翟衛(wèi)欣,潘家文,蘭玉彬,吳才聰
·農(nóng)業(yè)信息與電氣技術(shù)·
基于多元振蕩黏菌算法的田路分割模型參數(shù)優(yōu)化方法
翟衛(wèi)欣1,2,潘家文1,2,蘭玉彬3,4,吳才聰1,2※
(1. 中國(guó)農(nóng)業(yè)大學(xué)信息與電氣工程學(xué)院,北京 100083;2. 農(nóng)業(yè)農(nóng)村部農(nóng)機(jī)作業(yè)監(jiān)測(cè)與大數(shù)據(jù)應(yīng)用重點(diǎn)實(shí)驗(yàn)室, 北京 100083;3. 山東理工大學(xué)農(nóng)業(yè)工程與食品科學(xué)學(xué)院,淄博 255000;4. 國(guó)家精準(zhǔn)農(nóng)業(yè)航空施藥技術(shù)國(guó)際聯(lián)合研究中心,廣州 510642)
田路分割是農(nóng)機(jī)軌跡語(yǔ)義分割模型的重要任務(wù)之一,其目的是將軌跡自動(dòng)分割為田間作業(yè)軌跡和道路行駛軌跡。田路分割模型的參數(shù)是影響其分割準(zhǔn)確性及精度的關(guān)鍵因素,傳統(tǒng)的參數(shù)選擇方法效率較低且難以獲得較好的方案,限制了模型的分割性能。因此,該研究選用基于方向分布的田路分割模型(Field-Road Trajectory Segmentation Based on Direction Distribution,BDFRTS),嘗試從參數(shù)優(yōu)化的角度研究模型的性能提升,提出的方法主要包括兩個(gè)方面,首先建立了一種基于元啟發(fā)式算法(Metaheuristic Algorithms,MA)的田路分割模型參數(shù)優(yōu)化解決方案;其次,在黏菌算法(Slime Mould Algorithm,SMA)的基礎(chǔ)上提出了一種改進(jìn)的多元振蕩黏菌算法(Multiplex Oscillation Slime Mould Algorithm,MOSMA)求解參數(shù)優(yōu)化方案以更好地提高模型的分割性能。MOSMA分別提出一種動(dòng)態(tài)引導(dǎo)策略與多元振蕩策略強(qiáng)化了黏菌的振蕩收縮反應(yīng)及細(xì)胞質(zhì)的流動(dòng)過(guò)程,進(jìn)而增強(qiáng)了算法的優(yōu)化能力。為驗(yàn)證所提參數(shù)優(yōu)化方法的有效性,將博創(chuàng)聯(lián)動(dòng)提供的中國(guó)農(nóng)機(jī)在2021年9 月底—10月中下旬進(jìn)行水稻收割的作業(yè)軌跡作為數(shù)據(jù)集開(kāi)展試驗(yàn)。試驗(yàn)結(jié)果表明,該研究所提方法有效地提升了田路分割模型的準(zhǔn)確性和性能。MOSMA-BDFRTS分割模型在10組高密度軌跡中的平均準(zhǔn)確率相比網(wǎng)格搜索田路分割模型、粒子群田路分割模型分別提高了25和28個(gè)百分點(diǎn);在10組低密度軌跡中分割的平均準(zhǔn)確率分別提高了17和14個(gè)百分點(diǎn)。該研究可為田路分割技術(shù)提供合理的性能優(yōu)化方案,也為農(nóng)業(yè)機(jī)械運(yùn)動(dòng)軌跡分割技術(shù)的效率研究提供參考依據(jù)。
模型;參數(shù)優(yōu)化;田路分割;黏菌算法;動(dòng)態(tài)引導(dǎo);多元振蕩
田路分割的目的是將農(nóng)業(yè)機(jī)械上搭載的全球?qū)Ш叫l(wèi)星系統(tǒng)終端中記錄的時(shí)空軌跡數(shù)據(jù)分割成一系列田間作業(yè)-道路行駛的子軌跡[1]。作為農(nóng)業(yè)領(lǐng)域時(shí)空軌跡處理問(wèn)題重要研究?jī)?nèi)容,田路分割可從大數(shù)據(jù)角度為精準(zhǔn)農(nóng)業(yè)相關(guān)課題的深入研究與科學(xué)管理提供有力支撐[2],如鄉(xiāng)村路網(wǎng)構(gòu)建[3]、田間作業(yè)效率評(píng)估[4-5]、農(nóng)機(jī)跨區(qū)調(diào)度管理[6]、農(nóng)機(jī)作業(yè)路徑規(guī)劃[7-9]等。
