張 玨,田海清,趙志宇,張麗娜,張 晶,李 斐
?
基于改進(jìn)離散粒子群算法的青貯玉米原料含水率高光譜檢測(cè)
張 玨1,2,田海清1※,趙志宇1,張麗娜2,張 晶1,李 斐3
(1. 內(nèi)蒙古農(nóng)業(yè)大學(xué)機(jī)電工程學(xué)院,呼和浩特 010018;2. 內(nèi)蒙古師范大學(xué)物理與電子信息學(xué)院,呼和浩特 010020; 3. 內(nèi)蒙古農(nóng)業(yè)大學(xué)草原與資源環(huán)境學(xué)院,呼和浩特 010019)
快速、無(wú)損和準(zhǔn)確檢測(cè)青貯玉米原料含水率,對(duì)確保青貯玉米發(fā)酵品質(zhì)、推動(dòng)青貯產(chǎn)業(yè)健康快速發(fā)展有重要現(xiàn)實(shí)意義。為探究高光譜技術(shù)在青貯玉米原料含水率檢測(cè)方面的可行性,研究通過(guò)高光譜成像系統(tǒng)獲取青貯玉米原料高光譜圖像并利用烘箱加熱法測(cè)定實(shí)際含水率。在粒子更新方式和慣性權(quán)重2個(gè)方面對(duì)傳統(tǒng)離散粒子群算法(discrete binary particle swarm optimization,DBPSO)進(jìn)行優(yōu)化,提出基于改進(jìn)型離散粒子群算法(modified discrete binary particle swarm optimization,MDBPSO)的特征波段優(yōu)選方法,并利用相關(guān)系數(shù)分析法(correlation coefficient,CC)、DBPSO和MDBPSO法提取原料含水率高光譜特征變量,基于全波段反射光譜(total spectral reflectance,TSR)和特征波段反射光譜建立青貯玉米原料含水率預(yù)測(cè)模型。結(jié)果表明,MDBPSO優(yōu)選特征波段適應(yīng)度函數(shù)的收斂精度和收斂效率較DBPSO法均有明顯改善,最優(yōu)適應(yīng)度值由0.761 6提高至0.812 3,函數(shù)收斂迭代次數(shù)由280次降低至79次。MDBPSO-PLSR預(yù)測(cè)模型的建模精度和預(yù)測(cè)精度均高于CC-PLSR、DBPSO-PLSR和TSR-PLSR預(yù)測(cè)模型,其校正集決定系數(shù)R2和均方根誤差RMSEC (root mean square error of calibration)分別為0.81和0.032,預(yù)測(cè)集決定系數(shù)R2和均方根誤差RMSEP(root mean square error of prediction)分別為0.80和0.045。該研究表明,利用高光譜圖像技術(shù)檢測(cè)青貯玉米原料含水率具有較高的精度,研究可為后續(xù)開發(fā)青貯玉米原料水分快速檢測(cè)儀器提供借鑒方法。
粒子; 水分; 光譜分析; 高光譜; 粒子群; 青貯玉米; 特征波段
青貯玉米是把玉米乳熟后期至蠟熟期收獲的地上部分植株揉碎切短后,經(jīng)過(guò)加工、密封、貯藏發(fā)酵后制成的1種營(yíng)養(yǎng)豐富的飼料[1]。青貯玉米原料含水率影響整個(gè)青貯過(guò)程的排汁、壓實(shí)程度以及微生物活動(dòng),進(jìn)而影響青貯玉米飼料發(fā)酵品質(zhì)[2],因此,準(zhǔn)確、快速地進(jìn)行原料含水率檢測(cè)具有重要意義。
傳統(tǒng)含水率檢測(cè)通常采用電烘箱加熱、紅外加熱、微波加熱等物理干燥方法[3],該方法檢測(cè)精度較高,但存在耗時(shí)費(fèi)力、過(guò)程冗長(zhǎng)繁瑣、對(duì)樣本有損、時(shí)效性差等弊端[4],且難以滿足生產(chǎn)實(shí)踐中對(duì)大范圍玉米原料含水率實(shí)時(shí)、無(wú)損的檢測(cè)需求。利用光譜技術(shù)進(jìn)行含水率檢測(cè)具有快捷、無(wú)損的優(yōu)點(diǎn),而傳統(tǒng)的光譜監(jiān)測(cè)方法均采用非成像光譜儀進(jìn)行定量分析,該方法只局限于“點(diǎn)”尺度的研究,而以點(diǎn)代面采樣方式無(wú)法準(zhǔn)確反映整個(gè)樣本真實(shí)信息[5]。
高光譜圖像包含面層次的光譜信息與波段豐富的圖像信息,目前國(guó)內(nèi)外已有學(xué)者基于高光譜成像技術(shù)針對(duì)作物含水率進(jìn)行了無(wú)損檢測(cè)研究[6-10]。在應(yīng)用高光譜圖像技術(shù)進(jìn)行參數(shù)反演時(shí),由于高光譜數(shù)據(jù)存在信息量大、波段多且相鄰波段信息相關(guān)性高等特點(diǎn)[11],特征變量的有效提取則成為模型預(yù)測(cè)效果優(yōu)劣的關(guān)鍵環(huán)節(jié)。如何進(jìn)行光譜特征變量的有效提取已經(jīng)成為當(dāng)前研究熱點(diǎn),如孫俊等[12]采用競(jìng)爭(zhēng)性自適應(yīng)加權(quán)算法進(jìn)行高光譜特征波段選擇,建立基于人工蜂群算法的油麥菜葉片水分含量預(yù)測(cè)模型,研究結(jié)果顯示,該模型預(yù)測(cè)集決定系數(shù)2和均方根誤差RMSE分別為0.921 4和2.95%。Digman等[13]將一臺(tái)可移動(dòng)的光譜儀集成到自動(dòng)牧草收割機(jī)噴口中,現(xiàn)場(chǎng)采集牧草近紅外反射光譜及水分含量,利用主成分分析法提取光譜特征變量,分別建立全株青貯玉米和紫花苜蓿含水率模型,研究發(fā)現(xiàn),兩種牧草交叉驗(yàn)證均方根誤差RMSECV分別為3.3%和3.7%。Cozzolino等[14]采集400~2 500 nm青貯玉米樣品的反射光譜,并利用標(biāo)準(zhǔn)正態(tài)變量結(jié)合去趨勢(shì)法、多元散射校正等光譜預(yù)處理方法建立青貯玉米的干物質(zhì)、粗蛋白、酸性和中性洗滌纖維等品質(zhì)參數(shù)的偏最小二乘預(yù)測(cè)模型,研究表明,各品質(zhì)參數(shù)預(yù)測(cè)模型決定系數(shù)2均在0.8以上。Zhou等[15]利用第4層小波分解提取特征波段建立萵苣葉片含水率偏最小二乘回歸預(yù)測(cè)模型,有效實(shí)現(xiàn)了萵苣葉片含水率定量檢測(cè)及分布可視化。
上述研究主要利用原始光譜或衍生變量進(jìn)行特征變量選擇,多采用“單向”方式提取特征變量,特征變量只會(huì)單方面影響模型的反演精度,預(yù)測(cè)結(jié)果的優(yōu)劣卻不能干預(yù)特征變量的選擇,從而限制了光譜信息的有效提取,導(dǎo)致模型反演精度降低。因此,研究一種有效的“反饋型”特征變量智能提取方法,將模型的反演精度作為特征變量的提取標(biāo)準(zhǔn),利用少數(shù)關(guān)鍵變量代替全譜段信息,既可降低模型運(yùn)算量和復(fù)雜度,又可提高模型穩(wěn)定性和準(zhǔn)確性。
本文利用高光譜圖像系統(tǒng)采集青貯玉米原料反射光譜信息,并對(duì)原料反射光譜進(jìn)行標(biāo)準(zhǔn)正態(tài)變量校正。為提高特征波段提取效率,對(duì)經(jīng)典離散粒子群算法(discrete binary particle swarm optimization,DBPSO)進(jìn)行改進(jìn),建立基于改進(jìn)型離散粒子群算法(modified discrete binary particle swarm optimization, MDBPSO)優(yōu)選特征波段的原料含水率偏最小二乘回歸[16](partial least-squares regression, PLSR)模型。對(duì)比采用相關(guān)系數(shù)分析法(correlation coefficient, CC)和DBPSO法優(yōu)選出的特征波段對(duì)PLSR模型精度的影響,研究基于高光譜數(shù)據(jù)源的MDBPSO- PLSR模型在青貯玉米原料含水率反演中的適用性。
