基于機器學習的棉花葉面積指數(shù)監(jiān)測

馬怡茹，呂 新，易 翔，馬露露，祁亞琴，侯彤瑜，張 澤

基于機器學習的棉花葉面積指數(shù)監(jiān)測

馬怡茹，呂 新，易 翔，馬露露，祁亞琴，侯彤瑜，張 澤※

（石河子大學農(nóng)學院/新疆生產(chǎn)建設(shè)兵團綠洲生態(tài)農(nóng)業(yè)重點實驗室，石河子 832003）

為實現(xiàn)基于機器學習和無人機高光譜影像進行棉花全生育期葉面積指數(shù)（Leaf Area Index, LAI）監(jiān)測，該研究基于大田種植滴灌棉花，在不同品種及不同施氮處理的小區(qū)試驗基礎(chǔ)上，對無人機獲取的高光譜數(shù)據(jù)分別采用一階導（First Derivative, FDR）、二階導（Second Derivative, SDR）、SG（Savitzky-Golay）平滑和多元散射校正（Multiplicative Scatter Correction, MSC）進行預處理，并結(jié)合Pearson相關(guān)系數(shù)法、連續(xù)投影（Successive Projections Algorithm, SPA）、隨機蛙跳（Shuffled Frog Leaping Algorithm, SFLA）和競爭性自適應(yīng)重加權(quán)（Competitive Adaptive Reweighting, CARS）篩選敏感波段，將篩選出的波段，使用偏最小二乘回歸（Partial Least Squares Regression, PLSR）、支持向量回歸（Support Vector Regression, SVR）和隨機森林回歸（Random Forest Regression, RFR）3種機器學習算法構(gòu)建棉花LAI監(jiān)測模型。結(jié)果表明：棉花冠層LAI敏感響應(yīng)波段集中在可見光（400～780 nm）和近紅外（900 nm之后）波段；對比3種機器學習算法，各預處理下RFR建立的LAI監(jiān)測模型精度最高，穩(wěn)定性最好，其中以FDR-SFLA-RFR模型最佳，在建模集的決定系數(shù)為0.74，均方根誤差為1.648 3，相對均方根誤差為26.39%；驗證集的決定系數(shù)、均方根誤差分別為0.67和1.622 0，相對均方根誤差為25.97%。該研究基于無人機獲取的棉花冠層光譜反射率，從不同光譜預處理、波段篩選及建模方法建立的模型中篩選出最佳估算模型用于棉花全生育期LAI監(jiān)測，研究結(jié)果可為棉花大田精準管理及變量施肥提供依據(jù)。

棉花；無人機；高光譜；機器學習；葉面積指數(shù)

0 引 言

棉花是重要的經(jīng)濟作物[1]，不同施氮水平對棉花長勢有顯著影響[2-4]。葉面積指數(shù)（Leaf Area Index ,LAI）是反應(yīng)作物冠層結(jié)構(gòu)及長勢的重要指標之一[5-6]，通過監(jiān)測LAI變化可為棉花變量施肥提供依據(jù)[7-8]，因此快速、準確、無損的監(jiān)測棉花LAI對于指導作物施肥具有重要意義。傳統(tǒng)的LAI監(jiān)測主要靠人工取樣，需要投入大量人力和時間成本，存在滯后性，無法滿足實時監(jiān)測的需要。

遙感技術(shù)能夠?qū)崿F(xiàn)及時、動態(tài)、宏觀的監(jiān)測，成為監(jiān)測作物生長信息的重要手段。近年來，國內(nèi)外大量研究通過遙感技術(shù)對作物生物量[9-11]、葉綠素含量[12-14]、水氮含量[15-18]等生理生化參數(shù)進行反演。而針對作物LAI的監(jiān)測，也基于手持光譜儀、無人機和衛(wèi)星等遙感手段開展了大量研究[19-22]。地面光譜監(jiān)測具有無損、精確等優(yōu)點，但由于拍攝范圍以及儀器重量等因素限制，近地光譜不能實現(xiàn)空間尺度連續(xù)快速監(jiān)測[23]。此外，有研究表明衛(wèi)星影像在作物LAI監(jiān)測方面具有一定潛力[24]，但由于其影像分辨率在10～60m，多用于森林或大區(qū)域尺度的作物LAI監(jiān)測[25-26]。無人機在作物監(jiān)測方面具有快速、重復的捕獲能力，且與衛(wèi)星影像相比影像分辨率更高[27]，更適應(yīng)小地塊精確監(jiān)測。已有學者基于無人機獲取的光譜圖像對小麥、水稻、玉米等作物的LAI進行監(jiān)測[28-31]。無人機可快速獲取大量的高光譜數(shù)據(jù)，其中包含豐富的信息，同時也存在數(shù)據(jù)冗余的問題，機器學習算法因其強大的學習能力和對數(shù)據(jù)深層信息的挖掘和理解能力，越來越多與遙感技術(shù)相結(jié)合應(yīng)用于作物生長監(jiān)測[32-33]。國內(nèi)外學者多從光譜信息中提取植被指數(shù)，利用機器學習算法提高監(jiān)測模型精度[34-36]。

目前通過光譜數(shù)據(jù)進行LAI監(jiān)測多基于植被指數(shù)建模，而植物冠層的高光譜反射率是對植被特征最直接的反應(yīng)，與植被指數(shù)相比可以提供更詳細，更豐富的信息，合理的光譜變換也能夠在一定程度上消除光譜數(shù)據(jù)的背景和噪聲。但高光譜數(shù)據(jù)也具有多重共線性，偏最小二乘模型是多元線性模型的一種延伸，能夠減少數(shù)據(jù)變量間的共線性問題，支持向量機和隨機森林具有較高的學習和預測能力，能夠從不同角度克服變量間共線性的問題。因此，為提高棉花LAI監(jiān)測模型精度，本研究使用不同方法對光譜影像數(shù)據(jù)進行預處理，再分別篩選敏感波段，采用3種不同機器學習算法構(gòu)建LAI監(jiān)測模型，尋找最佳模型，以期為新疆棉花大田精準管理及變量施肥提供依據(jù)。