現(xiàn)有的田路分割方法大多利用邊界信息來(lái)區(qū)分農(nóng)田軌跡和道路軌跡[10-11],然而,由于邊界信息不夠可靠導(dǎo)致農(nóng)田的實(shí)際邊界難以自動(dòng)標(biāo)定,使得此類(lèi)方法的適用性較差[12]。為了解決這個(gè)問(wèn)題,Chen等[13]利用機(jī)器學(xué)習(xí)技術(shù)[14]提出一種基于方向分布的田路分割模型,通過(guò)聚合內(nèi)含相同空間特征的軌跡點(diǎn)實(shí)現(xiàn)了農(nóng)田軌跡與道路軌跡的自動(dòng)分割。模型的參數(shù)反映了其樣本學(xué)習(xí)的成效,同時(shí)也是影響其分割性能的關(guān)鍵因素,對(duì)參數(shù)進(jìn)行優(yōu)化可以使模型更好地表達(dá)軌跡的特征。然而,真實(shí)場(chǎng)景中的天氣、地形、采樣頻率、作業(yè)任務(wù)等因素都會(huì)影響軌跡變化,使軌跡空間特征呈現(xiàn)多元化現(xiàn)象,為模型參數(shù)的選擇帶來(lái)巨大的挑戰(zhàn)。目前,國(guó)內(nèi)外學(xué)者對(duì)田路分割問(wèn)題鮮有研究,現(xiàn)有模型在參數(shù)標(biāo)定方面仍處于較原始的階段,無(wú)法充分發(fā)揮模型性能,導(dǎo)致分割效果仍不夠理想,如何通過(guò)參數(shù)優(yōu)化提高模型準(zhǔn)確性與性能成為了本文研究的首要問(wèn)題。
參數(shù)優(yōu)化是指從一個(gè)巨大受限解決方案空間找到最佳參數(shù)結(jié)構(gòu),本質(zhì)上是函數(shù)優(yōu)化問(wèn)題。由于不同參數(shù)的交互影響,使分割模型參數(shù)優(yōu)化問(wèn)題呈現(xiàn)出非線性、多峰值的復(fù)雜數(shù)學(xué)特性,用傳統(tǒng)的方法求解費(fèi)時(shí)耗力,選出的參數(shù)結(jié)構(gòu)質(zhì)量差,限制了模型的性能。近年來(lái),元啟發(fā)式算法[15-18]因其魯棒性強(qiáng)、不依賴于目標(biāo)問(wèn)題的梯度信息等優(yōu)勢(shì)被國(guó)內(nèi)外眾多學(xué)者用于求解復(fù)雜非線性問(wèn)題,相比傳統(tǒng)方法在更短時(shí)間內(nèi)取得了相同或更好的結(jié)果[19-23]。張良安等[24]利用粒子群算法(Particle Swarm Algorithm,PSO)對(duì)四足激光除草機(jī)器人腿部結(jié)構(gòu)的相關(guān)參數(shù)進(jìn)行優(yōu)化,有效地提高了續(xù)航能力。Ding等[25]利用PSO算法提出了一種用于植物虛擬生長(zhǎng)模型的參數(shù)優(yōu)化方法。Jamei等[26]提出了一種基于鯨魚(yú)優(yōu)化算法的人工神經(jīng)網(wǎng)絡(luò)模型,增強(qiáng)了模型在地表蒸散指數(shù)預(yù)測(cè)任務(wù)中的準(zhǔn)確性。上述研究表明,利用元啟發(fā)式算法優(yōu)化參數(shù)能夠有效提升相關(guān)算法模型解決實(shí)際任務(wù)的能力。
本文主要從參數(shù)優(yōu)化的角度研究田路分割模型的性能提升問(wèn)題。首先通過(guò)確定編碼方案和適應(yīng)度函數(shù)將參數(shù)優(yōu)化問(wèn)題轉(zhuǎn)化為單目標(biāo)優(yōu)化問(wèn)題,進(jìn)而提出了一種基于元啟發(fā)式算法的田路分割模型參數(shù)優(yōu)化解決方案,旨在利用元啟發(fā)式算法作為優(yōu)化工具求解方案,在解空間內(nèi)自動(dòng)搜索模型的最優(yōu)參數(shù)結(jié)構(gòu),進(jìn)而實(shí)現(xiàn)準(zhǔn)確、快速的田路分割。由于傳統(tǒng)的元啟發(fā)式算法易陷入局部收斂,訓(xùn)練出來(lái)的參數(shù)難以為田路分割模型確定合適的空間特征來(lái)準(zhǔn)確地分割農(nóng)機(jī)軌跡。為提高模型分割效率,對(duì)基本黏菌算法(Slime Mould Algorithm,SMA)改進(jìn)提出了一種多元振蕩黏菌算法(Multiplex Oscillation Slime Mould Algorithm,MOSMA)用以求解模型的參數(shù)優(yōu)化問(wèn)題。為驗(yàn)證所提方法有效性,在真實(shí)的農(nóng)機(jī)作業(yè)軌跡數(shù)據(jù)集上進(jìn)行了實(shí)證研究。