本試驗(yàn)以青貯玉米原料為研究對(duì)象,樣品在2017年8月購(gòu)于內(nèi)蒙古呼和浩特市托克托縣,玉米品種為佳良99號(hào)。將采集到的青貯玉米原料置于真空袋,放入冰箱冷藏,共制備樣品238份。采用熱烘箱加熱法測(cè)定原料含水率,依據(jù)《飼料中水分的測(cè)定》[17]測(cè)定含水率。具體步驟如下:
1)將稱量瓶放入103 ℃的干燥箱中,干燥30 min后取出,在干燥器中冷卻至室溫,記錄稱量瓶質(zhì)量1;
2)稱取5 g(2)青貯玉米原料,置于已烘干至恒質(zhì)量的稱量瓶?jī)?nèi),放入103 ℃的干燥箱4 h,取出后將稱量瓶放在干燥器內(nèi)冷卻至室溫,記錄稱量瓶和試料干燥后質(zhì)量;
3)將稱量瓶再放入干燥箱1 h,冷卻稱其質(zhì)量,如此重復(fù),直至稱量值的變化小于試料質(zhì)量的0.1%。記錄稱量瓶和試料干燥后最終質(zhì)量3。
青貯玉米原料含水率計(jì)算公式如式(1)。
式中1為稱量瓶質(zhì)量,g;2為試料質(zhì)量,g;3為稱量瓶和試料干燥后最終質(zhì)量,g。
采用上述烘干法得到的青貯玉米原料整體含水率范圍為26.09%~82.91%,平均含水率為67.02%,樣本含水率值主要集中在65.62%~78.94%區(qū)間,占總樣本數(shù)的78.65%。
試驗(yàn)采用五鈴光學(xué)(ISUZU OPTICS)高光譜成像系統(tǒng),系統(tǒng)主要包括:高光譜圖像光譜儀(ImSpector V10E型,Spectral Imaging Ltd, Oulu,芬蘭)、CCD相機(jī)(IGV- B1620型,Imperx,美國(guó))、2個(gè)150 W的鹵素?zé)簦?900型,Illuminatior,Illumination科技)、1個(gè)直流可調(diào)節(jié)光源(2900型,Illumination,美國(guó))、移動(dòng)控制平臺(tái)(IRCP0076-1 COM, 臺(tái)灣)和計(jì)算機(jī)等部件組成。高光譜攝像機(jī)光譜范圍為383~1 004 nm,光譜分辨率為2.8 nm。
試驗(yàn)開始前,將系統(tǒng)預(yù)熱30 min以消除基線漂移的影響,然后對(duì)光譜儀進(jìn)行調(diào)焦。設(shè)置系統(tǒng)曝光時(shí)間為7.6 ms,電控移動(dòng)平臺(tái)速度為5.04 mm/s,高光譜圖像采集時(shí),將樣本置于容器內(nèi)攤平壓實(shí)放在移動(dòng)平臺(tái)上,圖像分辨率選擇800×428像素,通過(guò)高光譜圖像采集軟件得到383~100 4 nm范圍428個(gè)波段下的樣本高光譜圖像。
為減弱光照不均勻及相機(jī)暗電流對(duì)光譜信息的影響,在數(shù)據(jù)處理前對(duì)高光譜圖像進(jìn)行黑白校正[18]。于采集青貯玉米原料樣本相同的環(huán)境條件下,借助反射率為99%的標(biāo)準(zhǔn)白色校正板采集全白標(biāo)定圖像光譜,封閉鏡頭采集全黑標(biāo)定圖像光譜,對(duì)原始圖像光譜按照式(2)進(jìn)行校正[19]。
式中為黑白校正后樣本光譜反射率;I為原始樣本反射的光譜強(qiáng)度;I為標(biāo)準(zhǔn)校正黑板反射的光譜強(qiáng)度;I為標(biāo)準(zhǔn)校正白板反射的光譜強(qiáng)度。
使用ENVI 5.3軟件對(duì)采集到的高光譜圖像進(jìn)行光譜提取,設(shè)定葉片偽彩色圖像選擇譜的通道R為650 nm,G為550 nm,B為450 nm,避開樣本邊緣、反光嚴(yán)重及暗黑區(qū)域,手動(dòng)選取大小為20×20像素的正方形區(qū)域作為感興趣區(qū)域(region of interesting,ROI),區(qū)域內(nèi)包括圖1a所示的全株玉米老葉、新葉、嫩葉、秸稈表皮、破碎秸稈及籽粒等代表性部位。通過(guò)計(jì)算ROI內(nèi)所有像素點(diǎn)的平均值,最終得到每個(gè)樣本的平均高光譜數(shù)據(jù),圖1b為青貯玉米原料平均反射曲線。
標(biāo)準(zhǔn)正態(tài)變量校正(standard normal variate,SNV)方法是假設(shè)每個(gè)波段的光譜值均滿足正態(tài)分布等標(biāo)準(zhǔn)數(shù)據(jù)分布,利用假設(shè)數(shù)據(jù)分布信息對(duì)已知光譜進(jìn)行修正補(bǔ)償,用于消除或減弱粒子散射對(duì)光譜數(shù)據(jù)產(chǎn)生的影響[20]。SNV變換公式如式(3)。
1.5.1 相關(guān)系數(shù)(correlation coefficient,CC)分析法
CC分析法是一種常用的特征波段提取方法[21],將校正集光譜陣中的每個(gè)波長(zhǎng)對(duì)應(yīng)的吸光度向量與樣本矩陣向量進(jìn)行相關(guān)性計(jì)算,得到波長(zhǎng)-相關(guān)系數(shù)的變化曲線,再選擇一定閾值范圍的極值點(diǎn)作為敏感波長(zhǎng)。
1.5.2 離散粒子群優(yōu)化算法及改進(jìn)
粒子群算法是由J.Kennedy和R.C.Eberhart受鳥群覓食過(guò)程中的行為特征啟發(fā),于1995年提出來(lái)的1種群體智能隨機(jī)搜索算法[22]。假設(shè)由若干粒子構(gòu)成的1個(gè)種群在維空間搜索最優(yōu)位置,每個(gè)粒子由其速度和位置2方面向量信息表示,第個(gè)粒子的速度和位置分別表示為v=(v1,v2,…,v),x=(x1,x2,…,x)。
算法采用適應(yīng)度函數(shù)評(píng)價(jià)粒子當(dāng)前位置的優(yōu)劣,經(jīng)過(guò)多次迭代后,找到最優(yōu)解或近似最優(yōu)解。粒子在其第(1≤≤)維的位置和速度的更新方式見式(4)、(5)。
式中為慣性權(quán)重;1、2為學(xué)習(xí)因子;1、2為0~1隨機(jī)數(shù);1、1分別為粒子在+1次迭代更新后的速度和位置;、分別為粒子在次迭代后的速度和位置;為第次搜索時(shí)粒子的歷史最優(yōu)解對(duì)應(yīng)位置;為第次搜索時(shí)所有粒子全局最優(yōu)解對(duì)應(yīng)位置。
離散粒子群算法[23]的每個(gè)粒子均通過(guò)二進(jìn)制編碼表示,粒子速度決定粒子位置取0或1的概率,利用sigmoid函數(shù)將速度映射到[0, 1]區(qū)間計(jì)算對(duì)應(yīng)位置狀態(tài)的概率,(1)表示粒子位置取1的概率,(1)與粒子速度1的數(shù)學(xué)關(guān)系見式(6)。速度越大,粒子對(duì)應(yīng)位置為1的概率越大,反之,對(duì)應(yīng)位置為1的概率則越小。此時(shí)粒子速度更新公式不變,位置依據(jù)式(7)更新。
式中rand( )是產(chǎn)生(0, 1)隨機(jī)數(shù)的函數(shù)。
DBPSO算法依靠群體之間的合作與競(jìng)爭(zhēng)來(lái)迭代,一旦有粒子發(fā)現(xiàn)當(dāng)前最優(yōu)位置,其它粒子迅速向其靠攏,當(dāng)粒子速度接近零時(shí),種群多樣性會(huì)逐漸喪失,粒子群陷入局部最優(yōu)鄰域后停止搜索其它區(qū)域,從而容易陷入局部最優(yōu),發(fā)生“早熟”收斂[24]。為克服傳統(tǒng)DBPSO算法的上述劣勢(shì),保證優(yōu)選特征波段更具針對(duì)性且更為有效,本文提出MDBPSO算法,分別從粒子位置更新方式和慣性權(quán)重2個(gè)方面對(duì)傳統(tǒng)DBPSO法進(jìn)行改進(jìn),具體思路如下:
1)動(dòng)態(tài)調(diào)整粒子位置
為避免函數(shù)(1)靠近端點(diǎn)值出現(xiàn)“飽和現(xiàn)象”,須限定粒子飛行速度最大值max,將速度限定在[-max,max]區(qū)間范圍內(nèi)。當(dāng)粒子接近最優(yōu)解時(shí),粒子速度1將趨向于0,(1)值接近0.5,此時(shí)算法按照純隨機(jī)性模式搜索,局部搜索能力變差,收斂效率降低。為克服傳統(tǒng)算法的上述缺陷,在迭代后期對(duì)(7)式描述的粒子位置更新方式做如下改進(jìn):