1 材料與方法

1.1 試驗區(qū)及試驗設(shè)計

本試驗研究區(qū)域位于石河子大學農(nóng)試場二連（44°19′N，85°59′E）。研究區(qū)為干旱半干旱區(qū)域，年平均降水量125.9～207.7 mm，晝夜溫差大，前茬作物為棉花。試驗區(qū)域如圖1所示。

為使模型適應(yīng)于多種環(huán)境，試驗設(shè)置不同棉花品種和施氮處理。供試棉花品種為新陸早53號、新陸早45號和魯研棉24號；每個品種設(shè)置6個氮處理分別為N0(0 kg/hm2)、N1(120 kg/hm2)、N2(240 kg/hm2)、NC(360 kg/hm2)、N3(480 kg/hm2)、N4(600 kg/hm2)，每個處理重復3次，共54個小區(qū)，各小區(qū)面積為21 m2（2.1 m×10 m）。于2019年4月24日播種，2019年10月15日收獲。新陸早系列按照“一膜三管六行”的機采棉種植模式；魯研棉24號按“一膜三管三行”的模式種植。全生育期按新疆“矮、密、早、膜”的高產(chǎn)栽培技術(shù)進行大田管理，并注意預防病蟲草害。

1.2 無人機高光譜圖像獲取和處理

利用無人機搭載Nano-Hyperspecal（美國）傳感器獲取出苗后第57、66、76、88、98、112及120天的高光譜圖像。無人機使用大疆M600Pro（中國，深圳）六旋翼無人機，最大載負荷10 kg，配備6塊電池，數(shù)據(jù)采集時飛行高度為100 m。Nano-Hyperspecal為推掃式成像光譜儀，基本參數(shù)如表1所示。無人機獲取冠層光譜影像時每次航線一致，獲取的影像為.hdr格式，將影像數(shù)據(jù)導入Nano自帶的校正軟件SpectralView進行校正，校正后的影像導入到ENVI5.1中進行圖像拼接和并通過標準板計算反射率。

1.3 數(shù)據(jù)采集與預處理

1.3.1 棉花LAI采集

獲取無人機高光譜圖像后，在各小區(qū)內(nèi)隨機選擇連續(xù)3株具有代表性的樣株，取全株葉片利用LI-3000測量單株總?cè)~面積，依據(jù)公式（1）計算葉面積指數(shù)LAI：

表1 Nano-Hyperspecal傳感器主要參數(shù)

1.3.2高光譜數(shù)據(jù)預處理

無人機高光譜影像獲取過程中由于環(huán)境因素的影響影像會產(chǎn)生噪音，這種噪音干擾在數(shù)據(jù)獲取過程中是不可避免的，雖然在圖像拼接過程中進行大氣校正，但仍有部分干擾依然存在。為了有效提取對棉花LAI敏感的波段，常通過對原始光譜進行預處理以突出特征波段、去除背景噪音。本研究采用4種不同的方式：一階導（First Derivative, FDR）、二階導(Second Derivative, SDR)、SG（Savitzky-Golay）平滑及多元散射校正（Multiplicative Scatter Correction, MSC）進行光譜預處理。

1.3.3 特征波段篩選

高光譜影像中包括272個波段信息，使用全波段建模會出現(xiàn)數(shù)據(jù)冗余和共線性的問題，因此需要從中篩選出敏感波段以降低數(shù)據(jù)維度，減少冗余信息。本研究采用Pearson相關(guān)系數(shù)、連續(xù)投影算法（Successive Projections Algorithm, SPA）、隨機蛙跳（Shuffled Frog Leaping Algorithm, SFLA）和競爭性自適應(yīng)重加權(quán)（Competitive Adaptive Reweighting, CARS）4種方法篩選與棉花LAI相關(guān)性強的特征波段，其中相關(guān)系數(shù)法和SFLA選擇了相關(guān)性最高、選擇概率最高的10個波段進行建模。SPA是將各自波長投影到其他波長上計算其投影向量，并選擇投影向量長的為特征波段，其結(jié)果為信息最多、共線性現(xiàn)象最少的波段組合[37]。SFLA算法是一種基于青蛙社會行為的群體智能算法，結(jié)合了確定性方法和隨機性方法，是求解組合優(yōu)化問題的有效工具。CARS算法通過自適應(yīng)重加權(quán)采樣選擇出PLS模型中回歸系數(shù)絕對值大的波長，利用交互驗證選出RMSECV最低的子集，選擇出最優(yōu)變量組合[38]，可根據(jù)信息量確定特征波段個數(shù)。

1.4 模型構(gòu)建與驗證

1.4.1 模型構(gòu)建

為克服高光譜數(shù)據(jù)共線性問題，本文采用偏最小二乘回歸（Partial Least Squares Regression, PLSR）、支持向量機回歸（Support Vector Regression, SVR）和隨機森林回歸（Random Forest Regression, RFR）3種機器學習方法構(gòu)建回歸模型。機器學習被廣泛應(yīng)用于植物生理生化參數(shù)與遙感信息非線性關(guān)系建立，與簡單線性回歸相比，機器學習更適合基于多變量、多樣本的結(jié)果預測，基于Matlab2016a實現(xiàn)。

