柴宏紅,邵 科,于 超,邵金旺,王瑞利,隨 洋,白 凱,劉云玲,馬韞韜
基于三維點(diǎn)云的甜菜根表型參數(shù)提取與根型判別
柴宏紅1,邵 科2,于 超2,邵金旺2,王瑞利2,隨 洋2,白 凱3,劉云玲1,馬韞韜1※
(1. 中國(guó)農(nóng)業(yè)大學(xué)土地科學(xué)與技術(shù)學(xué)院農(nóng)業(yè)部華北耕地保育重點(diǎn)實(shí)驗(yàn)室,北京 100193;2. 內(nèi)蒙古自治區(qū)特色植物分子生物學(xué)重點(diǎn)實(shí)驗(yàn)室,內(nèi)蒙古自治區(qū)生物技術(shù)研究院,呼和浩特 010070;3. 內(nèi)蒙古自治區(qū)呼倫貝爾生態(tài)環(huán)境監(jiān)測(cè)站,呼倫貝爾 021008)
為挑選產(chǎn)糖量高且適合機(jī)械化收獲的甜菜根型,該文基于多視角圖像序列,構(gòu)建了207個(gè)基因型甜菜根的三維點(diǎn)云模型?;谌S點(diǎn)云提取了描述甜菜根形態(tài)特征的10個(gè)表型參數(shù):最大直徑、根長(zhǎng)、凸包體積、頂投影面積、緊湊度、凸起率、凸起角、根頭比、根尾比和根體漸細(xì)指數(shù)。與人工測(cè)定的最大直徑和根長(zhǎng)值進(jìn)行校驗(yàn),決定系數(shù)R均在0.95以上。其中根長(zhǎng)、凸包體積及頂投影面積與生產(chǎn)指標(biāo)呈極顯著(<0.01)相關(guān)關(guān)系。采用穩(wěn)定性較高的-medoids聚類算法將甜菜根型分為4類,結(jié)合專家知識(shí)獲取理想根構(gòu)型的主要特征為根型中等長(zhǎng)度、比例適中。采用線性判別、隨機(jī)森林、支持向量機(jī)、決策樹和樸素貝葉斯5種預(yù)測(cè)模型進(jìn)行根型判別。結(jié)果表明5種根系判別模型預(yù)測(cè)準(zhǔn)確率均在70.0%以上,隨機(jī)森林判別準(zhǔn)確率達(dá)到81.4%。研究結(jié)果將為培育高品質(zhì)和適應(yīng)機(jī)械化生產(chǎn)的甜菜品種提供依據(jù)。
圖像處理;機(jī)器學(xué)習(xí);三維點(diǎn)云;甜菜;根型;表型;分類
甜菜(L.)是中國(guó)制糖工業(yè)的重要原料之一,其產(chǎn)糖量約占全國(guó)食糖總產(chǎn)量的10%[1-2]。近年來(lái),隨著機(jī)械化程度及配套種植技術(shù)的逐步提高,甜菜在內(nèi)蒙古、新疆和黑龍江的種植面積不斷擴(kuò)大。發(fā)展甜菜產(chǎn)業(yè)對(duì)促進(jìn)邊疆和少數(shù)民族地區(qū)脫貧攻堅(jiān)有重要的推動(dòng)作用[3]。甜菜是以收獲塊根并從其中榨取糖分的經(jīng)濟(jì)作物。隨著甜菜種植業(yè)從人工種植步入機(jī)械化時(shí)代,如何篩選適合當(dāng)前規(guī)?;N植條件下大型機(jī)械收獲的甜菜根型成為新階段的首要任務(wù)。各種高通量傳感器和多源圖像處理技術(shù)的相繼推出,極大地提高了表型數(shù)據(jù)的獲取效率,保證了數(shù)據(jù)的客觀準(zhǔn)確性[4-5]。
基于圖像分析技術(shù)可以快速準(zhǔn)確地確定果實(shí)[6-7]、葉子[8-9]和根系形狀[10-11]參數(shù)。Nankar等[6]通過(guò)構(gòu)建番茄果實(shí)表型性狀對(duì)應(yīng)的數(shù)學(xué)模型,開發(fā)了可以半自動(dòng)精確測(cè)量番茄果實(shí)表型性狀的分析系統(tǒng)。吳正敏等[9]提出基于圖像自動(dòng)提取茶葉形態(tài)特征參數(shù)的方法,進(jìn)一步提高了茶葉分選精度??讖埖萚10]基于圖像提取馬鈴薯面積和周長(zhǎng),采用矩形度、圓形度、偏心率和不變矩等方法進(jìn)行形狀分類,可以對(duì)馬鈴薯質(zhì)量進(jìn)行初步篩選。Tsialtas等[2]基于圖像分析了不同地點(diǎn)和年型間6個(gè)甜菜品種根型的差異,初步建立了根形參數(shù)與產(chǎn)量和品質(zhì)之間的關(guān)系。
