管賢平,劉 寬,邱白晶,董曉婭,薛新宇
基于機(jī)載三維激光掃描的大豆冠層幾何參數(shù)提取
管賢平1,劉 寬1,邱白晶1※,董曉婭1,薛新宇2
(1. 江蘇大學(xué)農(nóng)業(yè)農(nóng)村部植保工程重點(diǎn)實(shí)驗(yàn)室,鎮(zhèn)江 212013; 2. 農(nóng)業(yè)農(nóng)村部南京農(nóng)業(yè)機(jī)械化研究所,南京 210014)
為了實(shí)現(xiàn)大田大豆單株植株幾何參數(shù)(高度、體積)準(zhǔn)確獲取,該文構(gòu)建了基于機(jī)載激光雷達(dá)(LiDAR)的農(nóng)作物表型探測(cè)系統(tǒng),并開(kāi)展了大田標(biāo)定和探測(cè)試驗(yàn)。針對(duì)大田大豆壟上種植模式下地面平整度差異大、植株枝葉交接難以區(qū)分的問(wèn)題,提出了一種LiDAR表型探測(cè)系統(tǒng)下的基于局部鄰域特征分割與均值漂移算法的提取方法。在獲取的點(diǎn)云中,首先使用基于局部鄰域特征的語(yǔ)義分割方法提取一壟植株行,然后采用均值漂移算法提取單株植株,最后進(jìn)行植株表面重建和植株幾何參數(shù)統(tǒng)計(jì)。LiDAR表型探測(cè)系統(tǒng)在沿探測(cè)系統(tǒng)前進(jìn)運(yùn)動(dòng)方向、垂直運(yùn)動(dòng)方向、垂直地面方向最大誤差分別為0.58%(5.8 cm)、?1.75%(?7.0 cm)、?1.74%(?3.4 cm)。該文采用的基于局部鄰域特征的分割方法,植株與地面分類(lèi)的效果良好,人工統(tǒng)計(jì)植株數(shù)量相比,檢測(cè)植株數(shù)量的平均相對(duì)誤差為11.83%。相對(duì)于常用的RANSAC(random sample consensus)方法,使用該文提出的高度計(jì)算方法,大豆植株高度平均相對(duì)誤差從9.05%下降到5.14%,利用alpha-shape算法重建后的冠層體積平均值為48.5 dm3。該文工作可為作物植株分割和體積統(tǒng)計(jì)提供借鑒。
激光;提?。粩?shù)據(jù)處理;單株點(diǎn)云;幾何參數(shù)
農(nóng)田信息獲取與分析是精準(zhǔn)農(nóng)業(yè)實(shí)施的前提[1]。作物形貌、位置信息是精確噴霧技術(shù)的基礎(chǔ)[2],探測(cè)作物幾何信息是對(duì)靶變量噴霧的重要研究?jī)?nèi)容[3]??蓪?shí)現(xiàn)探測(cè)作物形貌信息的技術(shù)手段越來(lái)越多[4],應(yīng)用于靶標(biāo)作物形貌探測(cè)的技術(shù)手段主要有超聲波探測(cè)、紅外探測(cè)、可見(jiàn)光成像與圖像處理、激光雷達(dá)(LiDAR)等。超聲波探測(cè)在獲取柑橘冠層形貌信息、計(jì)算冠層體積[5],估算果園靶標(biāo)冠層密度[6]等方面得到了應(yīng)用,Maghsoudi等[7]利用神經(jīng)網(wǎng)絡(luò)算法對(duì)超聲波數(shù)據(jù)進(jìn)行學(xué)習(xí),實(shí)現(xiàn)了果樹(shù)體積的可靠估計(jì)。超聲波傳感器成本較低,但存在靶向性低、對(duì)距離敏感、采樣頻率低等問(wèn)題。紅外探測(cè)可通過(guò)確定靶標(biāo)作物的有無(wú)[8]提升變量噴霧機(jī)效果,也可與其他傳感器數(shù)據(jù)結(jié)合,如顏色傳感器[9]實(shí)現(xiàn)只對(duì)綠色植物噴霧。但紅外探測(cè)受環(huán)境光干擾,穩(wěn)定性不足,田間應(yīng)用限制較多。可見(jiàn)光成像技術(shù)也被用在作物表型和參數(shù)測(cè)量中。如通過(guò)分析圖像獲取植株形態(tài)或表面模型信息可以?xún)?yōu)化移栽機(jī)結(jié)構(gòu)與移栽時(shí)間[10]、闡明作物與環(huán)境的協(xié)同關(guān)系[11]、估算大麥生物量[12]等。但可見(jiàn)光成像技術(shù)對(duì)光照條件要求高,不能滿(mǎn)足實(shí)際生產(chǎn)要求[13]。
激光雷達(dá)具有精度高、靶向性強(qiáng)、響應(yīng)時(shí)間短、受光照條件影響小等特點(diǎn)[14],被廣泛應(yīng)用在樹(shù)冠參數(shù)的獲取[15-16]、果樹(shù)生物量與產(chǎn)量的預(yù)測(cè)[17-19]、實(shí)時(shí)修正的靶標(biāo)作物表型[20]等方面。二維激光掃描儀與亞米級(jí)全球?qū)Ш叫l(wèi)星系統(tǒng)GNSS(global navigation satellite system)結(jié)合,可以獲取高精度的田間作業(yè)地理信息,提升施藥、灌溉、田間信息管理水平。Garrido等[21]實(shí)現(xiàn)了溫室發(fā)育初期玉米作物形態(tài)結(jié)構(gòu)重建。