1.4.2 精度驗證

單次采樣可獲取54個數(shù)據(jù)集，每個數(shù)據(jù)集包括54個地面實測數(shù)據(jù)和一架次無人機數(shù)據(jù)。全生育期共獲取345個樣本，按訓練集：驗證集=2：1進行數(shù)據(jù)集劃分，訓練集230個樣本，驗證集115個樣本。以決定系數(shù)（2）、均方根誤差（Root Mean Square Error, RMSE）和相對均方根誤差（Relative Root Mean Square Error, rRMSE）進行LAI估算模型的精度評估。其中，2越大，模型擬合性越好，RMSE和rRMSE越小，模型精度越高。其計算公式如下：

2 結(jié)果與分析

2.1 不同LAI值的棉花冠層反射率

圖2a為高光譜影像中不同LAI值對應(yīng)的冠層反射率，在760～1 000 nm內(nèi)LAI越高冠層反射率越高，且差異明顯。由圖2b可知：在490～760 nm LAI值與冠層反射率呈現(xiàn)負相關(guān)；760～1 000 nm呈現(xiàn)正相關(guān)。由此表明，無人機獲取的棉花冠層高光譜影像能夠有效反應(yīng)棉花LAI值變化。

2.2 特征波段篩選

以不同方法進行波段篩選，結(jié)果如表2所示，棉花LAI敏感波段在可見光及近紅外區(qū)域均有分布。其中，原始光譜及SG平滑處理后以Pearson篩選出的特征波段在紅光（700～720 nm）波段較為集中，多為相鄰波段；而經(jīng)過FDR、SDR及MSC預處理后以Pearson進行波段篩選后其特征波段在可見光（400～780 nm）波段均有分布。SFLA篩選出各預處理下的敏感波段均勻分布在可見光及近紅外（400～1 000 nm）波段。SPA篩選出的敏感波段多集中在近紅外（780～1 000 nm）波段，篩選結(jié)果較為集中。CARS在各預處理下篩選出的波段范圍較廣在可見光及近紅外（400～1 000 nm）波段皆有選擇，多集中在近紅外波段，篩選出的波段數(shù)最多。由此可見，不同篩選方法針對不同預處理后的光譜特征可在一定程度上實現(xiàn)數(shù)據(jù)降維。

表2 特征波段篩選結(jié)果

注：FDR為一階導，SDR為二階導，MSC為多元散射校正，Pearson為皮爾遜相關(guān)系數(shù)法，SPA為連投影法，SFLA為隨機蛙跳法，CARS為競爭性自適應(yīng)重加權(quán)算法，下同。

Note: FDR is the first derivative, SDR is the second derivative, MSC is the multiplicative scatter correction, Pearson is the Pearson correlation coefficient method, SPA is the successive projections algorithm, SFLA is the shuffled frog leaping algorithm, and CARS is the competitive adaptive reweigh ting algorithm, the same below.

2.3 基于PLSR的棉花LAI監(jiān)測模型

PLSR是多元線性回歸、主成分分析以及典型相關(guān)分析的結(jié)合，它要求各變量與估算目標間具有較好的線性關(guān)系。結(jié)合表2中篩選出的敏感波段，使用PLSR建模并驗證。如圖3，模型建模結(jié)果中2由0.17提升到0.59，RMSE從2.717 2降低到1.911 9，rRMSE從43.50%降低到30.61%。在PLSR模型中的最佳模型為FDR-SFLA組合獲取的光譜信息建立的模型，模型效果最差為SG-Pearson模型。模型驗證結(jié)果如圖4，與建模結(jié)果一致，以FDR-SFLA模型的擬合線更趨向于1∶1線，其2=0.59，RMSE=1.731 9，rRMSE=27.73%。

2.4基于SVR的棉花LAI監(jiān)測模型

SVR是支持向量機的一種重要形式，能夠有效、準確解決回歸問題。如圖5，不同處理下的SVR模型，其2由0.36提升到0.72，RMSE由2.579 2降低到1.570 8，rRMSE由41.29%降低到25.15%，SG-Pearson和SG-SFLA模型效果最好，但其驗證精度與模型結(jié)果存在一定差異。由圖6可知，SG-SFLA模型較SG-Pearson模型驗證結(jié)果更好，是由于相比SFLA，Pearson篩選的敏感波段較為集中，而SG平滑降噪同時使部分特征信息被消除，出現(xiàn)共線性問題，導致建模效果較好，而驗證結(jié)果較差。綜合對比模型結(jié)果及驗證結(jié)果，F(xiàn)DR-SFLA建模集的2=0.63，RMSE=1.890 8，rRMSE=30.27%，驗證集的2=0.63，RMSE=1.7137，rRMSE=27.44%。雖然FD-SFLA的建模效果不是最佳，但建模集與驗證集結(jié)果表現(xiàn)一致，且該模型真實值與預測值的線性擬合關(guān)系更趨向于1∶1線。因此，基于SVR算法構(gòu)建的模型中，F(xiàn)DR-SFLA模型效果最好。

2.5 基于RFR的棉花LAI監(jiān)測模型

隨機森林是一種廣泛應(yīng)用于分類、回歸等領(lǐng)域的機器學習算法，可以提供特征的重要性評估，從而能洞察特征選擇的過程。圖7所示，LAI監(jiān)測模型2由0.47提升到0.74，RMSE由2.152 6降低到1.648 3，rRMSE由34.46%降低到26.39%，其驗證與建模結(jié)果表現(xiàn)一致。綜合對比，RFR構(gòu)建的模型中FDR-SFLA模型效果最好，其2=0.74，RMSE=1.648 3，rRMSE=26.39%，從圖8可看出其真實值和預測值的線性擬合更趨向于1∶1線，其2=0.67，RMSE=1.622 0，rRMSE=25.97%。