基于作物三維結(jié)構(gòu)的植物表型分析進(jìn)一步擴(kuò)大了表型數(shù)據(jù)的維度,可以直接進(jìn)行植物表型參數(shù)的精準(zhǔn)提取,是建立植物表型-基因型研究最直接的橋梁[4]。研究人員已經(jīng)基于多視角圖像序列對(duì)溫室內(nèi)黃瓜、茄子、青椒等進(jìn)行了三維重建,提取的表型參數(shù)精度較高[12]。Zhu等[13]將該方法用于對(duì)大田玉米、大豆等植株個(gè)體及群體生長(zhǎng)動(dòng)態(tài)的監(jiān)測(cè)。盡管大田作物間遮擋嚴(yán)重,但通過(guò)去除待測(cè)植物周邊的遮擋,三維重建效果較好。Mortensen等[14]基于田間萵苣植株三維點(diǎn)云進(jìn)行植株分割及預(yù)測(cè)生物量,利用提取的植株表面積建立多種生物量預(yù)測(cè)模型,預(yù)測(cè)精度為84.0~94.0%。但基于三維模型對(duì)不同基因型甜菜根型表型的研究和精準(zhǔn)分類預(yù)測(cè)目前還未見報(bào)道。
應(yīng)用機(jī)器視覺技術(shù),采用三維重建的方法對(duì)甜菜根型進(jìn)行表型數(shù)字化處理與自動(dòng)分類是重新定義甜菜根型并進(jìn)行精準(zhǔn)篩選的需要,也是甜菜從機(jī)械化農(nóng)業(yè)步入精準(zhǔn)農(nóng)業(yè)階段的需要。因此,本文基于多視角圖像序列,構(gòu)建了207個(gè)不同基因型甜菜根的三維點(diǎn)云模型,開發(fā)了自動(dòng)提取表型參數(shù)的程序。采用-medoids算法實(shí)現(xiàn)了對(duì)甜菜根型的精準(zhǔn)分類。通過(guò)比較不同分組的表型參數(shù)差異,初步篩選出適合機(jī)械化種植的甜菜理想根型。以專家對(duì)甜菜根型調(diào)整后的分類結(jié)果為測(cè)試集,基于線性判別、隨機(jī)森林、支持向量機(jī)、決策樹、樸素貝葉斯5種模型對(duì)根型進(jìn)行精準(zhǔn)預(yù)測(cè)。以期用于今后甜菜根型的分類判別,快速篩選適合機(jī)械化種植的甜菜根型。
田間試驗(yàn)在內(nèi)蒙古生物院涼城試驗(yàn)基地(112.28E,40.29N)進(jìn)行,該地屬于中溫帶半干旱大陸性季風(fēng)氣候。年均氣溫為2~5 ℃,年日照時(shí)數(shù)3 026 h,有效積溫2 600 ℃。甜菜栽培土層為0~30 cm,土壤有機(jī)質(zhì)質(zhì)量分?jǐn)?shù)1.80%,全氮量1.03 g/kg,有效磷23.09 mg/kg。試驗(yàn)材料為207個(gè)不同基因型的甜菜根。其中73個(gè)是國(guó)內(nèi)品種,來(lái)源于新疆農(nóng)業(yè)科學(xué)院、內(nèi)蒙古自治區(qū)生物技術(shù)研究院、內(nèi)蒙古農(nóng)業(yè)大學(xué)和中國(guó)農(nóng)業(yè)科學(xué)院甜菜研究所,82個(gè)來(lái)源于荷蘭、20個(gè)來(lái)源于英國(guó)、30個(gè)來(lái)源于德國(guó)、2個(gè)源于俄羅斯。種植方式為機(jī)械覆膜打孔人工點(diǎn)播,行距40 cm,株距25 cm。甜菜根在出苗后140 d采用人工挖掘收獲,并測(cè)量甜菜根生物量和含糖率。采用iPhone 8 plus手機(jī)獲取甜菜根多視角圖像序列,相機(jī)鏡頭位置至甜菜根的距離不固定,以能拍攝到甜菜根清晰圖像為準(zhǔn)。本研究中相機(jī)鏡頭保持在距甜菜根50 cm左右范圍內(nèi),每張圖像角度間隔8°,環(huán)繞拍攝2圈,以保證圖像間的重疊度和重建三維點(diǎn)云的密度。手機(jī)像素為4 032×3 024,焦距為3.99 mm,快門速度為0.01 s,光圈數(shù)為/1.8,ISO感光度為32。每個(gè)甜菜根獲取的圖像數(shù)量在100~120之間。
1.2.1 三維點(diǎn)云重建
基于獲取的甜菜根多視角圖像序列(圖1a),采用3DF Zephyr Aerial (worldwide) - Version 4.353 (3DF Zephyr,https://www.3dflow.net/3df-zephyr-aerial-download-page/)重建拍攝場(chǎng)景的三維點(diǎn)云(圖1b),主要包括以下步驟:導(dǎo)入一個(gè)甜菜根對(duì)應(yīng)的多視角圖像序列;軟件自動(dòng)從這組圖像恢復(fù)其位置和方向;基于多視角立體成像技術(shù)提取密集且精確的三維點(diǎn)云;導(dǎo)出三維點(diǎn)云。三維重建過(guò)程相機(jī)定向預(yù)設(shè)類別選擇近景,其他參數(shù)均為默認(rèn)值。三維點(diǎn)云經(jīng)過(guò)去噪、旋轉(zhuǎn)、分離等系列處理后,量化甜菜根各表型參數(shù)值用于進(jìn)一步的分類。
圖1 基于多視角圖像序列量化甜菜根的流程圖
1.2.2 三維點(diǎn)云預(yù)處理