Sun等[22]將二維激光掃描儀與高精度(1 cm)RTK-GPS組成高通量信息獲取系統(tǒng)實(shí)現(xiàn)了對(duì)棉花高度的可靠估計(jì)。Escola等[23]將實(shí)時(shí)差分GNSS與二維激光掃描儀UTM30-LX-EW數(shù)據(jù)融合得到了橄欖樹(shù)冠層幾何參數(shù)和結(jié)構(gòu)。程曼等[24]針對(duì)花生田間特殊工作環(huán)境,設(shè)計(jì)了以二維激光掃描儀為核心的數(shù)據(jù)獲取系統(tǒng),通過(guò)對(duì)冠層剖面點(diǎn)云曲線(xiàn)擬合獲取邊界,進(jìn)而獲取花生高度,提高了高度獲取效率。三維LiDAR具有更高數(shù)據(jù)密度獲取能力,常被用于田間地理信息獲取以估算靶標(biāo)作物生物量。Martin等[25]將三維激光掃描儀數(shù)據(jù)融合GNSS與慣性測(cè)量單元(inertial measurement unit,IMU)繪制了小麥地圖并估算了高度與體積。Ravi等[26]基于無(wú)人機(jī)與三維LiDAR構(gòu)建的高通量探測(cè)系統(tǒng),用于檢測(cè)作物高度與冠層覆蓋率的變化,繪制作物地圖。
已有基于點(diǎn)云的獲取作物高度與體積的文獻(xiàn)中主要集中在不區(qū)分單株作物且植株無(wú)枝葉交接或地面相對(duì)平整的條件,但對(duì)于地面平整度差異大、枝葉交接較多條件下提取單株植株的文獻(xiàn)較少。為此本文提出一種利用局部鄰域特征分割與均值漂移算法的點(diǎn)云分割提取方法,進(jìn)行大豆作物的植株與地面分割及單株提??;采用alpha-shape算法進(jìn)行作物三維重建和幾何參數(shù)計(jì)算。
1.1.1 數(shù)據(jù)獲取系統(tǒng)組成
三維激光雷達(dá)農(nóng)作物表型探測(cè)系統(tǒng)(簡(jiǎn)稱(chēng)探測(cè)系統(tǒng))主要由移動(dòng)測(cè)量端、基準(zhǔn)站和PC端組成。移動(dòng)測(cè)量端包括激光掃描儀LASER和位姿測(cè)量單元,其中激光掃描儀采用美國(guó)Velodyne公司的VLP-16型三維激光掃描儀,位姿測(cè)量單元包括NovAtel公司的SPAN-IGM-A1型組合導(dǎo)航系統(tǒng)、GPS1000型天線(xiàn)、存儲(chǔ)控制器、電臺(tái)等。移動(dòng)測(cè)量端主要部件參數(shù)如表1所示;基準(zhǔn)站由北斗星通公司的C280-AT型接收機(jī)、數(shù)據(jù)記錄儀、GPS1000型天線(xiàn)、電源等組成;PC端可通過(guò)電臺(tái)實(shí)現(xiàn)對(duì)移動(dòng)測(cè)量端的遠(yuǎn)程控制。為高效獲取靶標(biāo)作物信息,移動(dòng)測(cè)量端搭載在大疆M600pro無(wú)人機(jī)上。為了得到田間地理環(huán)境三維點(diǎn)云數(shù)據(jù),需將激光掃描儀獲得的三維點(diǎn)云數(shù)據(jù)PCAP文件與通過(guò)后差分技術(shù)得到的位姿數(shù)據(jù)POS文件進(jìn)行數(shù)據(jù)融合。但基準(zhǔn)站與移動(dòng)測(cè)量端GNSS接收機(jī)采集的原始數(shù)據(jù)需解算后才能進(jìn)行GNSS與慣導(dǎo)系統(tǒng)(inertial navigation system,INS)數(shù)據(jù)的耦合。
表1 VLP-16和SPAN-IGM-A1參數(shù)表
使用NovAtel公司的位姿解算軟件Inertial Explorer 8.70完成數(shù)據(jù)的解算與耦合,解算后可得到GNSS數(shù)據(jù)、IMU數(shù)據(jù)、航向數(shù)據(jù)等,通過(guò)GNSS與IMU數(shù)據(jù)緊耦合得到激光掃描儀的位姿數(shù)據(jù)POS文件。使用北京北斗星通公司的點(diǎn)云數(shù)據(jù)處理軟件Li-Acquire完成PCAP文件與POS文件的數(shù)據(jù)集成生成標(biāo)準(zhǔn)的LAS格式點(diǎn)云文件。數(shù)據(jù)集成流程圖如圖1所示。
圖1 數(shù)據(jù)集成流程圖
1.1.2 試驗(yàn)環(huán)境與方案
田間試驗(yàn)于2017年8月至9月在農(nóng)業(yè)農(nóng)村部南京農(nóng)業(yè)機(jī)械化研究所白馬教學(xué)科研基地(31.62°N,119.18°E)育種試驗(yàn)田進(jìn)行,如圖2所示。試驗(yàn)田尺寸為長(zhǎng)度41 m,寬度18 m,共7壟,壟間有凹溝,每壟種植5行作物,株距0.45 m,行距0.55 m。