設(shè)定粒子飛行速度[25]最大值max為4,當(dāng)速度處于正負(fù)邊界時(shí),概率映射函數(shù)(1)則分別為0.982和0.018,函數(shù)值分別接近1和0;當(dāng)速度介于-max和max時(shí),概率映射函數(shù)為減函數(shù),在(0,1)區(qū)間范圍取值,特別地,當(dāng)1為0時(shí),映射函數(shù)(1)值為0.5。根據(jù)上述分析,sigmond函數(shù)和位置更新依照式(8)、(9)更新。
注設(shè)定邊界條件后,當(dāng)1≥4時(shí),(1)為0.982;當(dāng)1≤-4時(shí),(1)為0.018。
在粒子接近最優(yōu)解時(shí),MDBPSO算法依據(jù)式(9)對(duì)粒子位置進(jìn)行更新,壓縮了粒子的運(yùn)動(dòng)空間,使粒子運(yùn)動(dòng)范圍相對(duì)變窄,增強(qiáng)了算法的局部搜索能力,有利于算法實(shí)現(xiàn)快速收斂。
2)改進(jìn)慣性權(quán)重
慣性權(quán)重[26]對(duì)平衡算法的全局搜索能力和局部搜索能力有顯著作用,較大的慣性權(quán)重可加快粒子飛行速度,提高算法的全局搜索能力,但收斂性相對(duì)降低;較小的可減小粒子飛行速度,提高算法收斂效率,但容易陷入局部極值。
鑒于算法在運(yùn)行前期注重全局搜索,后期需盡快收斂,現(xiàn)對(duì)進(jìn)行線性遞減動(dòng)態(tài)調(diào)整,調(diào)整方式見式(10)。
式中max、min分別為的最大值與最小值,通常情況下,max和min分別取1.2和0.9[27];為當(dāng)前迭代次數(shù);max為最大迭代次數(shù)。
1.5.3 MDBPSO算法提取特征波段
1)粒子編碼設(shè)計(jì)
MDBPSO算法對(duì)特征波段選擇相當(dāng)于粒子編碼的過(guò)程,即把每個(gè)波長(zhǎng)定義為粒子的一維離散二進(jìn)制變量,粒子的長(zhǎng)度與光譜數(shù)據(jù)維數(shù)相同。對(duì)每個(gè)粒子,其取值可能為1或0,1表示相應(yīng)波長(zhǎng)被選中,反之表明該波長(zhǎng)未被選中,每個(gè)粒子的飛行位置即可代表波段選擇的1個(gè)解。對(duì)任何1種波長(zhǎng)組合,存在唯一的特征向量與之對(duì)應(yīng)。
2)適應(yīng)度函數(shù)設(shè)計(jì)
特征選擇的目的是找出預(yù)測(cè)能力最強(qiáng)的特征組合,因此需要一個(gè)定量準(zhǔn)則來(lái)度量特征組合的預(yù)測(cè)能力[28]。本文使用PLSR模型校正集決定系數(shù)2作為評(píng)判特征波段適用性標(biāo)準(zhǔn),根據(jù)式(11)構(gòu)造適應(yīng)度函數(shù),依照算法所處周期進(jìn)行分段優(yōu)化,提高搜索效率,實(shí)現(xiàn)波段的合理、高效選擇。
3)算法流程
MDBPSO算法流程如圖2所示,其搜索步驟簡(jiǎn)述如下:
①確定粒子群基本參數(shù),包括種群大小、學(xué)習(xí)因子1和2、慣性權(quán)重max和min和最大迭代次數(shù)max;
③粒子個(gè)體適應(yīng)度評(píng)估:對(duì)粒子進(jìn)行解碼,得到粒子個(gè)體對(duì)應(yīng)的特征變量的解,將其作為PLSR模型的輸入因子,并將校正集樣本實(shí)測(cè)值和預(yù)測(cè)值的決定系數(shù)作為粒子的適應(yīng)度函數(shù)值;
④根據(jù)粒子個(gè)體和種群歷史最優(yōu)適應(yīng)值,更新個(gè)體粒子歷史最優(yōu)位置和全局歷史最優(yōu)位置;
⑤計(jì)算迭代次數(shù)=+1,更新粒子速度,并根據(jù)確定粒子位置更新方式;
⑥判斷終止條件:若max,則跳轉(zhuǎn)到步驟(3);若=max,則終止迭代,對(duì)全局粒子最優(yōu)位置進(jìn)行解碼,得到特征波段提取結(jié)果。
上述流程中,為兼顧算法全局搜索能力及局部收斂效率,MDBPSO算法采用2種方式對(duì)粒子位置進(jìn)行更新:當(dāng)30%max,采用式(4)、(6)、(7)對(duì)粒子速度和位置進(jìn)行更新;為保證算法收斂效率,當(dāng)30%max 注:t為當(dāng)前迭代次數(shù);tmax為最大迭代次數(shù)。 為減少樣本結(jié)構(gòu)背景噪聲、表面紋理等因素的影響,對(duì)采集的青貯原料高光譜圖像原始光譜進(jìn)行SNV預(yù)處理,圖3為預(yù)處理后原始平均光譜響應(yīng)曲線。由圖3可知,青貯玉米原料光譜反射率在383~635 nm波段呈先降后升的趨勢(shì),隨后譜線緩慢下降,于680 nm附近形成“紅谷”。在680~780 nm的紅邊區(qū)域,光譜反射率迅速增加,并于780~890 nm近紅外波段形成1個(gè)較高的反射平臺(tái),隨后譜線下降至965 nm處再次出現(xiàn)吸收谷。 圖3 SNV預(yù)處理后青貯玉米原料樣本高光譜響應(yīng)曲線 鑒于高光譜數(shù)據(jù)維數(shù)和冗余度較高,且波段間存在較強(qiáng)的相關(guān)性,需對(duì)全波段光譜進(jìn)行降維處理。研究分別采用CC分析法、傳統(tǒng)DBPSO算法和MDBPSO算法進(jìn)行青貯玉米原料含水率特征波段的提取。 2.2.1 CC分析法提取特征波段 采用CC分析法對(duì)樣本反射光譜與原料含水率進(jìn)行相關(guān)性分析,找出相關(guān)特性曲線的全部極值點(diǎn),結(jié)果如圖4a所示。由圖可知,386~560和711~923 nm的相關(guān)系數(shù)為正,562~710和924~1 000 nm的相關(guān)系數(shù)為負(fù)。選擇相關(guān)系數(shù)絕對(duì)值高于0.4的極值點(diǎn)波長(zhǎng)變量,如圖4b所示,得到404、410 nm等25個(gè)特征向量作為特征波段變量。 圖4 相關(guān)系數(shù)分析法提取特征波段 2.2.2 MDBPSO算法提取特征波段 利用MDBPSO法進(jìn)行特征波提取,并與傳統(tǒng)DBPSO法進(jìn)行比較。具體設(shè)置參數(shù)如下:DBPSO慣性權(quán)重為1;最大迭代次數(shù)max為300;MDBPSO慣性權(quán)重max和min分別為1.2和0.9;max和-max分別為4和-4;最大迭代次數(shù)max為100。2種算法的粒子維數(shù)均為428,學(xué)習(xí)因子1和2均設(shè)為2。種群規(guī)模分別取20、30、40的條件下,算法均獨(dú)立運(yùn)行20次,表1統(tǒng)計(jì)了不同種群個(gè)數(shù)尋優(yōu)得到最優(yōu)適應(yīng)度(optimum fitness value,OFV)的最大值、最小值及平均值,及獲取到OFV最大值時(shí)算法的迭代次數(shù)。 由表1可知,2種算法的OFV值均隨種群規(guī)模的增加而增大,且在規(guī)模值為40時(shí)可獲得最優(yōu)結(jié)果。程序獨(dú)立測(cè)試運(yùn)行20次后,DBPSO法最優(yōu)適應(yīng)度最大值OFVmax、最小值OFVmin和平均值OFVave分別為0.761 6、0.680 4和0.731 8,其中OFVmax對(duì)應(yīng)的迭代次數(shù)為280次;算法改進(jìn)后,MDBPSO法的OFVmax介于0.786 7~0.812 3之間,OFVmin范圍為0.691 6~0.711 2,當(dāng)為40時(shí),OFVmax、OFVmin和OFVave分別為0.812 3、0.711 2和0.752 2,OFVmax對(duì)應(yīng)迭代次數(shù)為79次。綜上所述,相較傳統(tǒng)DBPSO算法,改進(jìn)后算法的OFV由0.761 6提高到0.812 3,迭代次數(shù)由280降低至79次,收斂效率提高了71.79%,表明MDBPSO的收斂精度和收斂效率均有明顯改善。 對(duì)2種算法測(cè)試后,適應(yīng)度收斂曲線如圖5所示。