本文選用基于方向分布的田路分割模型(Field-Road Trajectory Segmentation Based on Direction Distribution,BDFRTS)作為所提方法的優(yōu)化實(shí)例。農(nóng)機(jī)田間作業(yè)的軌跡點(diǎn)呈現(xiàn)兩種特點(diǎn):1)田內(nèi)軌跡因車(chē)速較低呈現(xiàn)聚集狀態(tài),使得單位空間內(nèi)的軌跡點(diǎn)密度較高。2)連續(xù)的作業(yè)軌跡帶的行駛方向角度近似平行分布且相反。分割方法利用上述特點(diǎn)對(duì)軌跡進(jìn)行分割,如圖1所示,主要包括三個(gè)步驟:1)為提供高質(zhì)量的軌跡數(shù)據(jù)進(jìn)行分割,對(duì)軌跡噪聲點(diǎn)進(jìn)行平滑處理,對(duì)重復(fù)軌跡點(diǎn)進(jìn)行過(guò)濾。2)使用基于密度的聚類(lèi)算法[27]分割軌跡,并分配不同的語(yǔ)義標(biāo)簽。3)針對(duì)步驟2)錯(cuò)誤分類(lèi)情況(“農(nóng)田”預(yù)測(cè)為“道路”或“道路”預(yù)測(cè)為“農(nóng)田”),采用基于方向分布的推理算法進(jìn)行糾正。具體地說(shuō),在基于軌跡簇的田轉(zhuǎn)路推理規(guī)則中,利用簇中軌跡點(diǎn)在區(qū)域方向分布的差異性將錯(cuò)誤分類(lèi)的假“田”點(diǎn)轉(zhuǎn)換為“路”;軌跡帶生成與軌跡帶噪聲過(guò)濾先將具有相同作業(yè)操作且連續(xù)的軌跡點(diǎn)連接為軌跡帶,再處理軌跡的噪聲點(diǎn);在基于軌跡帶的路轉(zhuǎn)田推理規(guī)則中,利用同一塊農(nóng)田中軌跡帶彼此平行關(guān)系將假“路”點(diǎn)轉(zhuǎn)換為“田”。
圖1 基于方向分布的田路分割模型
田路分割模型能夠在軌跡空間中構(gòu)建一種超平面來(lái)分離不同類(lèi)別的軌跡,為實(shí)現(xiàn)其分割性能的最大化,需對(duì)模型參數(shù)進(jìn)行優(yōu)化來(lái)提高超平面的泛化能力。本文首先針對(duì)田路分割任務(wù)參數(shù)優(yōu)化問(wèn)題的需求,通過(guò)確定問(wèn)題的優(yōu)化目標(biāo)、編碼方案,適應(yīng)度函數(shù)建立起參數(shù)優(yōu)化模型作為優(yōu)化方案;然后以最大化分割準(zhǔn)確率為優(yōu)化目標(biāo),提出一種改進(jìn)的多元振蕩黏菌算法來(lái)求解方案。整體優(yōu)化方法的基本思想是:首先將模型的參數(shù)向量抽象為搜索個(gè)體,在軌跡數(shù)據(jù)的驅(qū)動(dòng)下運(yùn)行田路分割模型,并以分割結(jié)果作為搜索個(gè)體在解空間內(nèi)適應(yīng)情況的評(píng)價(jià)指標(biāo),然后按照優(yōu)化算法的計(jì)算規(guī)則利用適應(yīng)度來(lái)引導(dǎo)個(gè)體的搜索方向,同時(shí)更新搜索個(gè)體所對(duì)應(yīng)參數(shù)向量的信息,最后將更新后的參數(shù)向量重新代入分割模型中計(jì)算適應(yīng)度,如此反復(fù),直到找到最佳參數(shù)結(jié)構(gòu)或滿足算法終止條件。
1.2.1 編碼方案
設(shè)田路分割模型含有個(gè)參數(shù),所有參數(shù)都被縮放到[]區(qū)域中形成一個(gè)統(tǒng)一的解空間,其中與分別表示解空間的上界與下界。解空間中的每個(gè)搜索個(gè)體都表示一項(xiàng)維度為的參數(shù)向量,其中每個(gè)元素都對(duì)應(yīng)模型的一個(gè)參數(shù),其值可根據(jù)模型實(shí)例設(shè)置為二進(jìn)制或浮點(diǎn)數(shù),為算法種群個(gè)體的數(shù)目。本文采用隨機(jī)初始化的方式構(gòu)建算法的種群。
1.2.2 適應(yīng)度函數(shù)
適應(yīng)度函數(shù)代表搜索個(gè)體在解空間的有效信息,同時(shí)也是評(píng)價(jià)參數(shù)向量泛化能力的重要準(zhǔn)則。田路分割任務(wù)屬于二分類(lèi)問(wèn)題,利用機(jī)器學(xué)習(xí)中的三種基于時(shí)間的性能指標(biāo)查準(zhǔn)率(Precision),查全率(Recall)與調(diào)和平均值(F1-score)評(píng)估其分類(lèi)精度。
參數(shù)優(yōu)化的目標(biāo)是通過(guò)持續(xù)調(diào)整每個(gè)參數(shù)向量的值來(lái)獲得泛化效果最好的超平面,進(jìn)而最大限度地提高軌跡的分割效果。本文以最大化分割精準(zhǔn)度為優(yōu)化目標(biāo),將參數(shù)優(yōu)化問(wèn)題轉(zhuǎn)換為單目標(biāo)優(yōu)化問(wèn)題。每個(gè)搜索代理使用F1-score作為其對(duì)應(yīng)參數(shù)向量的適用性標(biāo)準(zhǔn),最優(yōu)搜索代理適應(yīng)度應(yīng)滿足于最大調(diào)和平均值。
1.2.3 基本黏菌算法
黏菌優(yōu)化算法是由Li等[28]于2020年提出的一類(lèi)元啟發(fā)式算法,其靈感來(lái)源于黏菌的覓食行為和形態(tài)變化,數(shù)學(xué)模型可以表達(dá)為