靶標(biāo)作物為生長(zhǎng)了55~60 d處于結(jié)莢期階段的大豆。數(shù)據(jù)采集過(guò)程中,最大環(huán)境風(fēng)速為1.5 m/s,無(wú)人機(jī)路徑規(guī)劃在飛控軟件中進(jìn)行設(shè)置。設(shè)置帶寬為18 m,帶寬重疊率為40%,無(wú)人機(jī)的水平前進(jìn)速度不超過(guò)0.5 m/s,激光雷達(dá)距離地面平均高度為9.0 m,選擇測(cè)距范圍為2~20 m的點(diǎn)云數(shù)據(jù)用于分析,在18 m掃描帶寬范圍內(nèi),總體上掃描點(diǎn)密度約為1 600點(diǎn)/m2。
圖2 田間試驗(yàn)區(qū)與掃描路徑示意圖
手動(dòng)測(cè)量試驗(yàn)田一壟中6株大豆作物高度與體積,并與系統(tǒng)測(cè)量值對(duì)比。在進(jìn)行田間試驗(yàn)同時(shí),設(shè)置了大田數(shù)據(jù)精度驗(yàn)證試驗(yàn)方案。在試驗(yàn)田旁放置1塊長(zhǎng)方形板材和4根標(biāo)桿,布置方式及尺寸如圖3所示。
注:L1、L2、L3、L4分別為4根標(biāo)桿長(zhǎng)度,cm;L12、L23、L34、L41分別為標(biāo)桿之間的距離,m;a和b分別為長(zhǎng)方形板材寬度和長(zhǎng)度,cm。
通過(guò)點(diǎn)云中12,23,34,41的測(cè)量值與手工測(cè)量值的對(duì)比,驗(yàn)證試驗(yàn)過(guò)程中系統(tǒng)的測(cè)量精度。其中點(diǎn)云測(cè)量值的獲取方法是使用LiDAR360軟件,手工測(cè)量數(shù)據(jù)通過(guò)卷尺測(cè)量,均測(cè)量3次并求平均值。
數(shù)據(jù)處理的目標(biāo)是提取LAS文件中單株作物的三維點(diǎn)云,并完成靶標(biāo)作物的幾何參數(shù)提取和三維建模。該目標(biāo)分為點(diǎn)云數(shù)據(jù)預(yù)處理、分割地面與植株點(diǎn)云、獲取單株大豆植株及計(jì)算單株作物冠層高度與體積。
1.2.1 點(diǎn)云數(shù)據(jù)預(yù)處理
由于采集的田間地理點(diǎn)云包含大量的非靶標(biāo)作物信息和離群點(diǎn),導(dǎo)致處理時(shí)間和難度增加。本文通過(guò)設(shè)置感興趣區(qū)域ROI(region of interest)和基于鄰域平均距離的方法[27]完成點(diǎn)云去噪。
1)選取ROI點(diǎn)云。探測(cè)系統(tǒng)獲取的點(diǎn)云數(shù)據(jù),其坐標(biāo)系和田塊壟方向不一致,且平面坐標(biāo)數(shù)值較大,為此根據(jù)田塊中心和壟走向,建立田塊坐標(biāo)系,將原始點(diǎn)云數(shù)據(jù)轉(zhuǎn)換到田塊坐標(biāo)系下。通過(guò)設(shè)置田塊坐標(biāo)系下各坐標(biāo)軸的上下界,選取界內(nèi)點(diǎn)云數(shù)據(jù)得到ROI點(diǎn)云。
1.2.2 地面分割與作物點(diǎn)云提取
為了提取作物幾何參數(shù),通常采用設(shè)置分割閾值[22]或RANSAC(random sample consensus)算法[33]。但對(duì)于壟上種植模式,地面平整度差異大,設(shè)置統(tǒng)一閾值或使用RANSAC算法誤差較大。本文提出使用基于局部鄰域特征語(yǔ)義分割的算法,將靶標(biāo)作物點(diǎn)云與地面點(diǎn)云分離?;诰植苦徲蛱卣鞯恼Z(yǔ)義分割算法主要分為3步:
3)監(jiān)督分類(lèi)及評(píng)價(jià)指標(biāo)。根據(jù)不同特征組合對(duì)點(diǎn)云分類(lèi),選取隨機(jī)森林分類(lèi)器對(duì)測(cè)試集數(shù)據(jù)分類(lèi)標(biāo)簽進(jìn)行預(yù)測(cè)[31]。本文選擇ROC曲線(xiàn)的AUC值作為分類(lèi)效果的評(píng)價(jià)指標(biāo)。ROC的縱軸表示真正類(lèi)比率,橫軸表示負(fù)正類(lèi)比率,AUC值為ROC曲線(xiàn)下的面積,數(shù)值越接近1代表分類(lèi)器分類(lèi)效果越好。并將分類(lèi)效果與RANSAC算法的分類(lèi)效果進(jìn)行對(duì)比。
1.2.3 獲取單株大豆植株
1.2.4 計(jì)算大豆冠層高度與體積
1)冠層高度定義與測(cè)量
本文將大豆實(shí)際株高定義為莖稈與地面交點(diǎn)avg至最高葉片max的距離,如圖4a所示。由于激光束通過(guò)密集枝葉到達(dá)地面的概率較小,造成avg處采樣頻率低,本文通過(guò)avg點(diǎn)周?chē)牡匦喂烙?jì)其真值。
注:和¢分別為植株株高的實(shí)際測(cè)量值和系統(tǒng)測(cè)量值,cm;為植株點(diǎn)云的重心。
Note:and¢are the values of manual measurement and system measurement of plant height respectively, cm;is the gravity center of point cloud of plant.
