line discrimination of different strains of hybrid rice seeds. Combined with image processing technology,the visual prediction map could be generated by inputting the average spectra of each rice seed into the SPA-SVM model, and different colors would be employed to represent different kinds of seeds. It showed that the visual analysis of the sample could intuitively identify rice seeds of different strains by these methods. The overall results indicated that near infrared hyperspectral imaging technology can be used to identify and visually predict hybrid rice seeds. This research provides a new way for rapid screening and identification of seeds in the process of agricultural breeding.
image processing; spectral analysis; nondestructive detection; hyperspectral imaging; rice seed; SPA
10.11975/j.issn.1002-6819.2017.22.024
TP391.4
A
1002-6819(2017)-22-0189-06
2017-06-29
2017-09-20
國(guó)家重大儀器設(shè)備開(kāi)發(fā)專(zhuān)項(xiàng)(2014YQ470377),國(guó)家十三五重點(diǎn)研發(fā)計(jì)劃(2016YFD0200603)
劉小丹,女,博士生,研究方向?yàn)閿?shù)字農(nóng)業(yè)信息獲取與檢測(cè)技術(shù)的研究。Email:xdlww@zju.edu.cn
何 勇,男,教授,博士生導(dǎo)師,研究方向?yàn)閿?shù)字農(nóng)業(yè),3S技術(shù)與農(nóng)業(yè)物聯(lián)網(wǎng)等方面研究。Email:yhe@zju.edu.cn