1.2.4 改進(jìn)的多元振蕩黏菌算法
元啟發(fā)式算法的性能取決于其全局搜索能力、局部搜索能力及兩者之間的協(xié)調(diào)平衡。本文選用的SMA全局搜索能力較差,無(wú)法在田路分割模型參數(shù)約束范圍內(nèi)充分搜索。為滿足模型的參數(shù)優(yōu)化性能要求,本研究在SMA的基礎(chǔ)上,提出一種改進(jìn)的多元振蕩黏菌算法求解田路分割模型參數(shù)優(yōu)化問(wèn)題。
1)動(dòng)態(tài)引導(dǎo)策略
2)多元振蕩策略
1.2.5 計(jì)算流程
基于MOSMA參數(shù)優(yōu)化方法計(jì)算步驟可概括為:
1)數(shù)據(jù)處理:清洗軌跡數(shù)據(jù)(包括平滑噪聲點(diǎn)和過(guò)濾重復(fù)點(diǎn)),并劃分為訓(xùn)練集和驗(yàn)證集。
2)參數(shù)設(shè)定與算法初始化:設(shè)置算法相關(guān)的參數(shù),包括種群數(shù)目,參數(shù)維度、參數(shù)上界與下界、最大迭代次數(shù)等,初始化算法種群的位置。
4)參數(shù)優(yōu)化:以適應(yīng)度為引導(dǎo)信息更新算法的參數(shù),種群中的每個(gè)搜索個(gè)體分別通過(guò)式(1)、動(dòng)態(tài)引導(dǎo)策略與多元振蕩策略更新位置信息,最后更新迭代次數(shù)1。
為評(píng)估基于MOSMA的參數(shù)優(yōu)化方法對(duì)田路分割模型計(jì)算效率的提升效果,本文采用真實(shí)的谷物聯(lián)合收割機(jī)在2021年9月底—10月中下旬進(jìn)行水稻收割的作業(yè)軌跡展開(kāi)試驗(yàn),所有軌跡數(shù)據(jù)均由博創(chuàng)聯(lián)動(dòng)公司提供。地理空間上分布于中國(guó)的多個(gè)省級(jí)行政區(qū),具體包括青海省、陜西省、四川省、內(nèi)蒙古、河南省、湖北省、安徽省、江西省、江蘇省。為驗(yàn)證參數(shù)優(yōu)化方法的普適性,試驗(yàn)選用了兩組搭載不同型號(hào)定位終端的農(nóng)機(jī)軌跡數(shù)據(jù)集。為有效地區(qū)分兩組數(shù)據(jù)集,根據(jù)定位終端的采樣頻率不同及其之間的相對(duì)關(guān)系,定義了{(lán)低密度數(shù)據(jù)集,高密度數(shù)據(jù)集}語(yǔ)言集合。其中高密度數(shù)據(jù)集中相鄰軌跡點(diǎn)的時(shí)間間隔為3 s,低密度數(shù)據(jù)集相鄰軌跡點(diǎn)的時(shí)間間隔為5 s。每組數(shù)據(jù)集均包括10個(gè)不同的軌跡樣本,分別由D1~D10和S1~S10表示,詳細(xì)數(shù)據(jù)見(jiàn)表1。
表1 農(nóng)機(jī)運(yùn)動(dòng)軌跡數(shù)據(jù)基本信息
注:緯度、經(jīng)度取自軌跡樣本的起始點(diǎn);樣本規(guī)模是指一幅軌跡所包含的軌跡點(diǎn)數(shù)目。
Note: Latitude and longitude are extracted from the first point of the trajectory sample; the size refers to trajectory point number of a trajectory.
表2表示MOSMA-BDFRTS和GS-BDFRTS兩種模型在高密度軌跡數(shù)據(jù)集中的試驗(yàn)結(jié)果對(duì)比,可以看出,MOSMA-BDFRTS的計(jì)算結(jié)果均優(yōu)于GS-BDFRTS和PSO-BDFRTS。在10組高密度軌跡中,MOSMA-BDFRTS的平均準(zhǔn)確率相比GS-BDFRTS,PSO-BDFRTS分別提高了25和28個(gè)百分點(diǎn)。其中,MOSMA-BDFRTS在D03、D09等樣本中取得了明顯地提升,在農(nóng)田作業(yè)軌跡的F1-score值相比GS-BDFRTS分別提升了3和12個(gè)百分點(diǎn),相比PSO-BDFRTS分別提升了3和4個(gè)百分點(diǎn);在道路行駛軌跡的F1-score值比GS-BDFRTS分別提高了34和69個(gè)百分點(diǎn),相PSOPSO-BDFRTS分別提高了35和50個(gè)百分點(diǎn)。相比GS-BDFRTS和PSO-BDFRTS,MOSMA-BDFRTS在高密度軌跡樣本的分割識(shí)別中的農(nóng)田作業(yè)軌跡的F1-score值均提高了6個(gè)百分點(diǎn);在道路行駛軌跡的F1-score值平均提升了43和50個(gè)百分點(diǎn)。PSO-BDFRTS在D06、D08、D09的F1-score值優(yōu)于GS-BDFRTS,但在其他樣本中均低于GS-BDFRTS,表明PSO易陷入局部最優(yōu),優(yōu)化能力較差,而本文所提的MOSMA能夠擺脫局部最優(yōu),在全局范圍內(nèi)為BDFRTS搜索到更好的參數(shù)結(jié)構(gòu)。