圖4 高度測(cè)量示意圖
Fig.4 Schematic diagram of height measurement
2)冠層體積定義與測(cè)量
植株的冠層體積不包含其中的間隙,但VLP-16型激光掃描儀精度為±3 cm,難以準(zhǔn)確探測(cè)作物葉片厚度、枝葉寬度等參數(shù),并且由于枝葉阻擋,下層枝葉外形難以準(zhǔn)確完整獲取,無(wú)法準(zhǔn)確獲得嚴(yán)格意義上的冠層體積。為此本文將作物體積看為由點(diǎn)云邊界界定的3D實(shí)體體積。三維點(diǎn)云是對(duì)作物體積的離散化表示,恢復(fù)作物的體積需對(duì)點(diǎn)云進(jìn)行三維重建。本文使用alpha shape算法對(duì)作物進(jìn)行三維重建,alpha shape算法是Delaunay三角剖分算法的1種擴(kuò)展形式,可從散亂空間點(diǎn)集中求得點(diǎn)云輪廓[32]。
系統(tǒng)測(cè)量的長(zhǎng)方形板材和標(biāo)桿的幾何尺寸如表2所示。因部分長(zhǎng)方形板材被作物遮擋,故只列出了邊長(zhǎng)數(shù)據(jù),表2顯示了長(zhǎng)方形板材與標(biāo)桿的系統(tǒng)測(cè)量值與手工測(cè)量值的對(duì)比。其中沿作物行方向即掃描前進(jìn)方向(12、34)上尺寸為1 000 cm時(shí),最大相對(duì)誤差為0.58%,誤差值為5.8 cm;在垂直于作物行方向(23、41)上尺寸為400 cm時(shí),最大相對(duì)誤差?1.75%,誤差值為?7.0 cm;在垂直于地面方向上,尺寸為195 cm時(shí),最大的相對(duì)誤差為?1.74%,誤差值為?3.4 cm。所以大田試驗(yàn)精度可以達(dá)到10 cm以?xún)?nèi),空間精度關(guān)系為前進(jìn)方向精度最高,垂直于地面方向與垂直于作物行方向精度接近。部分相對(duì)誤差較小,一方面可能單次測(cè)量存在誤差較小的情況,另一方面可能激光掃描儀在較小測(cè)距范圍(<20 m)的精度較高[35]。
表2 長(zhǎng)方形板材與標(biāo)桿的手工測(cè)量值與系統(tǒng)測(cè)量值對(duì)比
最優(yōu)鄰域初始范圍設(shè)置為5~100個(gè)點(diǎn)。計(jì)算的最優(yōu)鄰域結(jié)果如圖5所示。由圖可知,97%的點(diǎn)最優(yōu)鄰域包含點(diǎn)數(shù)量小于20個(gè)。使用3D點(diǎn)云標(biāo)注工具對(duì)點(diǎn)云文件進(jìn)行手工標(biāo)注,建立標(biāo)簽數(shù)據(jù)集,以衡量算法對(duì)地面檢測(cè)分割的效果,標(biāo)注后的訓(xùn)練集如圖6所示,其中亮色表示作物。
圖5 鄰域分布柱狀圖
圖6 訓(xùn)練集標(biāo)記結(jié)果
本研究應(yīng)用Gini系數(shù)算法、Chi-square特征算法和ReliefF算法等常用算法對(duì)26個(gè)特征進(jìn)行評(píng)估,結(jié)果如圖7所示。特征順序值越小表示對(duì)應(yīng)特征對(duì)點(diǎn)云分類(lèi)效果越好,各算法選擇分類(lèi)效果最好的前5個(gè)特征作為分類(lèi)特征組合。
圖7 特征選擇結(jié)果
在以上3種算法的特征組合分類(lèi)的基礎(chǔ)上,本研究同時(shí)引入隨機(jī)森林方法,針對(duì)2D特征、3D特征及所有特征進(jìn)行選擇,同樣選取分類(lèi)效果最好的前5個(gè)特征作為特征組合。不同特征組合下的分類(lèi)的效果及其與RANSAC分類(lèi)算法效果對(duì)比如表3所示。在隨機(jī)森林算法基于所有特征確定的特征組合進(jìn)行分類(lèi)時(shí),分類(lèi)效果的評(píng)價(jià)指標(biāo)AUC值最大,為0.994,相比RANSAC算法本文方法不需人工調(diào)節(jié)參數(shù),利于自動(dòng)處理點(diǎn)云數(shù)據(jù)。
均值漂移算法需要設(shè)置的參數(shù)為帶寬,因大豆作物播種時(shí)按照一定的株距與行距,故在設(shè)置帶寬時(shí)考慮這些先驗(yàn)知識(shí)。本文大田大豆的行距為0.55 m,株距為0.45 m,設(shè)置了帶寬取值區(qū)間為19~21 cm的3個(gè)梯度帶寬以檢測(cè)帶寬與先驗(yàn)知識(shí)的關(guān)系。測(cè)試數(shù)據(jù)集中實(shí)際采集了121株大豆植株,每個(gè)帶寬與植株數(shù)量對(duì)應(yīng)關(guān)系如圖8所示。分別計(jì)算各作物行檢測(cè)植株數(shù)量與人工統(tǒng)計(jì)植株數(shù)量相對(duì)誤差,分析可知,當(dāng)=20 cm時(shí),檢測(cè)植株數(shù)量平均相對(duì)誤差為11.83%,且分布相關(guān)性最高,為0.675??紤]到作物生長(zhǎng)的差異性與農(nóng)藝水平等因素的影響,可以猜測(cè)均值漂移用于分割單株點(diǎn)云的最佳分割帶寬opt可在行距或株距最小值的一半附近得到。本文中最優(yōu)分割帶寬與猜想理論帶寬誤差為2.5 cm。
表3 不同算法分類(lèi)結(jié)果對(duì)比
注:CK為人工統(tǒng)計(jì);h為帶寬。
2.4.1 植株高度結(jié)果
植株高度測(cè)量誤差與重心投影點(diǎn)鄰域范圍的關(guān)系,如表4所示。表4分析了6株大豆植株系統(tǒng)測(cè)量高度與重心投影點(diǎn)鄰域范圍取值的關(guān)系??梢钥闯鲋苯硬捎弥匦耐队包c(diǎn)(k=1時(shí))時(shí),植株高度最大相對(duì)誤差出現(xiàn)在高度為36 cm的植株上,相對(duì)誤差為17.66%,總體平均相對(duì)誤差為9.05%。當(dāng)采用鄰近的方法計(jì)算植株高度時(shí),k=20時(shí),系統(tǒng)獲得的相對(duì)誤差有4株都顯著下降,平均相對(duì)誤差下降到5.14%;k=100時(shí),平均相對(duì)誤差下降到4.96%。雖然k=100時(shí)平均相對(duì)誤差略小于k=20時(shí),考慮到植株點(diǎn)云的規(guī)模較小,選擇k=20作為計(jì)算植株高度時(shí)的投影點(diǎn)鄰域范圍。