綜上所述，對比不同建模方法的模型精度，RFR模型的模型和驗證結(jié)果均優(yōu)于PLSR和SVR；SVR效果優(yōu)于PLSR，但信息冗余導致建模集和驗證集結(jié)果出現(xiàn)偏差。RFR模型克服了這一問題，去除了冗余信息干擾，有效提升了模型精度。對比可知，非線性模型性能優(yōu)于線性模型性能，而3種機器學習方法都以FDR-SFLA模型效果最好，相較于其他方法RFR的2提高了20.27%，RMSE降低了3.97%，rRMSE降低了3.90%；驗證集的2提高了11.94%，RMSE降低了3.97%，rRMSE降低了3.90%。

3 討 論

本研究對LAI變化和其光譜響應(yīng)進行分析，結(jié)果表明可見光區(qū)域LAI與冠層光譜反射率呈負相關(guān)，近紅外區(qū)域LAI與冠層光譜反射率呈正相關(guān)，這與前人在冬小麥[39]、油菜[40]和玉米[41]研究中的結(jié)果一致。這是由于植被光譜反射率，在350～800 nm差異主要是由于植物體內(nèi)葉綠素和其他色素的影響，800～1 000 nm的差異來源于植物細胞組織的散射，棉花生長茂盛多片葉子疊加輻射作用下，則會在近紅外波段產(chǎn)生較高的反射率，因此不同的LAI值的冠層光譜在近紅外區(qū)域差異更明顯。

冠層原始光譜受太陽輻射通量，作物結(jié)構(gòu)特征和土壤背景條件影響[42]，光譜預處理可減少背景噪聲信息，能夠有效提高光譜信息精度[43]。前人研究表明，F(xiàn)DR處理可減輕作物冠層重疊對反射率的影響，也可最小化土壤或大氣背景噪聲[44]。王玉娜等[45]以不同方法處理原始光譜后估算冬小麥生物量，以FDR處理后的光譜反射率與生物量相關(guān)性更高。Li等[46]發(fā)現(xiàn)750 nm波長處的一階導數(shù)與LAI具有較高的相關(guān)性，估算模型精度較高，與本研究結(jié)果表現(xiàn)一致。

高光譜分析包括特征波段篩選和回歸建模2個步驟。波段篩選能夠有效實現(xiàn)數(shù)據(jù)降維，緩解共線性的問題，本研究通過不同篩選方法篩選出的LAI敏感波段多集中在400～780 nm的可見光波段以及900 nm以后的近紅外波段，這與孫晶京等[47]通過隨機蛙跳法篩選出的敏感波段相似。本研究以SFLA法篩選出的敏感波段建模效果最好，Ren等[48]對比4種波段篩選方法進行紅茶評級，得到相同結(jié)果。已有研究中的模型，波段篩選有效降低了數(shù)據(jù)維數(shù)，但傳統(tǒng)的線性回歸建模仍然會出現(xiàn)共線性問題。本研究基于不同機器學習算法建立LAI監(jiān)測模型，其結(jié)果表現(xiàn)為：RFR最佳，SVR次之，PLSR最差。其中以FDR-SFLA-RFR模型精度最佳（模型的2=0.74，RMSE=1.648 3，rRMSE=26.39%；驗證集2=0.67，RMSE=1.622 0，rRMSE=25.97%），說明RFR對于棉花LAI監(jiān)測更有效。RFR是基于樹的集成學習技術(shù)，抗過擬合能力較強，被廣泛應(yīng)用于長勢指標監(jiān)測，并具有更優(yōu)的建模效果，如：Han等[49]通過機器學習算法估算玉米地上生物量；Lu等[50]基于RGB圖像建立小麥生物量估算模型；Wang等[51]以不同建模方法監(jiān)測氮營養(yǎng)指數(shù)，均以RFR模型的效果最優(yōu)。RFR模型在大樣本預測上要比其他算法具有優(yōu)勢，本研究中RFR模型也表現(xiàn)出較好的預測能力。

田明璐等[52]基于無人機獲取的光譜數(shù)據(jù)建立植被指數(shù)用于棉花盛蕾期LAI監(jiān)測，其模型驗證集的2=0.85，RMSE=0.02，其結(jié)果優(yōu)于本研究，但其模型僅適用于盛蕾期LAI監(jiān)測。而本研究建立的模型，可用于棉花全生育期LAI監(jiān)測，且涉及不同棉花品種。Chen等[53]基于無人機獲取的多光譜數(shù)據(jù)建立棉花不同生育期LAI監(jiān)測模型，其模型的2=0.65，RMSE=0.62，精度低于本研究基于棉花冠層光譜反射率建立的模型。近年來，為更好實現(xiàn)棉花生長信息監(jiān)測，有學者引入了機器視覺、深度學習等技術(shù)，有效提高了監(jiān)測模型精度[9,54]。為提高模型精度，未來可考慮引入更多監(jiān)測技術(shù)以及建模手段。

綜上所述，光譜數(shù)據(jù)采用FDR預處理，采用SFLA篩選敏感波段，可優(yōu)化模型變量，提高模型精度。RFR能夠有效對抗噪聲，更適合針對遙感數(shù)據(jù)進行建模，F(xiàn)DR-SFLA-RFR模型在棉花全生育期LAI監(jiān)測方面具有廣闊的應(yīng)用前景。本研究試驗設(shè)置了不同氮處理和不同棉花品種，但本研究的方法是基于特定地點同一年份的棉花冠層光譜數(shù)據(jù)，這限制了模型對其他數(shù)據(jù)集或其他地域的預測能力。因此，要將FDR-SFLA-RFR模型優(yōu)化至更穩(wěn)定精確，還需要從更多年份、種植模式和地區(qū)收集更多的數(shù)據(jù)集進行模型校正。