相同參數(shù)下,MDBPSO算法適應(yīng)度函數(shù)值在迭代前期迅速增大,由于迭代過(guò)程中通過(guò)對(duì)慣性權(quán)重的動(dòng)態(tài)調(diào)整,改進(jìn)后算法尋優(yōu)曲線適應(yīng)度函數(shù)值變化較快,在提高收斂精度的同時(shí),尋優(yōu)成功率亦顯著提高,30次迭代后,適應(yīng)度函數(shù)值已達(dá)到OFV的95%。在算法迭代后期,粒子群在局部尋找OFV值時(shí),慣性權(quán)重和粒子位置的動(dòng)態(tài)調(diào)整協(xié)同合作,壓縮粒子的搜索空間,提高了函數(shù)的局部搜索能力,從而降低了算法的時(shí)間復(fù)雜度,促使適應(yīng)度函數(shù)更為快速地收斂至最優(yōu)值。算法迭代至60次時(shí),適應(yīng)度函數(shù)值已經(jīng)基本接近OFV,可見算法改進(jìn)后收斂效率得到顯著提高。MDBPSO法求得函數(shù)最佳適應(yīng)度值為0.812 3,大于傳統(tǒng)DBPSO法的0.761 6,表明MDBPSO法優(yōu)選特征波段對(duì)青貯玉米含水率具有較強(qiáng)的表征力。綜上所述,MDBPSO算法能夠兼顧函數(shù)最優(yōu)適應(yīng)度的全局搜索和局部準(zhǔn)度和魯棒性。尋優(yōu),可有效避免“早熟”現(xiàn)象產(chǎn)生,具有較高的精準(zhǔn)度和魯棒性。 表1 傳統(tǒng)離散粒子群與改進(jìn)離散粒子群算法尋優(yōu)結(jié)果 注:OFVmax、OFVmin和OFVave分別為20次試驗(yàn)測(cè)試得到最優(yōu)適應(yīng)度的最大值、最小值和平均值。 Note: OFVmax, OFVminand OFVaveare the maximum, minimum and average of the optimal fitness respectively in the 20 test. 圖5 DBPSO和MDBPSO算法適應(yīng)度函數(shù)收斂曲線(種群個(gè)數(shù)為40) 通過(guò)DBPSO和MDBPSO法分別提取了188和62個(gè)光譜特征向量,2種算法優(yōu)選特征波段及其位置分布統(tǒng)計(jì)分別如圖6、7所示。分析發(fā)現(xiàn),2種方法提取的特征變量均在421~520 nm范圍分布最多,其次是571~670和871~920 nm。上述3個(gè)波段范圍內(nèi),DBPSO法優(yōu)選波段個(gè)數(shù)分別為51、45和15個(gè)。MDBPSO法優(yōu)選波段個(gè)數(shù)分別為15、11和12個(gè)。由此可以推斷,青貯玉米原料含水率在可見光區(qū)的敏感波段為421~520及571~670 nm,近紅外區(qū)的敏感波段為871~920 nm。另外,算法優(yōu)化后提取的特征波段在數(shù)量上有所減少,因此算法時(shí)間復(fù)雜度亦相應(yīng)降低。 圖6 DBPSO和MDBPSO算法優(yōu)選特征波段 研究采用Kennard-Stone算法對(duì)樣本進(jìn)行篩選以劃分校正集和預(yù)測(cè)集,167個(gè)樣本為校正集,71個(gè)樣本為預(yù)測(cè)集。分別以CC分析法、傳統(tǒng)DBPSO和MDBPSO法優(yōu)選特征波段與全波段光譜反射率(total spectral reflectance,TSR)為自變量,建立青貯玉米原料含水率CC-PLSR、DBPSO-PLSR及MDBPSO-PLSR、TSR-PLSR反演模型。繪制樣本含水率實(shí)測(cè)值與預(yù)測(cè)值分布關(guān)系的散點(diǎn)圖(圖8),并根據(jù)校正集、預(yù)測(cè)集實(shí)測(cè)值與預(yù)測(cè)值的決定系數(shù)R2、R2和均方根誤差RMSEC(root mean square error of calibration, RMSEC)、RMSEP(root mean square error of prediction, RMSEP)對(duì)各反演模型進(jìn)行精度評(píng)估。 圖7 DBPSO和MDBPSO算法優(yōu)選特征波段分布 比較4個(gè)模型性能發(fā)現(xiàn):TSR-PLSR和CC-PLSR 2個(gè)模型的擬合精度較低,校正集決定系數(shù)R2分別為0.69和0.70,預(yù)測(cè)集決定系數(shù)R2分別為0.67和0.64。DBPSO-PLSR模型較上述2個(gè)模型的性能指標(biāo)有明顯改善,其R2和R2分別為0.76和0.76,MDBPSO-PLSR模型擬合效果最好,散點(diǎn)分布相對(duì)貼近1﹕1線,校正集R2和均方根誤差RMSEC分別為0.81和0.032,預(yù)測(cè)集R2和均方根誤差RMSEP分別為0.80和0.045,相較其他3個(gè)模型在準(zhǔn)確度(2)和精確度(RMSE)方面均有顯著提高。 注:TSR為全波段光譜反射率,PLSR為偏最小二乘回歸法。 青貯玉米作為中國(guó)“糧改飼”政策的重要推手,將玉米跨區(qū)銷售轉(zhuǎn)向就地青貯,極大提高了農(nóng)業(yè)生產(chǎn)的利用效率[29]。原料含水率是青貯玉米品質(zhì)優(yōu)劣的關(guān)鍵影響因子,含水率過(guò)高容易導(dǎo)致可溶性營(yíng)養(yǎng)物質(zhì)隨滲出的汁液流失,產(chǎn)生梭酸發(fā)酵,含水率過(guò)低則不易壓實(shí),導(dǎo)致靑貯環(huán)境空氣含量超標(biāo)且易發(fā)生霉變[30]。因此,建立快速、無(wú)損、準(zhǔn)確的青貯玉米含水率測(cè)定方法,對(duì)推動(dòng)青貯產(chǎn)業(yè)健康快速發(fā)展有重要意義。 本文提取青貯玉米原料光譜信息,分析不同特征變量提取方法對(duì)原料含水率反演模型精度的影響,研究基于改進(jìn)型粒子群算法的PLSR模型在青貯玉米原料含水率反演預(yù)測(cè)方面的適用性。研究結(jié)果表明,利用全波段光譜和CC分析法建立PLSR模型的預(yù)測(cè)精度均比較低。分析認(rèn)為,全波段信息來(lái)源較為全面,但波段信息重疊帶來(lái)的繁雜冗余數(shù)據(jù)可能導(dǎo)致模型預(yù)測(cè)精度降低;CC分析法在很大程度上考慮了樣本實(shí)測(cè)值與光譜信息之間的相關(guān)度,提取25個(gè)特征變量的波長(zhǎng)范圍比較集中且相鄰間隔較小,波長(zhǎng)反射率之間表現(xiàn)為極顯著相關(guān)關(guān)系,相關(guān)系數(shù)最大可達(dá)0.996,由此可見,利用CC分析法提取的光譜特征變量之間存在多重共線性,降低了模型精度。此外,單一特征信息面窄且缺少互補(bǔ)信息,干擾信息對(duì)模型精度的影響也比較突出。粒子群算法將模型的反演精度作為特征變量的提取標(biāo)準(zhǔn),能夠更加“智能”地提取特征波段,相比原始光譜的428個(gè)波段,DBPSO提取的188個(gè)特征變量在很大程度上降低了模型的復(fù)雜度,但模型的OFV值較小且收斂效率頗低。本文提出的MDBPSO算法使用sigmoid映射函數(shù),對(duì)粒子的飛行速度加以限制,根據(jù)算法所處周期動(dòng)態(tài)調(diào)整粒子群慣性權(quán)重、改變粒子位置更新方式。研究表明,MDBPSO法提取特征變量個(gè)數(shù)由188減少至62個(gè),OFV值由0.761 6提高到至0.812 3,對(duì)應(yīng)迭代次數(shù)由280降低至79次。由此可見,算法優(yōu)化后進(jìn)一步降低了模型復(fù)雜度,在算法收斂效率和模型反演精度2個(gè)方面也均有顯著提高。此方法較為全面地考慮各波段光譜信息對(duì)反演參數(shù)的貢獻(xiàn)度,可在一定程度上克服或削弱利用傳統(tǒng)光譜參數(shù)進(jìn)行生化參數(shù)估測(cè)易受背景等干擾因素影響的弊端,提高參數(shù)反演模型的魯棒性和準(zhǔn)確性。 