表4 kter取值與相對(duì)誤差關(guān)系
除了從相對(duì)誤差角度闡述植株高度計(jì)算方法的準(zhǔn)確性,本文還對(duì)比了RANSAC和閾值法[22]2種方法,結(jié)果如表5所示。使用RANSAC算法和閾值法平均相對(duì)誤差分別為9.72%和31.07%,高于本文高度計(jì)算方法的相對(duì)誤差。所以,使用本文方法可以較高精度計(jì)算作物高度。對(duì)一壟大豆植株的高度統(tǒng)計(jì)結(jié)果如圖9所示。
表5 3種植株高度計(jì)算方法對(duì)比結(jié)果
圖9 植株高度與冠幅
大豆植株平均高度為65.9 cm,標(biāo)準(zhǔn)差為11.6 cm,冠幅的平均尺寸為40.1 cm,標(biāo)準(zhǔn)差為14.1 cm。數(shù)據(jù)顯示植株高度與冠幅差異較大,可為變量農(nóng)機(jī)提供決策信息。
2.4.2 植株體積結(jié)果
因定義冠層體積為外部輪廓所占的無(wú)空隙三維網(wǎng)格所占空間。所以在無(wú)空洞的情況下,選取越小,形成的網(wǎng)格化體積越小,越接近于真實(shí)體積。本文以單株橙樹(shù)冠層體積凹度=0.75為基礎(chǔ)[34],尋找適合結(jié)莢期大豆植株冠層的凹度值。當(dāng)=1.85時(shí),alpha shape算法重建的無(wú)縫隙包絡(luò)認(rèn)為是合理的表面重建,如圖10所示。
使用規(guī)則幾何體法(立方體法)估計(jì)作物體積,由于作物邊界與規(guī)則幾何體體間空隙大,造成作物體積被嚴(yán)重高估[36]。本文應(yīng)用優(yōu)化凹度值后的alpha shape算法重建了作物表面輪廓,減少了這種空隙的存在。
圖11顯示了立方體法和alpha shape算法對(duì)122棵大豆植株體積的統(tǒng)計(jì)結(jié)果。從圖中可以看出立方體法計(jì)算的體積遠(yuǎn)大于alpha shape算法計(jì)算的體積,立方體法包絡(luò)形成的體積平均值為125.6 dm3,后者平均值為48.5 dm3,本文計(jì)算的體積較接近植株的實(shí)際情況。但是,由于沒(méi)有對(duì)大豆植株進(jìn)行實(shí)際體積測(cè)量,未能將計(jì)算結(jié)果與實(shí)際測(cè)量結(jié)果進(jìn)行對(duì)比,未來(lái)將對(duì)定量的分析合理凹度值下alpha shape算法所得體積與真實(shí)體積之間的關(guān)系做進(jìn)一步研究。
圖10 作物表面重建結(jié)果
圖11 不同算法計(jì)算的單株植株體積
為了實(shí)現(xiàn)大田大豆單株植株高度與體積準(zhǔn)確自動(dòng)獲取,本文對(duì)基于大田大豆三維點(diǎn)云的單株植株高度與體積自動(dòng)提取進(jìn)行了研究,提出了以一種基于點(diǎn)云局部特征語(yǔ)義分割算法與均值漂移算法的單株高度與體積準(zhǔn)確獲取的方法。結(jié)果表明:
1)本文構(gòu)建的三維掃描探測(cè)系統(tǒng)可以提供前進(jìn)運(yùn)動(dòng)方向、垂直運(yùn)動(dòng)方向、垂直地面方向分別為0.58%(5.8 cm)、?1.75%(?7.0 cm)、?1.74%(?3.4 cm)的測(cè)量精度。
2)基于2D和3D局部特征組合的方法,可以實(shí)現(xiàn)大田大豆植株與地面點(diǎn)云分類(lèi),分類(lèi)效果指標(biāo)AUC值為0.994。
3)在高度計(jì)算方法上,使用作物點(diǎn)云重心在二次擬合曲面(地面)上投影取平均值作為估計(jì)植株莖部與地面的交點(diǎn)計(jì)算作物高度,比直接選取高度最小值計(jì)算作物高度,平均相對(duì)誤差可由9.05%下降到5.14%。
4)本文應(yīng)用優(yōu)化凹度值后的alpha shape算法重建了作物表面輪廓,減少了這種空隙的存在。計(jì)算的植株體積平均值為48.5 dm3,與立方體法包絡(luò)法的結(jié)果相比,較接近植株的實(shí)際情況。
[1]潘瑜春,趙春江. 地理信息技術(shù)在精準(zhǔn)農(nóng)業(yè)中的應(yīng)用[J]. 農(nóng)業(yè)工程學(xué)報(bào),2003,19(4):1-6.
Pan Yuchun, Zhao Chunjiang. Application of geographic information technologiesin precision agriculture[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2003, 19(4): 1-6. (in Chinese with English abstract)
[2]王萬(wàn)章,洪添勝,李捷,等. 果樹(shù)農(nóng)藥精確噴霧技術(shù)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2004,20(6):98-101.
Wang Wanzhang, Hong Tiansheng, Li Jie, et al. Review of the pesticide precision orchard spraying technologies[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2004, 20(6): 98-101. (in Chinese with English abstract)