表型參數(shù)提取前需要先對(duì)三維點(diǎn)云進(jìn)行預(yù)處理,主要包括點(diǎn)云去噪,將三維點(diǎn)云旋正及將甜菜根與桌面分割三部分。由于拍照時(shí)的復(fù)雜背景及拍照時(shí)手的抖動(dòng)均會(huì)給點(diǎn)云數(shù)據(jù)帶來(lái)噪點(diǎn),故先在3DF Zephyr中手動(dòng)去除噪點(diǎn)。拍攝時(shí)由于手機(jī)位置不固定,采用MATLAB讀入的三維點(diǎn)云處于傾斜狀態(tài)。提取表型參數(shù)前需要經(jīng)過(guò)三維坐標(biāo)位置變換,使甜菜根平行于平面,便于后期數(shù)據(jù)處理。在坐標(biāo)系中,桌面是一個(gè)光滑平面,采用平面擬合提取桌面并計(jì)算桌面法線,進(jìn)而推導(dǎo)出平面和平面的旋轉(zhuǎn)變換矩陣T(方程1)和T(方程2)[15]。將三維點(diǎn)云柵格化為平面的深度圖像,每個(gè)圖像像素代表點(diǎn)云中0.02í0.02網(wǎng)格的最大深度(值)。對(duì)深度圖像二值化,利用霍夫變換檢測(cè)桌邊緣直線并計(jì)算直線與軸正方向的夾角θ。根據(jù)θ可以推導(dǎo)出平面的旋轉(zhuǎn)矩陣T(方程3)。將傾斜的三維點(diǎn)云坐標(biāo)值乘以變換矩陣T,T和T得到旋正后的甜菜根三維點(diǎn)云。分離桌面和甜菜根,基于桌面的固定長(zhǎng)寬值進(jìn)行比例換算,獲得三維點(diǎn)云的實(shí)際坐標(biāo)值。
(1)
(2)
(3)
式中θ為桌面與空間坐標(biāo)系平面的夾角;θ為桌面與空間坐標(biāo)系平面的夾角;θ為桌面邊緣直線與空間坐標(biāo)系軸正方向的夾角,θ、θ、θ單位均為(°)。
1.2.3 基于三維點(diǎn)云的甜菜根表型參數(shù)提取
本研究綜合借鑒番茄[6]、馬鈴薯[16]、草莓[17-18]果實(shí)表型相關(guān)研究和甜菜專家的意見,基于預(yù)處理后的甜菜根三維點(diǎn)云提取了10個(gè)根表型參數(shù),如表1所示?;拘螒B(tài)特征包括根長(zhǎng)、最大直徑MD、凸包體積、頂投影面積以及各層段對(duì)應(yīng)的直徑。其中:根長(zhǎng)為根尾至根頭之間的距離,即三維坐標(biāo)系中最大值與最小值之差(如圖2);最大直徑MD為能包裹甜菜根最小圓柱體的直徑(如圖2);凸包體積為能包裹甜菜根最小多邊體的體積(如圖2);頂投影面積為根體在平面投影不規(guī)則形狀的面積(如圖2)。為量化甜菜根從根頭至根尾的各層段直徑,選擇0.1 cm為步長(zhǎng)進(jìn)行層切,計(jì)算各層點(diǎn)的最大距離,記為這一層的甜菜根直徑;找到最大直徑在甜菜根體上的位置,距離根頭的高度記為凸起高(如圖2)。為細(xì)化對(duì)甜菜根形態(tài)的數(shù)字描述,基于基本形態(tài)特征參數(shù)建立復(fù)雜根表型參數(shù)緊湊度、凸起率、凸起角、根頭比、根尾比、根體漸細(xì)指數(shù)[5,8]。
表1 甜菜根表型參數(shù)集
注:H為根長(zhǎng),cm;MD為最大直徑,cm;S為頂投影面積,cm2;h為最大直徑對(duì)應(yīng)的高度,cm;W1為距根頭5%高度的直徑,cm;W2為距根頭80%高度的直徑,cm;以最大直徑為分界線,W3為根下部平均直徑,cm,W4為根上部平均直徑,cm。
聚類分析采用-medoids算法,其工作流程是將所有表型參數(shù)的數(shù)據(jù)集標(biāo)準(zhǔn)化,確定聚類個(gè)數(shù),在所有數(shù)據(jù)集中選擇個(gè)聚族中心點(diǎn),計(jì)算其余點(diǎn)到這個(gè)中心點(diǎn)的距離,并把每個(gè)點(diǎn)到個(gè)中心點(diǎn)最短的聚簇作為自己所屬的聚簇。通過(guò)反復(fù)迭代計(jì)算,使得每個(gè)點(diǎn)都屬于離他最近的聚類中心對(duì)應(yīng)組且個(gè)中心點(diǎn)不再變化,最后確定最優(yōu)聚類結(jié)果[19-21]。
基于分類結(jié)果對(duì)表型參數(shù)采用LSD-t多重比較法檢驗(yàn)兩兩組間的顯著性差異,計(jì)算LSD-t值的計(jì)算公式(4)用字母標(biāo)記法表示差異顯著水平。