青貯玉米原料含水率是飼料發(fā)酵品質(zhì)的關(guān)鍵影響因子,為快速、無(wú)損檢測(cè)玉米原料含水率,本研究采集了383~1 004 nm范圍青貯玉米原料的高光譜數(shù)據(jù),采用SNV校正法對(duì)反射光譜進(jìn)行預(yù)處理。分別以CC分析法、傳統(tǒng)DBPSO和MDBPSO法優(yōu)選特征波段與全波段光譜反射率為自變量,建立青貯玉米原料含水率CC-PLSR、DBPSO-PLSR及MDBPSO-PLSR、TSR-PLSR反演模型。研究特征變量提取方法對(duì)模型預(yù)測(cè)精度的影響,探索應(yīng)用高光譜技術(shù)估算青貯玉米原料含水率的可行性,主要結(jié)論如下: 1)針對(duì)傳統(tǒng)DBPSO算法存在收斂性差及容易陷入“局部最優(yōu)”的缺陷,研究在粒子更新方式和慣性權(quán)重2個(gè)方面進(jìn)行改進(jìn),提出了一種改進(jìn)型離散粒子群算法。結(jié)果表明,MDBPSO法可有效提取青貯玉米原料含水率特征波段,算法優(yōu)化后,適應(yīng)度函數(shù)OFV值可由0.761 6提高到至0.812 3,對(duì)應(yīng)迭代次數(shù)由280降低至79次,收斂效率提高了71.79%。由此可見,MDBPSO法在收斂精度和收斂效率2個(gè)方面均得到顯著改善。 2)研究DBPSO和MDBPSO法優(yōu)選光譜特征波段的分布特征發(fā)現(xiàn),2種方法提取的特征變量均在421~520 nm范圍分布最多,其次是571~670和871~920 nm。上述3個(gè)波段范圍內(nèi),DBPSO法優(yōu)選波段個(gè)數(shù)分別為51、45和15個(gè),MDBPSO法優(yōu)選波段個(gè)數(shù)分別為15、11和12個(gè)。綜上所述,利用高光譜進(jìn)行青貯玉米原料含水率反演在可見光區(qū)的敏感波段為421~520及571~670 nm,在近紅外區(qū)的敏感波段為871~920 nm。 3)通過(guò)CC分析法、DBPSO和MDBPSO法優(yōu)選原料含水率高光譜特征變量,建立基于全波段和特征波段反射光譜的玉米原料含水率PLSR預(yù)測(cè)模型。對(duì)比各預(yù)測(cè)模型性能發(fā)現(xiàn),TSR-PLSR和CC-PLSR 2個(gè)模型的擬合精度較低,DBPSO-PLSR模型的性能指標(biāo)稍有改善,MDBPSO-PLSR的建模精度和預(yù)測(cè)精度均高于其他3個(gè)模型,其校正集決定系數(shù)R2和均方根誤差RMSEC分別為0.81和0.032,預(yù)測(cè)集決定系數(shù)R2和均方根誤差 RMSEP分別為0.80和0.045,該模型在準(zhǔn)確度和精確度方面均體現(xiàn)出顯著優(yōu)勢(shì)。結(jié)果表明,應(yīng)用高光譜圖像技術(shù)進(jìn)行青貯玉米原料含水率無(wú)損檢測(cè)具有較高的可行性,研究可為后續(xù)物料水分快速檢測(cè)儀器的開發(fā)提供理論依據(jù)。 [1] Ferraretto L F, Shaver R D, Luck B D, et al. Silage review: Recent advances and future technologies for whole-plant and fractionated corn silage harvesting[J]. Journal of Dairy Science, 2018, 101(5): 3937-3951. [2] Bruning D, Gerlach K, Wei? K, et al. Effect of compaction, delayed sealing and aerobic exposure on maize silage quality and on formation of volatile organic compounds[J]. Grass Forage Sci, 2018, 73: 53-66. [3] Pu Yuanyuan, Sun Dawen. Vis-NIR hyperspectral imaging in visualizing moisture distribution of mango slices during microwave-vacuum drying[J]. Food Chemistry, 2015, 188: 271-278. [4] 劉志剛,徐勤超. 基于高光譜技術(shù)的基質(zhì)含水率快速測(cè)定方法[J]. 灌溉排水學(xué)報(bào),2017,36(10):82-86. Liu Zhigang, Xu Qinchao. Rapid determination of matrix moisture content based on hyperspectral technology[J]. Journal of Irrigation and Drainage, 2017, 36(10): 82-86. (in Chinese with English abstract) [5] Zhu Yaodi, Zou Xiaobo, Shen Tingting, et al. Determination of total acid content and moisture content during solid-state fermentation processes using hyperspectral imaging[J]. Journal of Food Engineering, 2016, 174: 75-84. [6] Deng Shuiguang, Xu Yifei, Li Xiaoli, et al. Moisture content prediction in teal leaf with near infrared hyperspectral imaging[J]. Computers and Electronics in Agriculture, 2015, 118: 38-46. [7] Kobori H, Gorretta N, Rabatel G, et al. Applicability of Vis- NIR hyperspectral imaging for monitoring wood moisture content (MC)[J]. Holzforschung, 2013, 67 (3): 307-314. [8] Caporaso N, Whitworth M B, Grebby S, et al. Rapid prediction of single green coffee bean moisture and lipid content by hyperspectral imaging[J]. Journal of Food Engineering, 2018, 227: 18-29. [9] 吳見,譚靖,鄧凱,等. 基于優(yōu)化指數(shù)的玉米冠層含水量遙感估測(cè)[J]. 湖南農(nóng)業(yè)大學(xué)學(xué)報(bào):自然科學(xué)版,2015,41(6):685-690. Wu Jian, Tan Jing, Deng Kai, et al. Remote sensing monitoring of the corn canopy water content based on the optimized index[J]. Journal of Hunan Agricultural University: Natural Sciences, 2015, 41(6): 685-690. (in Chinese with English abstract) [10] 孫紅,陳香,孫梓淳,等. 