表3展現(xiàn)了不同方法在低密度軌跡數(shù)據(jù)集中的試驗(yàn)結(jié)果。MOSMA-BDFRTS分割模型在10組低密度軌跡中分割的平均準(zhǔn)確率相比GS-BDFRTS,PSO-BDFRTS模型分別提高了17和14個(gè)百分點(diǎn)。相比GS-BDFRTS和PSO-BDFRTS,MOSMA-BDFRTS在所有的低密度軌跡樣本分割識(shí)別中的農(nóng)田作業(yè)軌跡F1-score值平均提高了17和9個(gè)百分點(diǎn),在道路行駛軌跡的F1-score值平均提升了17和19個(gè)百分點(diǎn)。試驗(yàn)結(jié)果表明,MOSMA-BDFRTS在低密度軌跡數(shù)據(jù)中能夠取得更好的分割結(jié)果。此外,MOSMA-BDFRTS在S01、S08兩項(xiàng)軌跡樣本中相比GS-BDFRTS分別獲得了66與34個(gè)百分點(diǎn)的顯著提升;與PSO-BDFRTS比較,MOSMA-BDFRTS在S02、S04軌跡樣本有明顯的提升,其農(nóng)田作業(yè)軌跡的分割準(zhǔn)確率提升了17與1個(gè)百分點(diǎn),道路行駛軌跡的準(zhǔn)確率分別提升了36與31個(gè)百分點(diǎn),而在其他軌跡樣本中也都取得了更好的優(yōu)化效果,說(shuō)明MOSMA的性能優(yōu)于PSO。表4給出了不同田路分割模型參數(shù)選擇的執(zhí)行時(shí)間,可以看出,本文所提方法的計(jì)算開(kāi)銷(xiāo)明顯低于GS,略低于PSO。軌跡樣本的數(shù)據(jù)量越大,GS計(jì)算所需的時(shí)間就越久,此時(shí)使用MOSMA的計(jì)算收益就越大,因此MOSMA在處理大規(guī)模軌跡數(shù)據(jù)時(shí)更具優(yōu)勢(shì)。綜上,基于MOSMA的參數(shù)優(yōu)化方法能夠在更短的時(shí)間內(nèi)確定田路分割模型的最優(yōu)參數(shù)結(jié)構(gòu),提升模型分割準(zhǔn)確率,具有較好的適應(yīng)性和魯棒性。
圖2展現(xiàn)了三種模型在四組軌跡樣本參數(shù)訓(xùn)練過(guò)程中的收斂曲線,GS不具備元啟發(fā)式算法中常規(guī)意義上的迭代次數(shù)概念,為便于比較,將搜索次數(shù)按比例轉(zhuǎn)換成迭代次數(shù)。從圖2中可以看出,相比GS-BDFRTS和PSO-BDFRTS,MOSMA-BDFRTS能夠在較少的迭代次數(shù)內(nèi)收斂于最優(yōu)參數(shù)結(jié)構(gòu),證明了算法優(yōu)化性能的優(yōu)越性。從圖2整體來(lái)看,MOSMA-BDFRTS收斂曲線的變化呈單調(diào)上升趨勢(shì),且在更少的迭代次數(shù)內(nèi)取到最優(yōu)值,說(shuō)明算法有較好的收斂能力和優(yōu)化速度。
表2 不同田路分割模型在高密度軌跡的分割結(jié)果對(duì)比
注:GS-BDFRTS、PSO-BDFRTS和MOSMA-BDFRTS表示基于方向分布的田路分割模型分別配置網(wǎng)格搜索法、粒子群算法和多元振蕩黏菌算法,下同。
Note: GS-BDFRTS, PSO-BDFRTS and MOSMA-BDFRTS are represent the field-road trajectory segmentation model configures grid search method, particle swarm algorithm and multiplex oscillation slime mould algorithm, respectively, same below.
表3 不同田路分割模型在低密度軌跡的分割結(jié)果對(duì)比
表4 不同參數(shù)選擇方法的耗時(shí)對(duì)比
注:GS、PSO和MOSMA分別表示使用網(wǎng)格搜索法、粒子群算法和多元振蕩黏菌算法選擇模型的參數(shù)。
Note: GS, PSO and MOSMA refers the use grid search method, particle swarm algorithm and multiplex oscillation slime mould algorithm to select parameters for the model.