目前甜菜根型分類不明確,因此在本研究中甜菜專家基于視覺判別經(jīng)驗(yàn),結(jié)合聚類結(jié)果對(duì)各組離異值和臨近值進(jìn)行進(jìn)一步的調(diào)整。據(jù)此,采用調(diào)整后的分組作為根型分類真值,采用線性判別分析、隨機(jī)森林、支持向量機(jī)、決策樹和樸素貝葉斯判別5種方法對(duì)甜菜根型進(jìn)行分類建模與預(yù)測(cè)。以驗(yàn)證本試驗(yàn)提取的表型參數(shù)的可分性和聚類算法分類的實(shí)用性,并分析不同機(jī)器學(xué)習(xí)算法對(duì)復(fù)雜甜菜根樣本分類效果[22]。各類算法基本原理如下:
線性判別分析通過(guò)將高維空間的樣本投影到一維空間實(shí)現(xiàn)分類判斷,其優(yōu)勢(shì)在于對(duì)訓(xùn)練樣本分布、方差等均沒(méi)有限制,分類能力強(qiáng)大[23]。決策樹是一個(gè)樹結(jié)構(gòu),每個(gè)非葉節(jié)點(diǎn)表示一個(gè)特征屬性上的測(cè)試,每個(gè)分支代表特征屬性的輸出值,每個(gè)葉節(jié)點(diǎn)代表一個(gè)類別[9]。隨機(jī)森林是一種基于分類回歸樹的機(jī)器學(xué)習(xí)算法,將多種決策樹算法結(jié)合起來(lái),相較于傳統(tǒng)決策樹方法,其優(yōu)勢(shì)在于同等運(yùn)算率下具有更高的預(yù)測(cè)精度[24-25]。支持向量機(jī)是一種通用的前饋網(wǎng)絡(luò)類型,其主要影響因素是核函數(shù)的選擇和相應(yīng)參數(shù)的設(shè)置[8,26]。樸素貝葉斯采用先驗(yàn)概率來(lái)描述樣本特征,再用訓(xùn)練樣本來(lái)修正先驗(yàn)概率得到后驗(yàn)概率從而進(jìn)行統(tǒng)計(jì)推算[27-28]。
為了確定5種模型的最優(yōu)訓(xùn)練樣本數(shù)量,試驗(yàn)分別選取50、100、150、200個(gè)樣本的10個(gè)表型參數(shù)進(jìn)行訓(xùn)練,比較不同樣本數(shù)量的模型預(yù)測(cè)準(zhǔn)確率。結(jié)果表明線性判別分析、隨機(jī)森林、支持向量機(jī)、決策樹4種模型隨著訓(xùn)練樣本數(shù)增加,模型判別準(zhǔn)確率提高,而樸素貝葉斯在150個(gè)樣本數(shù)時(shí)判別準(zhǔn)確率已達(dá)到最大78.0%。因此,線性判別分析、隨機(jī)森林、支持向量機(jī)、決策樹4種模型選用80%數(shù)據(jù)作為訓(xùn)練集,20%數(shù)據(jù)作為測(cè)試集,而樸素貝葉斯模型選用70%數(shù)據(jù)作為訓(xùn)練集,30%數(shù)據(jù)作為測(cè)試集。5種模型均采用十折交叉驗(yàn)證[9],將數(shù)據(jù)集分成10份,輪流將其中9份用于訓(xùn)練,1份用于驗(yàn)證。
基于上述機(jī)器學(xué)習(xí)系統(tǒng)參數(shù)優(yōu)化過(guò)程獲得5種模型最優(yōu)訓(xùn)練測(cè)試數(shù)據(jù)集比例。根據(jù)模型輸出的混淆矩陣中真正(TP)、真負(fù)(TN)、假正(FP)、假負(fù)(FN)的樣本數(shù)量,利用公式(5)計(jì)算準(zhǔn)確率(Accuracy)、召回率(Recall)和調(diào)和平均值1(Harmonic average1)作為模型預(yù)測(cè)性能的評(píng)價(jià)指標(biāo)。
本文方差分析(ANOVA)、線性回歸、聚類分析及判別分析均基于R語(yǔ)言完成。模型計(jì)算值和測(cè)量值的吻合程度采用均方根誤差(Root Mean Square Error, RMSE)和決定系數(shù)(Coefficient of Determination,2)進(jìn)行描述:
式中y和x分別為第個(gè)計(jì)算值和測(cè)量值;為樣本個(gè)數(shù)。
基于多視角圖像序列重建的三維點(diǎn)云圖包含甜菜根相應(yīng)的顏色和紋理信息,具有較強(qiáng)的真實(shí)感?;谥亟ǖ娜S點(diǎn)云可用來(lái)提取甜菜根的各種表型參數(shù),甜菜根根長(zhǎng)和最大直徑計(jì)算值與測(cè)量值的比較如圖3所示。基于甜菜根三維模型計(jì)算的根長(zhǎng)和最大直徑與實(shí)測(cè)值的決定系數(shù)2均大于0.95,RMSE分別為1.78和2.68 mm。對(duì)比結(jié)果表明此方法能夠精確再現(xiàn)甜菜根的表型特征。
圖3 甜菜根根長(zhǎng)和最大直徑計(jì)算值與實(shí)測(cè)值的比較