基于透射光譜的玉米葉片含水率快速檢測(cè)儀研究[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2018,49(3):173-178. Sun Hong, Chen Xiang, Sun Zichun, et al. Rapid detection of moisture content in maize leaves based on transmission spectrum[J]. Transactions of the Chinese Society for Agricultural Machinery, 2018, 49(3): 173-178. (in Chinese with English abstract) [11] Yang Chen, Tan Yulei, Bruzzone L, et al. Discriminative feature metric learning in the affinity propagation model for band selection in hyperspectral images[J]. Remote Sensing, 2017, 9(8): 782-798. [12] 孫俊,叢孫麗,毛罕平,等. 基于高光譜的油麥菜葉片水分CARS-ABC-SVR預(yù)測(cè)模型[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017, 33(5):178-184. Sun Jun, Cong Sunli, Mao Hanping, et al. CARS-ABC-SVR model for predicting leaf moisture of leaf-used lettuce based on hyperspectral[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(5): 178-184. (in Chinese with English abstract) [13] Digman M F, Shinners K J. Real-time moisture measurement on a forage harvester using near-infrared reflectance spectroscopy[J]. Transactions of the ASABE, 2008, 51 (5): 1801-1810. [14] Cozzolino D, Fassio A, Ferna?Ndez E, et al. Measurement of chemical composition in wet whole maize silage by visible and near infrared reflectance spectroscopy[J]. Animal Feed Science & Technology, 2006, 129 (3): 329-336. [15] Zhou Xin, Sun Jun, Mao Hanping, et al. Visualization research of moisture content in leaf lettuce leaves based on WT-PLSR and hyperspectral imaging technology[J]. Journal of Food Process Engineering, 2017, 41 (2): 1-7. [16] Waseem Amjad, Crichton S O J, Munir A, et al. Hyperspectral imaging for the determination of potato slice moisture content and chromaticity during the convective hot air drying process[J]. Biosystems Engineering, 2018, 166: 180-183. [17] 中華人民共和國(guó)國(guó)家質(zhì)量監(jiān)督檢驗(yàn)檢疫總局. 飼料中水分的測(cè)定:GB/T 6435—2014[S]. 北京:標(biāo)準(zhǔn)出版社,2014: 7. [18] Brook A, Polinova M, Bendor E. Fine tuning of the SVC method for airborne hyperspectral sensors: The BRDF correction of the calibration nets targets[J]. Remote Sensing of Environment, 2018, 204: 861-871. [19] 趙茂程,楊君榮,陸丹丹,等. 基于高光譜成像的青梅酸度檢測(cè)方法[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2017,48(9):318-323. Zhao Maocheng, Yang Junrong, Lu Dandan, et al. Detection methods of greengage acidity based on hyperspectral imaging [J]. Transactions of the Chinese Society for Agricultural Machinery, 2017, 48(9): 318-323. (in Chinese with English abstract) [20] 孫紅,鄭濤,劉寧,等. 高光譜圖像檢測(cè)馬鈴薯植株葉綠素含量垂直分布[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(1):149-156. Sun Hong, Zheng Tao, Liu Ning, et al. Vertical distribution of chlorophyll in potato plants based on hyperspectral imaging[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(1): 149-156. (in Chinese with English abstract) [21] Sun Jun, Lu Xinzi, Mao Hanping, et al. Quantitative determination of rice moisture based on hyperspectral imaging technology and BCC-LS-SVR algorithm[J]. Journal of Food Process Engineering, 2017, 40 (3): 1-8. [22] Kennedy J, Eberhart R. Particle swarm optimization[C]// IEEE International Conference on Neural Networks, Perth, Australia, Proceedings, IEEE, 1995: 1942-1948. [23] Kennedy J, Eberhart R.A discrete binary version of the particle swarm algorithm[C]// IEEE International Conference on Systems, Man, and Cybernetics, Computational Cybernetics and Simulation, IEEE, 1997, 5: 4104-4108. [24] Yang Jun, Zhang Hesheng, Ling Yun, et al. Task allocation for wireless sensor network using modified binary particle swarm optimization[J]. IEEE Sensor Journal, 2014, 14(3): 882-891. [25] 孟常亮,李衛(wèi)忠,廖勇,等. 基于改進(jìn)離散二進(jìn)制粒子群的SVM選擇集成算法[J]. 計(jì)算機(jī)工程與應(yīng)用,2011, 47(28):166-169. Meng Changliang, Li Weizhong, Liao Yong, et al. SVM selection ensemble algorithm based on improved binary particle swarm optimization[J]. Computer Engineering and Applications, 2011, 47(28): 166-169. (in Chinese with English abstract) [26] 胡清,張強(qiáng). 