為更加直觀地分析本文所提方法的優(yōu)化效果,圖3展現(xiàn)了部分軌跡樣本用GS-BDFRTS與MOSMA- BDFRTS進(jìn)行分割的結(jié)果。由表2和表3可知,MOSMA-BDFRTS在D01、D02、S01、S02樣本中取得了明顯地提升,在農(nóng)田作業(yè)軌跡的F1-score值分別提升了13、14、9與14個(gè)百分點(diǎn);在道路行駛軌跡的F1-score值分別比GS-BDFRTS提高了68、69、66與5個(gè)百分點(diǎn)。從圖3展現(xiàn)的軌跡分割圖像可以清晰地看出,經(jīng)MOSMA-BDFRTS分割后的軌跡分類(lèi)在相應(yīng)的地圖背景下的準(zhǔn)確率更高,說(shuō)明本文所提方法能夠提供更優(yōu)越的分割結(jié)果。
a. 軌跡樣本D01a. Trajectory sample D01b. 軌跡樣本D02b. Trajectory sample D02c. 軌跡樣本S01c. Trajectory sample S01d. 軌跡樣本S02d. Trajectory sample S02
圖3 不同田路分割模型的分割結(jié)果
為了提升田路分割模型的性能,本研究從參數(shù)優(yōu)化的角度切入,提出了一種面向田路分割模型的參數(shù)優(yōu)化方法。該方法以最大化分割準(zhǔn)確率為優(yōu)化目標(biāo),為田路分割模型建立了一種通用的參數(shù)優(yōu)化方案,其能夠使用元啟發(fā)式算法求解優(yōu)化方案來(lái)自動(dòng)標(biāo)定模型的參數(shù)結(jié)構(gòu)。為了解決傳統(tǒng)元啟發(fā)式算法因性能低下求解方案困難的問(wèn)題,提出了一種多元振蕩黏菌算法,該算法提出了動(dòng)態(tài)引導(dǎo)策略與多元振蕩策略提升了優(yōu)化性能。在20組不同密度的農(nóng)機(jī)作業(yè)軌跡數(shù)據(jù)中開(kāi)展試驗(yàn),結(jié)果表明,基于多元振蕩黏菌算法的參數(shù)優(yōu)化方法能有效提升田路分割模型的性能。在分割結(jié)果方面,多元振蕩黏菌算法優(yōu)化的田路分割模型相比網(wǎng)格搜索法和粒子群算法優(yōu)化的田路分割模型在高密度軌跡數(shù)據(jù)中的分割準(zhǔn)確率分別提高了25和28個(gè)百分點(diǎn);在低密度軌跡數(shù)據(jù)中的分割準(zhǔn)確率分別提高了17和14個(gè)百分點(diǎn)。從分割效果和平均處理時(shí)間等指標(biāo)綜合來(lái)看,基于多元振蕩黏菌算法的參數(shù)優(yōu)化方法能夠有效地提升田路分割模型的性能。本文所提方法為田路分割模型提供了一種通用的參數(shù)優(yōu)化解決方案,可直接應(yīng)用于其他類(lèi)型的模型實(shí)例中。
[1] 肖敬,劉卉,魏學(xué)禮,等. 基于時(shí)空立方體的農(nóng)機(jī)運(yùn)動(dòng)軌跡分割方法[J]. 江蘇農(nóng)業(yè)科學(xué),2018,46(20):244-247.
Xiao Jing, Liu Hui, Wei Xueli, et al. Segmentation of agricultural machinery trajectories based on space-time cube[J]. Jiangsu Agricultural Sciences, 2018, 46(20): 244-247. (in Chinese with English abstract)
[2] 孫忠富,杜克明,鄭飛翔,等. 大數(shù)據(jù)在智慧農(nóng)業(yè)中研究與應(yīng)用展望[J]. 中國(guó)農(nóng)業(yè)科技導(dǎo)報(bào),2013,15(6):63-71.
Sun Zhongfu, Du Kengming, Zheng Feixiang, et al. Perspectives of research and application of big data on smart agriculture[J]. Journal of Agricultural Science and Technology, 2013, 15(6): 63-71. (in Chinese with English abstract)
[3] 馮曉莉,王永興,仇寶云. 基于混合狼群算法參數(shù)優(yōu)選的泵站群運(yùn)行優(yōu)化[J]. 農(nóng)業(yè)工程學(xué)報(bào),2020,36(3):30-36.
Feng Xiaoli, Wang Yongxing, Chou Baoyun. Optimization of pumping station group operation based on optimized parameters using hybrid wolf pack algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(3): 30-36. (in Chinese with English abstract)
[4] 王培,孟志軍,尹彥鑫,等. 基于農(nóng)機(jī)空間運(yùn)行軌跡的作業(yè)狀態(tài)自動(dòng)識(shí)別試驗(yàn)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2015,31(3):56-61.
Wang Pei, Meng Zhijun, Yin Yanxin, et al. Automatic recognition algorithm of field operation status based on spatial track of agricultural machinery and corresponding experiment[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(3): 56-61. (in Chinese with English abstract)
[5] 劉卉,孟志軍,王培,等. 基于農(nóng)機(jī)空間軌跡的作業(yè)面積的緩沖區(qū)算法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2015,31(7):180-184.
Liu Hui, Meng Zhijun, Wang Pei, et al. Buffer algorithms for operation area measurement based on global navigation satellite system trajectories of agricultural machinery[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(7): 180-184. (in Chinese with English abstract)
[6] 吳才聰,蔡亞平,羅夢(mèng)佳,等. 基于時(shí)間窗的農(nóng)機(jī)資源時(shí)空調(diào)度模型[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2013,44(5):237-241,231.
Wu Caicong, Cai Yaping, Luo Mengjia, et al. Time-windows based temporal and spatial scheduling model for agricultural machinery resources[J]. Transactions of the Chinese Society for Agricultural Machinery, 2013, 44(5): 237-241, 231. (in Chinese with English abstract)
[7] 翟衛(wèi)欣,王東旭,陳智博,等. 無(wú)人駕駛農(nóng)機(jī)自主作業(yè)路徑規(guī)劃方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(16):1-7.