表2為3個(gè)生產(chǎn)指標(biāo)和10個(gè)表型參數(shù)兩兩變量間的相關(guān)系數(shù)。最大直徑、頂投影面積及凸包體積間有極顯著相關(guān)性,相關(guān)系數(shù)均在0.90以上,表明三者間存在強(qiáng)共線性。凸起率與凸起角都是由凸起高計(jì)算得到,兩者間為0.91,存在強(qiáng)共線性。最大直徑、頂投影面積及凸包體積與生物量呈極顯著正相關(guān),均在0.85以上,與含糖率呈負(fù)相關(guān),為0.21~0.24。該結(jié)果與Tsialtas[2]對(duì)最大直徑與含糖率間相關(guān)關(guān)系的研究結(jié)論一致。生物量與含糖率間為0.25,結(jié)果與高妙真[29]提出的甜菜根在500 g以上時(shí),其生物量與含糖率呈負(fù)相關(guān)的結(jié)論一致。含糖量由生物量和含糖率之積求得,其中生物量有更高數(shù)量級(jí),與含糖量有極顯著相關(guān)關(guān)系,達(dá)0.97。含糖量與投影面積、凸包體積及最大直徑有正相關(guān)關(guān)系,均在0.80以上。
表2 表型參數(shù)與生產(chǎn)指標(biāo)相關(guān)分析
注:***表示相關(guān)性極顯著(<0.01), **表示相關(guān)性顯著(<0.05)。
Note: *** means that the correlation is extremely significant (<0.01), ** means that the correlation is significant (<0.05)。
以上結(jié)果表明,在10個(gè)表型參數(shù)中,對(duì)甜菜根生產(chǎn)指標(biāo)影響顯著的表型參數(shù)依次為凸包體積、頂投影面積、最大直徑、根長(zhǎng)和凸起率,其他表型參數(shù)對(duì)生產(chǎn)指標(biāo)影響均不顯著。本試驗(yàn)數(shù)據(jù)存在共線性和數(shù)量級(jí)差異,故在聚類分析之前需進(jìn)行主成分分析消除共線性,并對(duì)數(shù)據(jù)進(jìn)行歸一化。
傳統(tǒng)甜菜根根型依靠人為觀察根莖、根體上端和根體下端的寬窄來(lái)對(duì)根型進(jìn)行分類。這種分類方法要求研究人員具有一定的甜菜專業(yè)知識(shí),且分類標(biāo)準(zhǔn)不明確。本研究采用-medoids算法,基于計(jì)算的甜菜根表型數(shù)據(jù)[30-31]對(duì)甜菜根型進(jìn)行識(shí)別與客觀分類。圖4為對(duì)207個(gè)基因型甜菜根的分類結(jié)果。圖中由不同顏色和不同形狀分別劃分的4個(gè)多邊形代表4個(gè)不同的分類組。多邊形包圍的數(shù)字是207個(gè)基因型甜菜根材料序號(hào)。由于部分甜菜根的表型特征差異較細(xì)微,故組與組之間存在少量重疊。這4組分別記為組1(圓錐形)、組2(錘形)、組3(楔形)、組4(長(zhǎng)楔形),各組的成份數(shù)目依次為66、21、75、45。
在此分類結(jié)果基礎(chǔ)上對(duì)各組間在表型參數(shù)上的差異顯著性(<0.05)進(jìn)行檢驗(yàn),采用字母標(biāo)記法標(biāo)注組間差異顯著性強(qiáng)弱,結(jié)果如表3所示。在4組甜菜根中長(zhǎng)楔形組4根最長(zhǎng),最大直徑、凸包體積和頂投影面積均小于圓錐形組1,但顯著(<0.05)大于錘形組2和楔形組3,緊湊度、凸起率、根尾比及根體漸細(xì)指數(shù)均顯著(<0.05)小于其他組,表明長(zhǎng)楔形組4屬于根頭寬大,根尾細(xì)小且根體從根頭至根尾的縮小速率最快的根體。錘形組2根最短,緊湊度、根尾比和根體漸細(xì)指數(shù)均顯著(<0.05)大于其他組,表明錘形組2根體是短小形且根頭至根尾的縮小程度不大。楔形組3根長(zhǎng)顯著(<0.05)大于錘形組2而顯著(<0.05)小于圓錐形組1,最大直徑、凸包體積和投影面積是最小的,根頭比顯著(<0.05)高于其他組,且凸起率顯著(<0.05)低于其它組,表明楔形組3最大直徑接近根頭處。圓錐形組1最大直徑、凸包體積和頂投影面積最大而根頭比最小,圓錐形組1屬于根頭寬度較小,最大直徑位于根體中部,根體整個(gè)寬度的變化是先增加后減小的過(guò)程。
注:組代表分類序號(hào),數(shù)目代表各組成份數(shù)。
表3 甜菜根組間表型差異的統(tǒng)計(jì)分析
注:同一參數(shù)不同小寫字母表示在0.05水平差異顯著。
Note: Different lowercase letters for the same parameter indicate significant differences at the level of 0.05.