基于改進(jìn)二進(jìn)制粒子群算法的配電網(wǎng)故障定位[J]. 南京工程學(xué)院學(xué)報(bào):自然科學(xué)版,2016,14(3):77-81. Hu Qing, Zhang Qiang. Fault location of distribution networks based on improved binary particle swarm optimization algorithm[J]. Journal of Nanjing Institute of Technology: Natural Science Edition, 2016, 14(3): 77-81. (in Chinese with English abstract) [27] 劉朔,周康,張杰,等. 基于二進(jìn)制粒子群的圖像分割算法[J]. 武漢工業(yè)學(xué)院學(xué)報(bào),2011,30(4):42-44. Liu Suo, Zhou Kang, Zhang Jie, et al. Image segmentation algorithm based on binary PSO[J]. Journal of Wuhan Polytechnic University, 2011, 30(4): 42-44. (in Chinese with English abstract) [28] 曹引,冶運(yùn)濤,趙紅莉,等. 基于離散粒子群和偏最小二乘的水源地濁度高光譜反演[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2018,49(1):173-182. Cao Yin, Ye Yuntao, Zhao Hongli, et al. Satellite hyperspectral retrieval of turbidity for water source based on discrete particle swarm and partial least squares[J]. Transactions of the Chinese Society for Agricultural Machinery, 2018, 49(1): 173-182. (in Chinese with English abstract) [29] 郭勇慶,曹志軍,李勝利,等. 全株玉米青貯生產(chǎn)與品質(zhì)評(píng)定關(guān)鍵技術(shù)[J]. 中國(guó)畜牧雜志,2012,48(18):39-43. Guo Yongqing, Cao Zhijun, Li Shengli, et al. Key technologies for silage production and quality evaluation of whole plant corn[J]. Journal of Animal Science Chinese, 2012, 48(18): 39-43. (in Chinese with English abstract) [30] Huart F, Malumba P, Odjo S, et al. In vitro and in vivo assessment of the effect of initial moisture content and drying temperature on the feeding value of maize grain[J/OL]. British Poultry Science, 2018, 1-11. http:// www.tandfonline. com/loi/cbps20. Moisture content dectection in silage maize raw material based on hyperspectrum and improved discrete particle swarm Zhang Jue1,2, Tian Haiqing1※, Zhao Zhiyu1, Zhang Lina2, Zhang Jing1, Li Fei3 (1.010018,;2.010020,; 3.010019,) Moisture content of silage maize raw material affects juice discharge, compaction degree and microbial activity during the whole silage process, and it has further influence on silage fermentation quality. Rapid, non-destructive and accurate detection of moisture content in silage maize raw material issignificant for ensuring the silage maizequality and promoting the silage industry healthy and rapidly. Hyperspectral imagesof silage maize raw material in the visible and near infrared (383-1 004 nm) regions were acquired by the hyperspectral imaging system, and then corresponding moisture content in silage maize raw material were obtained by oven heating method successfully. Thehyperspectral information was extracted from the images by selecting the region of interest (ROI) using the ENVI software. The standard normalized variate (SNV) was applied for eliminating or weakening the effect of particle scattering on original hyperspectral data. The hyperspectral imaging provides much more information including spectral and image information for all the samples of silage maize raw material, however, hyperspectral imagery contains more noise and redundancy. These disturbances made it difficult to meet the needs of fast and effective detection of certain objects. Therefore, it was difficult to apply online industrial applications in daily life directly, and the feature band effective selection for hyperspectral images was very critical. In view of the disadvantages as poor efficiency and easy premature, the traditional discrete particle swarm optimization (DBPSO) was optimized in terms of particle updating method and inertia weight. A modified discrete particle swarm optimization (MDBPSO) was proposed to extract the hyperspectral feature bands effectively. The hyperspectral characteristic variables of raw material