Zhai Weixin, Wang Dongxu, Chen Zhibo, et al. Autonomous operation path planning method for unmanned agricultural machinery[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(16): 1-7. (in Chinese with English abstract)
[8] 翟志強(qiáng),王秀倩,王亮,等. 面向主從跟隨協(xié)同作業(yè)的導(dǎo)航路徑規(guī)劃方法[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2021,52(S1):542-547.
Zhai Zhiqiang, Wang Xiuqian, Wang Liang, et al. Collaborative path planning for autonomous agricultural machinery of master-slave cooperation[J]. Transactions of the Chinese Society for Agricultural Machinery, 2021, 52(S1): 542-547. (in Chinese with English abstract)
[9] 姚竟發(fā),劉靜,張璠,等. 基于多普勒與貪心策略的農(nóng)機(jī)作業(yè)路徑優(yōu)化研究[J]. 中國(guó)農(nóng)機(jī)化學(xué)報(bào),2020,41(4):130-137.
Yao Jingfa, Liu Jing, Zhang Fan, et al. Optimization of agricultural machinery operation path based on doppler and greedy strategy[J]. Journal of Chinese Agricultural Mechanization, 2020, 41(4): 130-137. (in Chinese with English abstract)
[10] Kortenbruck D, Griepentrog H W, Paraforos D S. Machine operation profiles generated from ISO 11783 communication data[J]. Computers and Electronics in Agriculture, 2017, 140: 227-236.
[11] Hanke S, Tr?sken L, Frerichs L. Development and parameterization of an object-oriented model for describing agricultural process steps[J]. Landtechnik, 2018, 73(2): 22-36.
[12] Kilic T, Zezza A, Carletto C, et al. Missing (ness) in action: selectivity bias in GPS-based land area measurements[J]. World Development, 2017, 92: 143-157.
[13] Chen Y, Zhang X, Wu C, et al. Field-road trajectory segmentation for agricultural machinery based on direction distribution[J]. Computers and Electronics in Agriculture, 2021, 186: 106180.
[14] Bertolini M, Mezzogori D, Neroni M, et al. Machine Learning for industrial applications: A comprehensive literature review[J]. Expert Systems with Applications, 2021, 175: 114820.
[15] Ding X, Zheng M, Zheng X. The application of genetic algorithm in land use optimization research: A review[J]. Land, 2021, 10(5): 526.
[16] Houssein E H, Gad A G, Hussain K, et al. Major advances in particle swarm optimization: theory, analysis, and application[J]. Swarm and Evolutionary Computation, 2021, 63: 100868.
[17] Yu X, Xu W, Li C. Opposition-based learning grey wolf optimizer for global optimization[J]. Knowledge-Based Systems, 2021, 226: 107139.
[18] Dhiman G, Singh K K, Soni M, et al. MOSOA: A new multi-objective seagull optimization algorithm[J]. Expert Systems with Applications, 2021, 167: 114150.
[19] Tam N T, Hung T H, Binh H T T. A decomposition-based multi-objective optimization approach for balancing the energy consumption of wireless sensor networks[J]. Applied Soft Computing, 2021, 107: 107365.
[20] Dhal P, Azad C. A multi-objective feature selection method using newton’s law based pso with gwo[J]. Applied Soft Computing, 2021, 107: 107394.
[21] D’Angelo G, Palmieri F. GGA: A modified genetic algorithm with gradient-based local search for solving constrained optimization problems[J]. Information Sciences, 2021, 547: 136-162.
[22] Kapoor K, Rani S, Kumar M, et al. Hybrid local phase quantization and grey wolf optimization based SVM for finger vein recognition[J]. Multimedia Tools and Applications, 2021, 80(10): 15233-15271.
[23] Xu L, Hou L, Zhu Z, et al. Mid-term prediction of electrical energy consumption for crude oil pipelines using a hybrid algorithm of support vector machine and genetic algorithm[J]. Energy, 2021, 222: 119955.
[24] 張良安,唐鍇,趙永杰,等. 四足激光除草機(jī)器人腿部結(jié)構(gòu)參數(shù)優(yōu)化[J]. 農(nóng)業(yè)工程學(xué)報(bào),2020,36(2):7-15.
Zhang Liang’an, Tang Kai, Zhao Yongjie, et al. Optimization of leg structure parameter of quadruped laser weeding robot[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(2): 7-15. (in Chinese with English abstract)
[25] Ding W L, Zhao Y L, Xin W T, et al. Parameter extraction method of virtual plant growth model based on Improved Particle Swarm Optimization[J]. Computers and Electronics in Agriculture, 2021, 191: 106470.
[26] Jamei M, Elbeltagi A, Maroufpoor S, et al. Combined terrestrial evapotranspiration index prediction using a hybrid artificial intelligence paradigm integrated with relief algorithm-based feature selection[J]. Computers and Electronics in Agriculture, 2022, 193: 106687.
[27] Ester M, Kriegel H P, Sander J, et al. A density-based algorithm for discovering clusters in large spatial databases with noise[C]// Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. Portland, America: ACM, 1996: 226-231.