從生產(chǎn)指標(biāo)來(lái)看,圓錐形組1生物量和含糖量最大,顯著高于錘形組2和楔形組3,而楔形組3含糖率最高且顯著高于圓錐形組1。結(jié)合實(shí)際生產(chǎn)情況和甜菜專家育種知識(shí)及田間種植經(jīng)驗(yàn)分析,錘形組2過(guò)于短小,生物量小,長(zhǎng)楔形組4根長(zhǎng)太長(zhǎng),機(jī)械化生產(chǎn)過(guò)程中易斷。因此,圓錐形組1和楔形組3為甜菜理想根型,產(chǎn)糖量高,尺寸適宜,利于機(jī)械化收割。其中圓錐形組1更適合于內(nèi)蒙古自治區(qū)北部生育期較短的地區(qū),楔形組3更適于內(nèi)蒙古自治區(qū)中西部地區(qū),與文獻(xiàn)[2-3, 32]結(jié)果一致。
根據(jù)聚類分析和統(tǒng)計(jì)分析結(jié)果,在4組中分別挑選2個(gè)有代表性的甜菜根進(jìn)行展示,如圖5。這8個(gè)根所在坐標(biāo)系的刻度均為甜菜根的實(shí)際大小。
為快速準(zhǔn)確地挑選高品質(zhì)甜菜根,開發(fā)基于多視角圖像序列的判別模型。本研究采用5種機(jī)器學(xué)習(xí)方法對(duì)甜菜根型的判別結(jié)果如表4所示。線性判別、隨機(jī)森林和支持向量機(jī)3種判別模型精準(zhǔn)確率、召回率和調(diào)和平均數(shù)均在75.0%以上,能夠較好的用于甜菜根型判別預(yù)測(cè)。其中隨機(jī)森林判別效果最佳,判別準(zhǔn)確率為81.4%,決策樹模型判別效果最差,判別準(zhǔn)確率為70.0%,召回率和調(diào)和平均數(shù)分別為64.3%和65.5%。
圖5 4組根型中典型甜菜根的三維點(diǎn)云圖
表4 5種模型對(duì)甜菜根型的預(yù)測(cè)結(jié)果
決策樹法在每個(gè)屬性測(cè)試節(jié)點(diǎn)上都會(huì)產(chǎn)生分枝,由聚類分析圖可以看出本試驗(yàn)數(shù)據(jù)存在較多離異值,故采用決策樹算法判別準(zhǔn)確率較低,而隨機(jī)森林法能夠?qū)⒍喾N決策樹算法結(jié)合起來(lái),其基本思想是通過(guò)bootstrap 重采樣的方法在原始訓(xùn)練集中抽取多個(gè)樣本,對(duì)每個(gè)抽取出的樣本均進(jìn)行決策樹建模,最后通過(guò)多數(shù)投票法得到最終的預(yù)測(cè)結(jié)果,有效避免了決策樹采樣算法的缺點(diǎn)從而提高了判別的準(zhǔn)確率。機(jī)器學(xué)習(xí)過(guò)程中采用十折交叉驗(yàn)證方法選取每次循環(huán)的訓(xùn)練樣本集,能夠顯著提高模型的準(zhǔn)確率和普適性。
1)本文以207個(gè)基因型甜菜根為研究對(duì)象,采用手機(jī)獲取甜菜根多視角圖像序列并重建三維點(diǎn)云。
2)基于重建的三維點(diǎn)云提取了甜菜根根型的基本表型參數(shù),包括根最大直徑、根長(zhǎng)、凸包體積、頂投影面積和各層段直徑。在基本形態(tài)參數(shù)的基礎(chǔ)上提出復(fù)雜表型參數(shù),包含緊湊度、凸起率、凸起角、根頭比、根尾比和根體漸細(xì)指數(shù),其中根長(zhǎng)、頂投影面積及凸包體積與生產(chǎn)指標(biāo)有極顯著(<0.01)相關(guān)性。
3)采用-medoids算法將根型劃分為4組,結(jié)合專家知識(shí)及多年種植經(jīng)驗(yàn)分析得出圓錐形和楔形甜菜根產(chǎn)糖量高、根型長(zhǎng)度比例適中、適合機(jī)械化生產(chǎn),為理想根型。
4)基于最適訓(xùn)練樣本量,建立線性判別分析、隨機(jī)森林、支持向量機(jī)、決策樹和樸素貝葉斯5種根型判別模型,并進(jìn)行準(zhǔn)確率比較。隨機(jī)森林分類效果最佳,判別準(zhǔn)確率為81.4%;決策樹分類效果最低,判別準(zhǔn)確率為70.0%。該研究結(jié)果可用于基因型與表型的關(guān)聯(lián)分析,為培育高品質(zhì)和適應(yīng)機(jī)械化生產(chǎn)的甜菜品種提供依據(jù)。
[1]Geng G, Yang J. Sugar beet production and industry in China[J]. Sugar Tech, 2015, 17(1): 13-21.
[2]Tsialtas J T, Maslaris N. Sugar beet root shape and its relation with yield and quality[J]. Sugar Tech, 2010, 12(1): 47-52.
[3]陳藝文,李用財(cái),余凌羿,等. 中國(guó)三大主產(chǎn)區(qū)甜菜糖業(yè)發(fā)展分析[J]. 中國(guó)糖料,2017,39(4):74-76.
Chen Yiwen, Li Yongcai, Yu Lingyi, et al. Analysis on the development of sugar beet industry in three main producing areas in China[J]. Sugar Crops of China, 2017, 39(4): 74-76. (in Chinese with English abstract)
[4]Yang W N, Feng H, Zhang X H, et al. Crop phenomics and high-throughput phenotyping: Past decades, current challenges, and future perspectives[J]. Molecular Plant, 2020, 13(2): 187-214.
[5]Sankaran S, Khot L R, Espinoza C Z, et al. Low-altitude, high-resolution aerial imaging systems for row and field crop phenotyping: A review[J]. European Journal of Agronomy, 2015, 70: 112-123.
[6]Nankar A N, Tringovska I, Grozeva S, et al. Application of high-throughput phenotyping tool tomato analyzer to characterize balkan capsicum fruit diversity[J]. Scientia Horticulturae, 2020, 260: 108862-108874.
[7]Mohan V J, Gupta S, Thomas S, et al. Tomato fruits show wide phenomic diversity but fruit developmental genes show low genomic diversity[J]. Plos One, 2016, 11(4): e152907-e152931.
[8]Li M, An H, Angelovici R, et al. Topological data analysis as a morphometric method: Using persistent homology to demarcate a leaf morphospace[J]. Frontiers in Plant Science, 2018, 9: 553-567.