moisture content were extracted using the correlation coefficient (CC), DBPSO and MDBPSO method. Partial least squares regression (PLSR) prediction model for silage maize moisture content was established by using full band and characteristic band. The results indicated that the convergence accuracy and efficiency of MDBPSO had a significantly improvement compared with the DBPSO method. When the population number was 40 and the program independent test ran 20 times, for DBPSO, the maximum value of optimal fitness (OFVmax), the minimum value of optimal fitness (OFVmin), and the mean value of optimal fitness (OFVave) were 0.761 6, 0.680 4 and 0.731 8 respectively, and the number of iterations corresponding to the OFVmaxwas 280 times. The OFVmax, OFVmin, and OFVavewere 0.812 3, 0.711 2 and 0.752 2 for MDBPSO, respectively, and the number of iterations corresponding to the OFVmaxwas 79 times. After the improvement of DBPSO method, OFV of the fitness function was increased from 0.761 6 to 0.812 3, the number of iterations was reduced from 280 to 79, and the convergence efficiency was increased by 71.79%. 188 and 62 eigenvectors were extracted by DBPSO and MDBPSO respectively. The characteristic bands selected by the DBPSO method were mainly distributed in 421-520 nm, followed by 571-670 nm and 871-920 nm, and the number of bands was 51, 45 and 15 respectively. The characteristic bands selected by the MDBPSO method were also mainly distributed in the above band, and the number of the wave segments was 15, 11 and 12 respectively. It could be inferred that the sensitive bands of moisture content of silage maize in visible light region are 421-520, 571-670 nm and 871-920 nm in near infrared region. Comparing the performance of the 4 models, the fitting accuracies of TSR-PLSR and CC-PLSR were lower, and the verification set determination coefficients (R2) were 0.69 and 0.70 respectively, and the prediction set determination coefficients (R2) were 0.67 and 0.64, respectively. The DBPSO-PLSR model was improved significantly, and theR2andR2was 0.76 and 0.76 respectively. The DBPSO-PLSR model performed better than the other 3 model: TSR-PLSR, CC-PLSR and DBPSO-PLSR, achieving the highest accuracy withR2of 0.81, RMSEC of 0.032,R2of 0.80, RMSEP of 0.045. The study demonstrated that the application of hyperspectral image technology to the nondestructive testing of the moisture content of silage maize raw material content had high feasibility, and could provide efficient guidance for rapid detecting instrument development. particles; moisture; spectral analysis; hyperspectrum; particle swarm optimization; silage maize; feature band 2018-08-13 2018-11-22 國(guó)家自然科學(xué)基金項(xiàng)目(41261084);內(nèi)蒙古自然科學(xué)基金項(xiàng)目(2016MS0346) 張玨,博士生,主要從事基于光、電特性的農(nóng)作物營(yíng)養(yǎng)及農(nóng)產(chǎn)品品質(zhì)診斷研究。Email:zhangjue0428@163.com 田海清,教授,博士生導(dǎo)師,主要從事基于光、電特性的農(nóng)作物營(yíng)養(yǎng)、農(nóng)產(chǎn)品品質(zhì)診斷研究及農(nóng)牧業(yè)機(jī)械智能化研究。Email:hqtian@126.com 10.11975/j.issn.1002-6819.2019.01.035 S816.5; TP391 A 1002-6819(2019)-01-0285-09 張 玨,田海清,趙志宇,張麗娜,張 晶,李 斐.基于改進(jìn)離散粒子群算法的青貯玉米原料含水率高光譜檢測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(1):285-293. doi:10.11975/j.issn.1002-6819.2019.01.035 http://www.tcsae.org Zhang Jue, Tian Haiqing, Zhao Zhiyu, Zhang Lina, Zhang Jing, Li Fei.Moisture content dectection in silage maize raw material based on hyperspectrum and improved discrete particle swarm [J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(1): 285-293. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2019.01.035 http://www.tcsae.org2 結(jié)果與分析
2.1 青貯玉米原料反射光譜與預(yù)處理
2.2 特征波段提取
2.3 青貯玉米原料含水率預(yù)測(cè)模型的建立
3 討 論
4 結(jié) 論