4 結(jié) 論

本研究基于無人機獲取棉花冠層高光譜數(shù)據(jù)，通過不同預處理和波段篩選方法篩選波段組合，使用偏最小二乘回歸（Partial Least Squares Regression, PLSR）、支持向量回歸（Support Vector Regression, SVR）和隨機森林回歸（Random Forest Regression, RFR）對棉花全生育期葉面積指數(shù)LAI進行估算，結(jié)果表明：不同LAI的冠層光譜在760～1 000 nm存在明顯差異，冠層光譜與LAI存在明顯的相關(guān)性。對比不同預處理下的波段篩選方法可知，基于相關(guān)系數(shù)進行波段篩選，篩選出的波段過于集中，會出現(xiàn)信息冗余和信息提取不全的現(xiàn)象；而隨機蛙跳（Shuffled Frog Leaping Algorithm, SFLA）算法篩選出的敏感波段分布均勻，對棉花LAI敏感的波段多集中在400～780 nm的可見光波段以及900 nm以后的近紅外波段。不同建模方法的棉花LAI估算模型結(jié)果表現(xiàn)為：RFR最佳，SVR次之，PLSR最差，F(xiàn)DR-SFLA-RFR模型最佳，其建模結(jié)果的2為0.74，RMSE為1.648 3，rRMSE為26.39%；驗證結(jié)果的2為0.67，RMSE為1.622 0，rRMSE為25.97%。

[1] Khan M A, Wahid A, Ahmad M, et al. World Cotton Production and Consumption: An Overview[M]. Cotton Production and Uses. Pakistan: Multan, 2020: 1-7.

[2] 忠智博，翟國亮，鄧忠，等. 水氮施量對膜下滴灌棉花生長及水氮分布的影響[J]. 灌溉排水學報，2020，39(1)：67-76.

Zhong Zhibo, Zhai Guoliang, Deng Zhong, et al. The impact of N application and drip irrigation amount on cotton growth and water and N distributions in soil mulched with film[J]. Journal of Irrigation and Drainage, 2020, 39(1): 67-76. (in Chinese with English abstract）

[3] 馬騰飛，李杰，陳志，等. 施氮對膜下滴灌棉花生長發(fā)育及土壤硝態(tài)氮的影響[J]. 新疆農(nóng)業(yè)科學，2020，57(2)：245-253.

Ma Tengfei, Li Jie, Chen Zhi, et al. Simulating effects of nitrogen application on growth and development and soil nitrate nitrogen of cotton under mulched drip irrigation[J]. Xinjiang Agricultural Sciences, 2020, 57(2): 245-253. (in Chinese with English abstract)

[4] Muharam F M, Delahunty T, Maas S J. Evaluation of nitrogen treatment effects on the reflectance of cotton at different spatial scales[J]. International Journal of Remote Sensing, 2018, 39(23): 8482-8504.

[5] 張宏鳴，劉雯，韓文霆，等. 基于梯度提升樹算法的夏玉米葉面積指數(shù)反演[J]. 農(nóng)業(yè)機械學報，2019，50(5)：258-266.

Zhang Hongming, Liu Wen, Han Wenting, et al. Inversion of summer maize leaf area index based on gradient boosting decision tree algorithm[J]. Transactions of the Chinese Society for Agricultural Machinery, 2019, 50(5): 258-266. (in Chinese with English abstract)

[6] 陶惠林，徐良驥，馮海寬，等. 基于無人機高光譜長勢指標的冬小麥長勢監(jiān)測[J]. 農(nóng)業(yè)機械學報，2020，51(2)：180-191.

Tao Huilin, Xu Liangji, Feng Haikuan, et al. Monitoring of winter wheat growth based on UAV hyperspectral growth index[J]. Transactions of the Chinese Society for Agricultural Machinery, 2020, 51(2): 180-191. (in Chinese with English abstract)

[7] 尹佳碩. 變量施肥夏玉米高光譜與農(nóng)學參數(shù)關(guān)聯(lián)模型研究[D]. 保定：河北農(nóng)業(yè)大學，2018.

Yin Jiashuo. Study on Correlation Model of Hyperspectral and Agronomic Parameters of Summer Maize with Variable Fertilization[D]. Baoding: Hebei Agricultural University, 2018. (in Chinese with English abstract)

[8] 鄧江，谷海斌，王澤，等. 基于無人機遙感的棉花主要生育時期地上生物量估算及驗證[J]. 干旱地區(qū)農(nóng)業(yè)研究，2019，37(5)：55-61，69.

Deng Jiang, Gu Haibin, Wang Ze, et al. Estimation and validation of above-ground biomass of cotton during main growth period using Unmanned[J]. Agricultural Research in the Arid Areas, 2019, 37(5): 55-61, 69. (in Chinese with English abstract)

[9] 劉金然. 基于無人機遙感影像的棉花主要生長參數(shù)反演[D]. 濟南：山東師范大學，2019.

Liu Jinran. Inversion of Cotton Main Growth Parameters Based on Unmanned Aerial Vehicle (UAV) Sensing Image[D]. Jinan: Shandong Normal University, 2019. (in Chinese with English abstract)

[10] Zhu Y, Zhao C, Yang H, et al. Estimation of maize above-ground biomass based on stem-leaf separation strategy integrated with LiDAR and optical remote sensing data[J]. PeerJ, 2019, 7: e7593.

[11] Haldar D, Dave V, Misra A, et al. Radar vegetation index for assessing cotton crop condition using RISAT-1 data[J]. Geocarto International, 2020, 35(4): 364-375.

[12] 依爾夏提·阿不來提，白燈莎·買買提艾力，買買提·沙吾提等. 基于高光譜和BP神經(jīng)網(wǎng)絡(luò)的棉花冠層葉綠素含量聯(lián)合估算[J]. 光學學報，2019，39(9)：372-380.