[3]邱白晶,閆潤(rùn),馬靖,等. 變量噴霧技術(shù)研究進(jìn)展分析[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2015,46(12):59-72. Qiu Baijing, Yan Run, Ma Jing, et al. Research progress analysis of variable rate sprayer technology[J]. Transactions of the Chinese Society for Agricultural Machinery, 2015, 46(3): 59-72. (in Chinese with English abstract)
[4]劉建剛,趙春江,楊貴軍,等. 無(wú)人機(jī)遙感解析田間作物表型信息研究進(jìn)展[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(24):98-106. Liu Jiangang, Zhao Chunjiang, Yang Guijun, et al. Review of field-based phenotyping by unmanned aerial vehicle remote sensing platform[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(24): 98-106. (in Chinese with English abstract)
[5]Tumbo S D, Salyani M, Whitney J D, et al. Investigation of laser and ultrasonic ranging sensors for measurements ofcitrus canopy volume[J]. Transaction of the ASAE, 2002, 18(3): 367-372.
[6]Li Hanzhe, Zhai Changyuan,Weckler P, et al. A canopy density model for planar orchard target detection based on ultrasonic sensors[J]. Sensors, 2016, 17(1): 31-45.
[7]Maghsoudi H, Minaei S, Ghobadian B, et al. Ultrasonic sensing of pistachio canopy for low-volume precision spraying[J]. Computers and Electronics in Agriculture, 2015, 112: 149-160.
[8]Zou Wei, Wang Xiu, Deng Wei, et al. Design and test of automatic toward-target sprayer used in orchard[C]. sprayer used in orchard[C]. Shenyang: IEEE, 2015: 697-702.
[9]李麗,李恒,何雄奎,等. 紅外靶標(biāo)自動(dòng)探測(cè)器的研制及試驗(yàn)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(12):159-163. Li Li, Li Heng, He Xiongkui, et al. Development and experiment of automatic detection device for infrared target[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(12): 159-163. (in Chinese with English abstract)
[10]劉明峰,胡先朋,廖宜濤,等. 不同油菜品種適栽期機(jī)械化移栽植株形態(tài)特征研究[J]. 農(nóng)業(yè)工程學(xué)報(bào),2015,31(增刊1):79-88. Liu Mingfeng, Hu Xianpeng, Liao Yitao, et al. Morphological parameters characteristics of mechanically transplanted plant in suitable transplanting period for different rape varieties[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2015, 31(Supp.1): 79-88. (in Chinese with English abstract)
[11]王傳宇,杜建軍,郭新宇,等. 基于時(shí)間序列圖像的玉米植株干旱脅迫表型檢測(cè)方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(21):189-195. Wang Chuanyu, Du Jianjun, Guo Xinyu, et al. Maize crop drought stress phenotype testing method based on time-series images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(21): 189-195. (in Chinese with English abstract)
[12]Juliane B, Andreas B, Simon B, et al. Estimating biomass of barley using crop surface models (CSMs) derived from UAV-based RGB imaging[J]. Remote Sensing, 2014, 6(11): 10395-10412.