[9]吳正敏,曹成茂,王二銳,等. 基于形態(tài)特征參數(shù)的茶葉精選方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(11):315-321.
Wu Zhengmin, Cao Chengmao, Wang Errui, et al. Tea selection method based on morphology feature parameters[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(11): 315-321. (in Chinese with English abstract)
[10]孔彥龍,高曉陽(yáng),李紅玲,等. 基于機(jī)器視覺的馬鈴薯質(zhì)量和形狀分選方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(17):143-148.
Kong Yanlong, Gao Xiaoyang, Li Hongling, et al. Potato grading method of mass and shapes based on machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(17): 143-148. (in Chinese with English abstract)
[11]惠放,馬韞韜,朱晉宇,等. 利用多視角圖像法分析番茄幼苗根構(gòu)型對(duì)氮水平的響應(yīng)[J]. 植物營(yíng)養(yǎng)與肥料學(xué)報(bào),2016,22(5):1418-1424.
Hui Fang, Ma Yuntao, Zhu Jinyu, et al. Analysis of nitrogen effects on root system architecture of tomato seedlings using the multi-view images method[J]. Journal of Plant Nutrition and Fertilizer, 2016, 22(5): 1418-1424. (in Chinese with English abstract)
[12]Hui F, Zhu J Y, Hu P C, et al. Image-based dynamic quantification and high-accuracy 3D evaluation of canopy structure of plant populations[J]. Annals of Botany, 2018, 121(5): 1079-1088.
[13]Zhu B L, Liu F S, Xie Z W, et al. Quantification of light interception within image-based 3D reconstruction of sole and intercropped canopies over the entire growth season[J]. Annals of Botany, [2020-01-30], https://doi.org/10.1093/aob/ mcaa046.
[14]Mortensen A K, Bender A, Whelan B, et al. Segmentation of lettuce in coloured 3D point clouds for fresh weight estimation[J]. Computers and Electronics in Agriculture, 2018, 154: 373-381.
[15]Jiang Y, Li C Y, Takeda F, et al. 3D point cloud data to quantitatively characterize size and shape of shrub crops[J]. Hortic Res, 2019, 6(1): 1-17.
[16]Su Q H, Kondo N S, Li M Z, et al. Potato quality grading based on machine vision and 3D shape analysis[J]. Computers and Electronics in Agriculture, 2018, 152: 261-268.
[17]Oo L M, Aung N Z. A simple and efficient method for automatic strawberry shape and size estimation and classification[J]. Biosystems engineering, 2018, 170: 96-107.
[18]Xu L M, Zhao Y C. Automated strawberry grading system based on image processing[J]. Computers and Electronics in Agriculture, 2010, 71: S32-S39.
[19]Dhanachandra N, Manglem K, Chanu Y J. Image segmentation using K-means clustering algorithm and subtractive clustering algorithm[J]. Procedia Computer Science, 2015, 54: 764-771.
[20]Bhusare B B, Bansode S M. Centroids initialization for K-means clustering using improved pillar algorithm[J]. Int J Adv Res Comput Eng Technol (IJARCET), 2014, 3(4): 1317-1322.
[21]Yu D H, Liu G J, Guo M Z, et al. An improved K-medoids algorithm based on step increasing and optimizing medoids[J]. Expert Systems with Applications, 2018, 92: 464-473.
[22]Ozguven M M, Adem K. Automatic detection and classification of leaf spot disease in sugar beet using deep learning algorithms[J]. Physica A: Statistical Mechanics and its Applications, 2019, 535: 122537-122544.
[23]Gurubelli Y, Ramanathan M, Ponnusamy P. Fractional fuzzy 2DLDA approach for pomegranate fruit grade classification[J]. Computers and Electronics in Agriculture, 2019, 162: 95-105.
[24]Belgiu M, Dr?gu? L. Random forest in remote sensing: A review of applications and future directions[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2016, 114: 24-31.
[25]競(jìng)霞,白宗璠,高媛,等. 利用隨機(jī)森林法協(xié)同SIF和反射率光譜監(jiān)測(cè)小麥條銹病[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(13):154-161.
Jing Xia, Bai Zongfan, Gao Yuan, et al. Wheat stripe rust monitoring by random forest algorithm combined with SIF and reflectance spectrum[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(13): 154-161. (in Chinese with English abstract)
[26]潘春華,肖德琴,林探宇,等. 基于SVM和區(qū)域生長(zhǎng)結(jié)合算法的南方主要蔬菜害蟲分類識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(8):192-199.
Pan Chunhua, Xiao Deqin, Lin Tanyu, et. Classification and recognition for major vegetable pests in Southern China using SVM and region growing algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(8): 192-199. (in Chinese with English abstract)
[27]Atoui M A, Cohen A, Verron S, et al. A single Bayesian network classifier for monitoring with unknown classes[J]. Engineering Applications of Artificial Intelligence, 2019, 85: 681-690.
[28]翟治芬,徐哲,周新群. 基于樸素貝葉斯分類器的棉花盲椿象危害等級(jí)識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2015,31(1):204-211.
Zhai Zhifen, Xu Zhe, Zhou Xinqun. Recognition of hazard grade for cotton blind stinkbug based on Na?ve Bayesian classifier[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(1): 204-211. (in Chinese with English abstract)
[29]高妙真. 甜菜根重與含糖率相關(guān)關(guān)系的研究[J]. 中國(guó)甜菜,1995(1):16-19.