Ershat Ablet, Baidengsha Maimaitiail, Mamat Sawut, et al. Combined estimation of chlorophyll content in cotton canopy based on hyperspectral parameters and back propagation neural network[J]. Acta Optica Sinica, 2019, 39(9): 372-380. (in Chinese with English abstract)

[13] Xie Q, Dash J, Huete A, et al. Retrieval of crop biophysical parameters from Sentinel-2 remote sensing imagery[J]. International Journal of Applied Earth Observation and Geoinformation, 2019, 80: 187-195.

[14] 張卓然，常慶瑞，張廷龍，等. 基于支持向量機的棉花冠層葉片葉綠素含量高光譜遙感估算[J]. 西北農(nóng)林科技大學學報：自然科學版，2018，46(11)：39-45.

Zhang Zhuoran, Chang Qingrui, Zhang Yanlong, et al. Hyperspectral estimation of chlorophyll content of cotton canopy leaves based on support vector machine [J]. Journal of Northwest A&F University: Natural Science Edition, 2018, 46(11): 39-45. (in Chinese with English abstract)

[15] 劉馨月. 基于熒光參數(shù)的棉花水分狀況的高光譜遙感監(jiān)測研究[D]. 石河子：石河子大學，2018.

Liu Xinyue.Hyperspectral Remote Sensing Monitoring of Cotton Moisture Status Based on Fluorescence Parameters[D]. Shihezi: Shihezi University, 2018. (in Chinese with English abstract)

[16] Tavakoli H, Gebbers R. Assessing nitrogen and water status of winter wheat using a digital camera[J]. Computers and Electronics in Agriculture, 2019, 157: 558-567.

[17] 馬巖川. 基于高光譜遙感的棉花冠層水氮參數(shù)估算[D].北京：中國農(nóng)業(yè)科學院，2020.

Ma Yanchuan. Estimation of Water and Nitrogen Parameters in Cotton Canopy Based on Hyperspectral Remote Sensing[D]. Beijng: Chinese Academy of Agricultural Sciences, 2020. (in Chinese with English abstract)

[18] Thorp K R, Thompson A L, Harders S J, et al. High-throughput phenotyping of crop water use efficiency via multispectral drone imagery and a daily soil water balance model[J]. Remote Sensing, 2018, 10(11): 1682.

[19] Fang H, Baret F, Plummer S, et al. An overview of global leaf area index (LAI): Methods, products, validation, and applications[J]. Reviews of Geophysics, 2019, 57(3): 739-799.

[20] Yu L, Shang J, Cheng Z, et al. Assessment of cornfield LAI retrieved from multi-source satellite data using continuous field LAI measurements based on a wireless sensor network[J]. Remote Sensing, 2020, 12(20): 3304.

[21] Singh P, Pandey P C, Petropoulos G P, et al. Hyperspectral remote sensing in precision agriculture: Present status, challenges, and future trends[M]. Hyperspectral Remote Sensing. Netherlands: Elsevier, 2020: 121-146.

[22] Ca?ete-Salinas P, Zamudio F, Yá?ez M, et al. Evaluation of models to determine LAI on poplar stands using spectral indices from Sentinel-2 satellite images[J]. Ecological Modelling, 2020, 428: 109058.

[23] Caballero D, Calvini R, Amigo J M. Hyperspectral Imaging in Crop Fields: Precision Agriculture[M]. Data Handling in Science and Technology. Netherlands: Elsevier, 2020: 453-473.

[24] 易秋香. 基于Sentinel-2多光譜數(shù)據(jù)的棉花葉面積指數(shù)估算[J]. 農(nóng)業(yè)工程學報，2019，35(16)：189-197.

Yi Qiuxiang. Remote estimation of cotton LAI using Sentinel-2 multispectral data[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(16): 189-197. (in Chinese with English abstract)

[25] Pasqualotto N, Delegido J, Van Wittenberghe S, et al. Multi-crop green LAI estimation with a new simple sentinel-2 LAI index (SeLI)[J]. Sensors, 2019, 19(4): 904.

[26] Li S, Yuan F, Ata-UI-Karim S T, et al. Combining color indices and textures of UAV-based digital imagery for rice LAI estimation[J]. Remote Sensing, 2019, 11(15): 1763.

[27] Wang J, Xiao X, Bajgain R, et al. Estimating leaf area index and aboveground biomass of grazing pastures using Sentinel-1, Sentinel-2 and Landsat images[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 154: 189-201.

[28] Tao H, Feng H, Xu L, et al. Estimation of crop growth parameters using UAV-based hyperspectral remote sensing data[J]. Sensors, 2020, 20(5): 1296.

[29] Zhu W, Sun Z, Huang Y, et al. Improving Field-scale wheat LAI retrieval based on UAV remote-sensing observations and optimized VI-LUTs[J]. Remote Sensing, 2019, 11(20): 2456.

[30] 陳曉凱，李粉玲，王玉娜，等. 無人機高光譜遙感估算冬小麥葉面積指數(shù)[J]. 農(nóng)業(yè)工程學報，2020，36(22)：40-49.

Chen Xiaokai, Li Fenling, Wang Yuna, et al. Estimation of winter wheat leaf area index based on UAV hyperspectral remote sensing[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(22): 40-49. (in Chinese with English abstract)

[31] Peng Y, Li Y, Dai C, et al. Remote prediction of yield based on LAI estimation in oilseed rape under different planting methods and nitrogen fertilizer applications[J]. Agricultural and Forest Meteorology, 2019, 271: 116-125.