[13]孫智慧,陸聲鏈,郭新宇,等. 基于點(diǎn)云數(shù)據(jù)的植物葉片曲面重構(gòu)方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(3):184-190. Sun Zhihui, Lu Shenglian, Guo Xinyu, et al. Surfaces reconstruction of crop leaves based on point cloud data[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(3): 184-190. (in Chinese with English abstract)
[14]劉慧,李寧,沈躍,等. 模擬復(fù)雜地形的噴霧靶標(biāo)激光檢測(cè)與三維重構(gòu)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(18):84-91. Liu Hui, Li Ning, Shen Yue, et al. Spray target laser scanning detection and three-dimensional reconstruction under simulated complex terrain[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(18): 84-91. (in Chinese with English abstract)
[15]Babcock C, Finley A O, Andersen H E, et al. Geostatistical estimation of forest biomass in interior Alaska combining Landsat-derived tree cover, sampled airborne LiDAR and field observations[J]. Remote Sensing of Environment, 2018, 212: 212-230.
[16]Méndez V, Rosell-Polo J R, Pascual M, et al. Multi-tree woody structure reconstruction from mobile terrestrial laser scanner point clouds based on a dual neighbourhood connectivity graph algorithm[J]. Biosystems Engineering, 2016, 148: 34-47.
[17]郭彩玲,宗澤,張雪,等. 基于三維點(diǎn)云數(shù)據(jù)的蘋(píng)果樹(shù)冠層幾何參數(shù)獲取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017,33(3):175-181. Guo Cailing, Zong Ze, Zhang Xue, et al. Apple tree canopy geometric parameters acquirement based on 3D point clouds[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(3): 175-181. (in Chinese with English abstract)
[18]Sanz R, Rosella J, Llorensb J, et al. Relationship between tree row LIDAR-volume and leaf area density for fruit orchards and vineyards obtained with a LIDAR 3D dynamic measurement system[J]. Agricultural and Forest Meteorology, 2013, 172(3): 153-162.
[19]Sanz R, Llorensb J, Escolà A, et al. LIDAR and non-LIDAR-based canopy parameters to estimate the leaf area in fruit trees and vineyard[J]. Agricultural and Forest Meteorology, 2018, 260: 229-239.
[20]劉慧,李寧,沈躍,等. 融合激光三維探測(cè)與IMU姿態(tài)角實(shí)時(shí)矯正的噴霧靶標(biāo)檢測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017,33(15):88-97. Liu Hui, Li Ning, Shen Yue, et al. Spray target detection based on laser scanning sensor and real-time correction of IMU attitude angle[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2017, 33(15): 88-97. (in Chinese with English abstract)
[21]Garrido M, Paraforos D, Reiser D, et al. 3D maize crop reconstruction based on georeferenced overlapping LiDAR point clouds[J]. Remote Sensing, 2015, 7(12): 17077-17096.
[22]Sun Shangpeng, Li Changying, Paterson A H, et al. In-field high throughput phenotyping and cotton crop growth analysis using LiDAR[J]. Frontiers in Crop Science, 2018, 9: 16-37.
[23]Escola A, Martínez-Casasnovas J A, Rufat J, et al. Mobile terrestrial laser scanner applications in precision fruticulture/ horticulture and tools to extract information from canopy point clouds[J]. Precision Agriculture, 2017, 18(1): 111-132.
[24]程曼,蔡振江,Ning Wang,等. 基于地面激光雷達(dá)的田間花生冠層高度探測(cè)系統(tǒng)研制[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(1):180-187. Cheng Man, Cai Zhenjiang, Wang Ning, et al. System design for peanut canopy height information acquisition based on LiDAR[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(1): 180-187. (in Chinese with English abstract)
[25]Martin C, Morten L, Rasmus J, et al. Designing and testing a UAV mapping system for agricultural field surveying[J]. Sensors, 2017, 17(12): 2703-2722.
[26]Ravi, R, Lin Y J, Shamseldin T, et al. Implementation of UAV-Based lidar for high throughput phenotyping[C]. Valencia: IEEE, 2018: 2018: 8761-8764.
[27]夏春華,施瀅,尹文慶. 基于TOF深度傳感的植物三維點(diǎn)云數(shù)據(jù)獲取與去噪方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(6):168-174.
Xia Chunhua, Shi Ying, Yin Wenqing. Obtaining and denoising method of three-dimensional point cloud data of plants based onTOF depth sensor[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(6): 168-174. (in Chinese with English abstract)
[28]Dittrich A, Weinmann M, Hinz S. Analytical and numerical investigations on the accuracy and robustness of geometric features extracted from 3D point cloud data[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 126: 195-208.
[29]Raileanu L E, Stoffel K. Theoretical comparison between the gini index and information gain criteria[J]. Annals of Mathematics and Artificial Intelligence, 2004, 41(1): 77-93.
[30]Spolaor N, Cherman E A, Monard M C, et al. ReliefF for multi-label feature selection[C]. Fortaleza: IEEE, 2013: 6-11.
[31]Ru? G, Brenning A. Data mining in precision agriculture: management of spatial information[C]. Berlin: Springer, 2010: 350-359.