Gao Miaozhen. A study on the correlation between root weight and sugar content in sugar beet[J]. China Sugar beet, 1995(1): 16-19.(in Chinese with English abstract)
[30]Sun S P, Li C Y, Chee P W, et al. Three-dimensional photogrammetric mapping of cotton bolls in situ based on point cloud segmentation and clustering[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2020, 160: 195-207.
[31]Li M, Klein L L, Duncan K E, et al. Characterizing 3D inflorescence architecture in grapevine using X-ray imaging and advanced morphometrics: Implications for understanding cluster density[J]. Journal of Experimental Botany, 2019, 70(21): 6261-6276.
[32]Stevanato P, Chiodi C, Broccanello C, et al. Sustainability of the sugar beet crop[J]. Sugar Tech, 2019, 21(5): 703-716.
Extraction of phenotypic parameters and discrimination of beet root types based on 3D point cloud
Chai Honghong1, Shao Ke2, Yu Chao2, Shao Jinwang2, Wang Ruili2, Sui Yang2, Bai Kai3, Liu Yunling1, Ma Yuntao1※
(1,,,100193,; 2.010070,; 3.,021008,)
Sugar beet is one of the main crops for sugar production in the world, and originated from the western and southern coasts of Europe. Selecting and breeding of varieties of sugar beet based on plant phenotyping are the key factors for the development of sugar beet industry on a large-scale cultivation. In China, sugar beet was widely planted in arid and semi-arid regions, particularly for poverty alleviation of farmers living in border areas and ethnic minority areas. The type of beet root with great different genotypes directly determines the sugar yield and mechanization efficiency in modern agriculture. The traditional classification of beet root type depends mainly on manual separation, and thereby greatly limits industry production and breeding of the sugar beet due to heavy workload and relatively large errors. In order to meet the requirements of high-throughput analysis, a three-dimensional (3D) phenotyping technique with multi-view images was recently developed to facilitate the classification of fruit and vegetable with high accuracy and efficiency. In this study, the beet roots with 207 genotypes were selected as experimental materials. Multi-view images were obtained by moving mobile phone around beet root. Three-dimensional point clouds were reconstructed in 3DF Zephyr Aerial software, which can restore position and direction from a dataset of multi-views images to extract for the matching feature points between each pair of images. After the postprocessing of the matching images, including noise reduction, rotating and segment, the detailed features of beet root shape, color, and texture can be achieved in the 3D point cloud. Ten phenotypic parameters can be used to clarify the morphological characteristics of beet roots, the maximum diameter, root length, convex hull volume, top projection area, compactness, convex index, convex angle, distal root end ratio, proximal root end ratio and root taper index. There was a good agreement between the measured maximum diameter and root length, with coefficient of determinationR> 0.95. The-medoids clustering algorithm with high stability was selected to classify the beet root into four groups. Group 1, namely as cone beet root, indicates that the maximum root diameter located at the middle of the root body. Group 2, namely as hammer beet root, shows the shortest body of root, the smallest root head ration while larger root tail ration. Group 3, namely as wedge beet root, has the maximum diameter of root body close to the root head, whereas, the width of root from head to tail gradually decreased. Group 4, namely as long wedge beet root, has longer root body than that in group 3, wider root head and smaller root tail. The reduction rate of root body from head to tail was the greatest. Based on the combination of phenotypic traits and experts’ knowledge, Group 1 (cone beet root) and Group 3 (wedge beet root) were recommended due to their high sugar yield, medium root length and moderate proportion. After adjusting the categories by the experts as the true values, five prediction models were established to discriminate beet root type, including linear discrimination, random forest, support vector machine, decision tree, and naive Bayes. The results showed that the prediction accuracies of the five models were above 70.0%, where accuracy of random forest reached 81.4%. These results demonstrated that 3D point cloud reconstructed by multi-view image sequences can be used for the identification of beet root shape, and thereby to effectively improve the yield prediction of sugar beet and the selection of high-quality beet varieties. Since 207 genotypes have been selected for the classification of root types during this time, much more genotypes at different environments can be expected to enrich the 3D phenotyping library, and thereby further improve the accuracy of classification. This finding can provide a potential practical basis for the beet root type screening and breeding.
image procession; machine learning; three dimensional point cloud; beet; root type; phenotype; classification
柴宏紅,邵科,于超,等. 基于三維點(diǎn)云的甜菜根表型參數(shù)提取與根型判別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2020,36(10):181-188.doi:10.11975/j.issn.1002-6819.2020.10.022 http://www.tcsae.org
Chai Honghong, Shao Ke, Yu Chao, et al. Extraction of phenotypic parameters and discrimination of beet root types based on 3D point cloud[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(10): 181-188. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2020.10.022 http://www.tcsae.org
2020-02-15
2020-04-21
內(nèi)蒙古自治區(qū)科技重大專項(xiàng)和科技成果重大轉(zhuǎn)化項(xiàng)目
柴宏紅,博士生,主要研究方向?yàn)樽魑锶S表型研究。Email:honghong.chai@cau.edu.cn
馬韞韜,副教授,博士生導(dǎo)師,主要研究方向?yàn)橹参锕δ?結(jié)構(gòu)-環(huán)境互作的模型研究。Email:yuntao.ma@cau.edu.cn
10.11975/j.issn.1002-6819.2020.10.022
TP391.4
A
1002-6819(2020)-10-0181-08