[28] Li S, Chen H, Wang M, et al. Slime mould algorithm: A new method for stochastic optimization[J]. Future Generation Computer Systems, 2020, 111: 300-323.
[29] 王鵬新,齊璇,李俐,等. 基于粒子群優(yōu)化投影尋蹤的玉米單產(chǎn)估測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(13):145-153.
Wang Pengxin, Qi Xuan, Li Li, et al. Estimation of maize yield based on projection pursuit with particle swarm optimization[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(13): 145-153. (in Chinese with English abstract)
Parameter optimization of field-road trajectory segmentation model using multiplex oscillation slime mould algorithm
Zhai Weixin1,2, Pan Jiawen1,2, Lan Yubin3,4, Wu Caicong1,2※
(1.,,100083,;2.,,100083,;3.,,255000,;4.,510642,)
Field-Road Trajectory Segmentation (FRTS) is one of the important tasks of agricultural machinery. A sequence of field-road segments of a trajectory can be automatically divided for the big data in precision agriculture. The parameter of the FRTS model can also determine the segmentation accuracy and precision. However, the traditional parameter selection cannot obtain the superior solution, limiting the segmentation performance of the model. Therefore, this study aims to investigate the performance improvement of the FRTS model from the perspective of parameter optimization. Two aspects were mainly contained as follows. Firstly, the metaheuristic algorithms were used to determine the parameter configuration of the model. The classification accuracy was considered as an objective to transform the parameter into a single-objective optimization. Specifically, the parameter structure of the model was abstracted as the searched individual of the optimization. The reasonable fitness function was set, according to the metrics of FRTS evaluation. Then, the fitness was used to evaluate the search of the individual in the solution space. The location of the searched individual was also continuously adjusted, according to the calculation rules of the optimization. As such, the global optimal parameter structure was achieved to converge. Secondly, a Multiplex Oscillation Slime Mould Algorithm (MOSMA) was proposed to realize the parameter optimization with the nonlinear characteristics and multiple locally optimal solutions. A dynamic guidance strategy was also established to adaptively change the individual movement for the better exploitation capability of the model, according to the search process of the population. Then, a strategy (called multivariate oscillation) was proposed to improve the segmentation performance and exploration capability of the model. Different search paths were utilized to produce multiple oscillations before the individual moves, and the priori rule was then to evaluate the qualities of paths. As such, the path with the highest quality was selected to move. The synergy of the two strategies enhanced the optimization capability of the model. Dynamic guidance and a multiplex oscillation strategy enhanced the oscillation contraction patterns of the slime mould and the process of the cytoplasm flows, thereby improving the optimization performance of the model. The experiments were also performed on real agricultural trajectory datasets with different sampling frequencies. A comparison was then made with the Grid Search (GS) and Particle Swarm Optimization (PSO) to validate the effectiveness of the model. The experiment results show that the new optimization effectively improved the accuracy and performance of the FRTS model using direction distribution (BDFRTS). The average accuracy of the MOSMA-BDFRTS on high-density trajectory data was increased by 25 percentage points and 28 percentage points compared with GS-BDFRTS and PSO-BDFRTS. MOSMA-BDFRTS achieved more competitive results on low-density trajectory data, whose average accuracy of segmentation was improved by 17 percentage points and 14 percentage points compared with GS-BDFRTS and PSO-BDFRTS. The proposed method provides a generalized parameter optimization solution for field-road trajectory segmentation models, and it can be applied directly to other types of model instances. The study also provides a reference for the research of the agricultural machinery movement trajectory segmentation technology.
model; parameter optimization; field-road trajectory segmentation; slime mould algorithm; dynamic guidance; multiplex oscillation
10.11975/j.issn.1002-6819.2022.18.019
S126
A
1002-6819(2022)-18-0176-08
翟衛(wèi)欣,潘家文,蘭玉彬,等. 基于多元振蕩黏菌算法的田路分割模型參數(shù)優(yōu)化方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(18):176-183.doi:10.11975/j.issn.1002-6819.2022.18.019 http://www.tcsae.org
Zhai Weixin, Pan Jiawen, Lan Yubin, et al. Parameter optimization of field-road trajectory segmentation model using multiplex oscillation slime mould algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(18): 176-183. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2022.18.019 http://www.tcsae.org
2022-06-26
2022-09-12
國(guó)家精準(zhǔn)農(nóng)業(yè)應(yīng)用項(xiàng)目(JZNYYY001);北京市科技計(jì)劃項(xiàng)目(Z201100008020008);中國(guó)科協(xié)科技智庫(kù)青年人才計(jì)劃項(xiàng)目(20220615ZZ07110141)
翟衛(wèi)欣,博士,副教授,博士生導(dǎo)師,研究方向?yàn)闀r(shí)空大數(shù)據(jù)、地圖學(xué)與地理信息系統(tǒng)、智能農(nóng)機(jī)等。Email:zhaiweixin@cau.edu.cn
吳才聰,博士,教授,博士生導(dǎo)師,研究方向?yàn)檗r(nóng)機(jī)導(dǎo)航與位置服務(wù)等。Email:wucc@cau.edu.cn
中國(guó)農(nóng)業(yè)工程學(xué)會(huì)會(huì)員:翟衛(wèi)欣(E041201339S)