[32]Bernardini F, Mittleman J, Rushmeier H, et al. The ball-pivoting algorithm for surface reconstruction[J]. IEEE Transactions on Visualization and Computer Graphics, 1999, 5(4): 349-359.
[33]Weiss U, Biber P. Plant detection and mapping for agricultural robots using a 3D LIDAR sensor[J]. Robotics and Autonomous Systems, 2011, 59(5): 265-273.
[34]Cola?o A, Trevisan R, Molin J, et al. A method to obtain orange crop geometry information using a mobile terrestrial laser scanner and 3D modeling[J]. Remote Sensing, 2017, 9(8): 763-784.
[35]Heinz E, Eling C, Wieland M, et al. Development, calibration and evaluation of a portable and direct georeferenced laser scanning system for kinematic 3D mapping[J]. Journal of Applied Geodesy, 2015, 9(4): 227-243.
[36]Yan Zhaojin, Liu Rufei, Cheng Liang, et al. A concave hull methodology for calculating the crown volume of individual trees based on vehicle-borne LiDAR data[J]. Remote Sensing. 2019, 11(6): 623-642.
Extraction of geometric parameters of soybean canopy by airborne 3D laser scanning
Guan Xianping1, Liu Kuan1, Qiu Baijing1※, Dong Xiaoya1, Xue Xinyu2
(1.,,,212013,; 2.,,210014,)
Accurate acquisition and analysis of crop geometric information is an important basis for the implementation of precision agriculture. Canopy height and volume are important decision parameters for variable sprayer application rate. In the field environment, the large change of ambient light has an important influence on the measurement of canopy geometry information by sensors. At the same time, there are few researches on the problems of remove the effect of ground roughness and difficulty in distinguishing individual plants due to branches and leaves crossing under the ridge planting mode of field soybean. Therefore, it is necessary to design an information acquisition system that is less affected by light conditions and an algorithm to improve the ability to extract geometric information from individual crops. In this study, a crop phenotype detection system based on airborne lidar was constructed and its accuracy was verified. A method of extracting individual plant based on local geometric feature segmentation and mean shift algorithm was proposed. In the process of soybean plant and ground classification, firstly, the local geometric features constructed in the optimal neighborhood are classified into 2D and 3D local shape features according to their dimensions. Secondly, in order to select 5 feature combinations that are strongly related to classification, all features were evaluated using Gini index algorithm, Chi-square algorithm, ReliefF algorithm, and random forest method. Finally, according to different feature combinations, a random forest classifier is selected to predict the test set data. In the process of extracting a single soybean plant, the point cloud data of different plants were used to obtain the point cloud data of a single plant using the mean shift algorithm to complete the extraction of a single soybean plant. In the process of obtaining geometric information of single plants, the height of plants was defined as the height difference from the intersection point of soybean stem and ground to the highest point of crops. In actual measurement, because the laser beam was blocked by branches and leaves, it is difficult to obtain the intersection point of soybean stem and ground, so the paper used the method of projecting the center of gravity of single point cloud to the ground fitting surface to estimate the intersection point. Furthermore, the plant height of single plant was obtained by subtracting the estimated intersection point from the maximum point. The experimental results showed that the maximum relative errors of the lidar scanning measurement system along the carrier moving direction, vertical moving direction and vertical ground direction were 0.58% (5.8 cm), ?1.75% (?7.0 cm) and ?1.74% (?3.4 cm), respectively. In the process of soybean crop and ground classification, the AUC (area under curve) value of the classification index ROC (receiver operating characteristic) curve was 0.994, achieving a good classification effect based on feature combination which was selected from 26 features using random forest algorithms. The relative error was 11.83% between the number of artificially counted plants and the number of manual measurements, and the distribution correlation was the highest with0.675 when the mean shift algorithm parameter is 20 cm. The average relative error of the height estimated method in this paper was 5.14%, which was better than RANSAC algorithm. This paper can provide reference for crop segmentation and yield statistics. Future research should focus on converting the obtained target crop information into a prescription map and storing it in a server for application in online spraying.
laser; extraction; data processing; point cloud of single crop; geometric parameters
管賢平,劉 寬,邱白晶,董曉婭,薛新宇. 基于機(jī)載三維激光掃描的大豆冠層幾何參數(shù)提取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(23):96-103.doi:10.11975/j.issn.1002-6819.2019.23.012 http://www.tcsae.org
Guan Xianping, Liu Kuan, Qiu Baijing, Dong Xiaoya, Xue Xinyu. Extraction of geometric parameters of soybean canopy by airborne 3D laser scanning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(23): 96-103. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2019.23.012 http://www.tcsae.org
2019-05-11
2019-10-07
國(guó)家重點(diǎn)研發(fā)計(jì)劃項(xiàng)目課題(2017YFD0701403)
管賢平,副研究員,主要從事精確施藥技術(shù)研究。Email:xpguan@ujs.edu.cn
邱白晶,教授,博士生導(dǎo)師,主要從事農(nóng)業(yè)植保機(jī)械領(lǐng)域研究。Email:qbj@ujs.edu.cn
10.11975/j.issn.1002-6819.2019.23.012
S24
A
1002-6819(2019)-23-0096-08
農(nóng)業(yè)工程學(xué)報(bào)2019年23期