[32] Rehman T U, Mahmud M S, Chang Y K, et al. Current and future applications of statistical machine learning algorithms for agricultural machine vision systems[J]. Computers and electronics in agriculture, 2019, 156: 585-605.

[33] Zha H, Miao Y, Wang T, et al. Improving unmanned aerial vehicle remote sensing-based rice nitrogen nutrition index prediction with machine learning[J]. Remote Sensing, 2020, 12(2): 215.

[34] Duan B, Liu Y, Gong Y, et al. Remote estimation of rice LAI based on fourier spectrum texture from UAV image[J]. Plant Methods, 2019, 15(1): 1-12.

[35] Mao H, Meng J, Ji F, et al. Comparison of machine learning regression algorithms for cotton leaf area index retrieval using Sentinel-2 spectral bands[J]. Applied Sciences, 2019, 9(7): 1459.

[36] Cao J, Zhang Z, Tao F, et al. Integrating Multi-Source Data for Rice Yield Prediction across China using Machine Learning and Deep Learning Approaches[J]. Agricultural and Forest Meteorology, 2021, 297: 108275.

[37] Araújo M C U, Saldanha T C B, Galvao R K H, et al. The successive projections algorithm for variable selection in spectroscopic multicomponent analysis[J]. Chemometrics and Intelligent Laboratory Systems, 2001, 57(2): 65-73.

[38] Li H, Liang Y, Xu Q, et al. Key wavelengths screening using competitive adaptive reweighted sampling method for multivariate calibration[J]. Analytica chimica acta, 2009, 648(1): 77-84.

[39] 郭建彪，馬新明，時雷，等. 冬小麥葉面積指數(shù)的品種差異性與高光譜估算研究[J]. 麥類作物學報，2018，38(3)：340-347.

Guo Jianbiao, Ma Xinming, Shi Lei, et al. Variety variation and hyperspectral estimate model of leaf area index of winter wheat[J]. Journal of Triticeae Crops, 2018, 38(3): 340-347. (in Chinese with English abstract)

[40] 李嵐?jié)铎o，明金，等. 冬油菜葉面積指數(shù)高光譜監(jiān)測最佳波寬與有效波段研究[J]. 農(nóng)業(yè)機械學報，2018，49(2)：156-165.

Li Lantao, Li Jing, Ming Jin, et al. Selection optimization of hyperspectral bandwidth and effective wavelength for predicting leaf area index in winter oilseed rape[J]. Transactions of the Chinese Society for Agricultural Machinery, 2018, 49(2): 156-165. (in Chinese with English abstract)

[41] 程雪，賀炳彥，黃耀歡，等. 基于無人機高光譜數(shù)據(jù)的玉米葉面積指數(shù)估算[J]. 遙感技術(shù)與應(yīng)用，2019，34(4)：775-784.

Cheng Xue, He Bingyan, Huang Yaohuan, et al. Estimation of corn leaf area index based on UAV hyperspectral image[J]. Remote Sensing Technology and Application, 2019, 34(4): 775-784. (in Chinese with English abstract)

[42] Baldocchi D D, Ryu Y, Dechant B, et al. Outgoing near‐infrared radiation from vegetation scales with canopy photosynthesis across a spectrum of function, structure, Physiological Capacity, and Weather[J]. Journal of Geophysical Research: Biogeosciences, 2020, 125(7): 1-17.

[43] Li L, Lin D, Wang J, et al. Multivariate analysis models based on full spectra range and effective wavelengths using different transformation techniques for rapid estimation of leaf nitrogen concentration in winter wheat[J]. Frontiers in plant science, 2020, 11: 755.

[44] Ihuoma S O, Madramootoo C A. Sensitivity of spectral vegetation indices for monitoring water stress in tomato plants[J]. Computers and Electronics in Agriculture, 2019, 163: 104860.

[45] 王玉娜，李粉玲，王偉東，等. 基于連續(xù)投影算法和光譜變換的冬小麥生物量高光譜遙感估算[J]. 麥類作物學報，2020，40(11)：1389-1398.

Wang Yuna, Li Fenling, Wang Weidong, et al. Hyper-spectral remote sensing estimation of shoot biomass of winter wheat based on SPA and transformation spectra[J]. Journal of Triticeae Crops, 2020, 40(11): 1389-1398. (in Chinese with English abstract)

[46] Li X, Zhang Y, Bao Y, et al. Exploring the best hyperspectral features for LAI estimation using partial least squares regression[J]. Remote Sensing, 2014, 6(7): 6221-6241.

[47] 孫晶京，楊武德，馮美臣，等. 基于隨機蛙跳和支持向量機的冬小麥葉面積指數(shù)估算[J]. 山西農(nóng)業(yè)大學學報：自然科學版，2020，40(5)：120-128.

Sun Jingjing, Yang Wude, Feng Meichen, et al. Estimation of winter wheat leaf area index based on random leapfrog and support vector regression approach[J]. Journal of Shanxi Agricultural University: Natural Science Edition, 2020, 40(5): 120-128. (in Chinese with English abstract)

[48] Ren G, Wang Y, Ning J, et al. Highly identification of keemun black tea rank based on cognitive spectroscopy: Near infrared spectroscopy combined with feature variable selection[J]. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2020, 230: 118079.

[49] Han L, Yang G, Dai H, et al. Modeling maize above-ground biomass based on machine learning approaches using UAV remote-sensing data[J]. Plant Methods, 2019, 15(1): 1-19.

[50] Lu N, Zhou J, Han Z, et al. Improved estimation of aboveground biomass in wheat from RGB imagery and point cloud data acquired with a low-cost unmanned aerial vehicle system[J]. Plant Methods, 2019, 15(1): 1-16.

[51] Wang X, Miao Y, Dong R, et al. Machine learning-based in-season nitrogen status diagnosis and side-dress nitrogen recommendation for corn[J]. European Journal of Agronomy, 2021, 123: 126193.

[52] 田明璐，班松濤，常慶瑞，等. 基于低空無人機成像光譜儀影像估算棉花葉面積指數(shù)[J]. 農(nóng)業(yè)工程學報，2016，32(21)：102-108.

Tian Minglu, Ban Songtao, Changqingrui, et al. Use of hyperspectral images from UAV-based imaging spectroradiometer to estimate cotton leaf area index[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(21): 102-108. (in Chinese with English abstract).

[53] Chen P. Cotton Leaf Area Index Estimation Using Unmanned Aerial Vehicle Multi-Spectral Images[C]. IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2019: 6251-6254.

[54] Paul A, Ghosh S, Das A K, et al. A review on agricultural advancement based on computer vision and machine learning[M]. Emerging Technology in Modelling and Graphics. Springer, Singapore, 2020: 567-581.

Monitoring of cotton leaf area index using machine learning

Ma Yiru, Lyu Xin, Yi Xiang, Ma Lulu, Qi Yaqin, Hou Tongyu, Zhang Ze※

(/,832003,)

Leaf area index (LAI) is one of the most important indicators that characterize canopy structure and growth of crops. LAI changes can therefore greatly contribute to the variable rate fertilization of cotton. It is of great significance to monitor LAI quickly, accurately, and non-destructively, thereby guiding crop fertilization in modern agriculture. The traditional LAI monitoring relies mainly on manual sampling with high labor intensity and time-consuming. Furthermore, the lagging data cannot meet the needs of real-time monitoring. Most studies on crop LAI have also been made using remote sensing in recent years, such as hand-held spectrometers, unmanned aerial vehicles, and satellites. Nevertheless, the near-earth surface spectrum cannot be used to continuously and rapidly monitor at the spatial scale, due to the limited shooting range and the weight of the instrument. Satellite images are mostly used for the plant LAI monitoring at forest or large regional scale, particularly on the resolution of 10-60m. Alternatively, an Unmanned Aerial Vehicle (UAV) has the potential to fast capture high resolution images repeatedly, suitable for accurate crop monitoring of small plots. Many efforts have been made to monitor the LAI of wheat, rice, corn and others using spectral images under UAVs. Since spectral technology can monitor timely and dynamically, and in macro mode, the resulting LAI spectral data really determines the vegetation index. As such, the hyperspectral reflectance of plant canopy can provide much richer information of vegetation characteristics, compared with vegetation index. However, a large amount of hyperspectral data under UAVs normally presents data redundancy and high multicollinearity. Reasonable spectral transformation can also be utilized to remove the background and noise of hyperspectral data. Correspondingly, machine learning has widely been applied to crop growth monitoring for deep information in data, particularly combined with remote sensing. Great ability of learning and prediction can be achieved using the partial least squares (PLS) model (an extension of multicollinearity model), Support Vector Machine (SVM), and Random Forest (RF), in order to reduce the collinearity between variables in different ways. In this study, the UAV hyperspectral data was preprocessed using the First Derivative (FDR), the Second Derivative (SDR), Savitzky-Golay(SG) smoothing, and Multiple Scatter Correction (MSC) under the plot experiments of different varieties and nitrogen treatments. Sensitive bands were also selected using the Pearson correlation coefficient, Successive Projections Algorithm (SPA), Shuffled Frog Leaping Algorithm (SFLA), and Competitive Adaptive Reweighting (CARS). A cotton LAI monitoring model was finally constructed to calculate the reflectance of selected bands using the Partial Least Square Regression(PLSR), Support Vector Regression (SVR), and Random Forest Regression (RFR). The results showed that the canopy spectra of different LAI were significantly different from 760-1000 nm, where there was a significant correlation between the canopy spectrum and LAI. The sensitive response band of LAI in the cotton canopy was concentrated in the visible light (400-780 nm) and near-infrared (after 900 nm). The highest precision and stability were achieved in the RFR model under each pretreatment for LAI monitoring. Among them, the FDR-SFLA-RFR model performed the best, where the determination coefficient, Root Mean Square Error (RMSE), and relative RMSE for the modeling dataset were 0.74, 1.648 3, and 26.39%, respectively. In the verification dataset, the determination coefficient, RMSE and relative RMSE were 0.67, 1.622 0, and 25.97%, respectively. Consequently, the optimal estimation model can be rationally selected to represent the UAV spectral reflectance of the canopy using various pretreatments, band selecting, and modeling. The findings can provide the potential basis to accurately manage the variable fertilization in cotton fields.

cotton; UAV; hyperspectral; machine learning; leaf area index

馬怡茹，呂新，易翔，等. 基于機器學習的棉花葉面積指數(shù)監(jiān)測[J]. 農(nóng)業(yè)工程學報，2021，37(13)：152-162.

10.11975/j.issn.1002-6819.2021.13.018 http://www.tcsae.org

Ma Yiru, Lyu Xin, Yi Xiang, et al. Monitoring of cotton leaf area index using machine learning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(13): 152-162. (in Chinese with English abstract) doi：10.11975/j.issn.1002-6819.2021.13.018 http://www.tcsae.org

2021-01-07

2021-06-10

兵團重點領(lǐng)域科技攻關(guān)計劃（2020AB005）；兵團重大科技計劃項目（2018AA004）

馬怡茹，研究方向為農(nóng)業(yè)信息化。Email：mayiru@stu.shzu.edu.cn

張澤，博士，副教授，碩士生導師，研究方向為農(nóng)業(yè)信息化技術(shù)及應(yīng)用。Email：zhangze1227@163.com

10.11975/j.issn.1002-6819.2021.13.018

S147.2

A

1002-6819(2021)-13-0152-11