国产日韩欧美一区二区三区三州_亚洲少妇熟女av_久久久久亚洲av国产精品_波多野结衣网站一区二区_亚洲欧美色片在线91_国产亚洲精品精品国产优播av_日本一区二区三区波多野结衣 _久久国产av不卡

?

基于無人機(jī)多光譜遙感的臺(tái)風(fēng)災(zāi)后玉米倒伏信息提取

2021-03-17 07:59閆春雨楊東建溫昱婷黎文華魯力群蘭玉彬
關(guān)鍵詞:植被指數(shù)反射率波段

趙 靜,閆春雨,楊東建,溫昱婷,黎文華,魯力群,蘭玉彬※

·農(nóng)業(yè)航空工程·

基于無人機(jī)多光譜遙感的臺(tái)風(fēng)災(zāi)后玉米倒伏信息提取

趙 靜1,2,閆春雨1,2,楊東建1,2,溫昱婷1,2,黎文華1,2,魯力群2,3,蘭玉彬1,2※

(1. 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;2. 山東理工大學(xué)國(guó)際精準(zhǔn)農(nóng)業(yè)航空應(yīng)用技術(shù)研究中心,淄博 255000;3. 山東理工大學(xué)交通與車輛工程學(xué)院,淄博 255000)

為快速獲取臺(tái)風(fēng)過后玉米倒伏信息,該研究以生態(tài)無人農(nóng)場(chǎng)大田玉米作為研究對(duì)象,利用無人機(jī)搭載多光譜相機(jī)獲取玉米田塊圖像。采用主成分分析(Principal Component Analysis,PCA)變換多光譜圖像,保留信息量最多的前3 個(gè)主成分波段;應(yīng)用最小噪聲分離變換(Minimum Noise Fraction Rotation,MNF)對(duì)48項(xiàng)紋理特征降維,保留信息量最多的前6項(xiàng)特征;計(jì)算選擇10種植被指數(shù);對(duì)多光譜圖像進(jìn)行低通、高通濾波,將以上特征作為全特征集。使用支持向量機(jī)遞歸(Support Vector Machines-Recursive Feature Elimination,SVM-RFE)、 ReliefF和套索算法(Least Absolute Shrinkage and Selection Operator,Lasso)篩選出3種特征子集,建立5種監(jiān)督分類模型,對(duì)4種數(shù)據(jù)集進(jìn)行訓(xùn)練。ReliefF特征子集訓(xùn)練的5種監(jiān)督分類模型測(cè)試集最低分類準(zhǔn)確率為89.02%,SVM-RFE和Lasso特征子集訓(xùn)練的5種監(jiān)督分類模型測(cè)試集最低分類準(zhǔn)確率均為95.38%,與全特征相比僅相差0.58%,表明通過特征篩選方法可在取得較高分類精度同時(shí)大幅減少特征輸入數(shù)量;運(yùn)用3種特征篩選方法與不同分類模型的最佳組合提取驗(yàn)證區(qū)域玉米倒伏信息,通過混淆矩陣驗(yàn)證結(jié)果可知,K最近鄰模型結(jié)合SVM-RFE特征篩選方法分類精度最高,達(dá)93.49%,Kappa系數(shù)為0.90,表明了分類模型普適性較強(qiáng)。該研究使用較少特征數(shù)量參與分類,且獲得較高分類識(shí)別精度,可為無人機(jī)多光譜技術(shù)快速、準(zhǔn)確提取臺(tái)風(fēng)災(zāi)后玉米倒伏信息提供技術(shù)支持。

無人機(jī);遙感;提??;多光譜;玉米;倒伏信息;臺(tái)風(fēng)災(zāi)害

0 引 言

玉米是全球種植總產(chǎn)量最高的農(nóng)作物之一,種植面積僅次于小麥和水稻[1]。玉米產(chǎn)量除受自身遺傳因素影響,還受播種密度、栽培方式及氣候環(huán)境的影響。由于玉米莖稈比較高大,易受暴風(fēng)雨等外力作用發(fā)生倒伏,導(dǎo)致減產(chǎn)、減質(zhì)、無法進(jìn)行機(jī)械化采收[2-4]。近幾年臺(tái)風(fēng)天氣偏多,暴風(fēng)雨不時(shí)發(fā)生,對(duì)玉米產(chǎn)量影響極大,輕則會(huì)對(duì)玉米造成15%~20%的減產(chǎn),嚴(yán)重時(shí)可能造成一半以上的減產(chǎn)[5-6]。及時(shí)、準(zhǔn)確提取玉米倒伏信息,能為災(zāi)后農(nóng)業(yè)生產(chǎn)、政府決策及保險(xiǎn)理賠提供數(shù)據(jù)和技術(shù)支持[7]。

隨著遙感技術(shù)發(fā)展,為快速提取作物倒伏信息提供了多種方法[8-9]。在地面監(jiān)測(cè)方面,傳統(tǒng)獲取作物倒伏信息的方法主要是現(xiàn)場(chǎng)測(cè)量,在作物發(fā)生倒伏災(zāi)害后,測(cè)量人員使用手持GPS、卷尺等工具測(cè)量倒伏面積和位置信息,該方法存在客觀影響大、準(zhǔn)確度低和低效等問題[10]。若農(nóng)田中已產(chǎn)生積水,則無法及時(shí)進(jìn)行地面測(cè)量。倒伏會(huì)影響玉米的冠層結(jié)構(gòu),進(jìn)而改變玉米冠層的光譜特征和輻射傳輸特性,這為地面監(jiān)測(cè)玉米倒伏提供了條件。王猛等[11]通過手持地物光譜儀獲取正常(未倒伏)與倒伏玉米的光譜反射率,結(jié)果表明倒伏玉米的冠層光譜反射率,在可見光和近紅外波段均比正常玉米光譜反射率低;包玉龍等[12]對(duì)USB2000高光譜儀采集的玉米倒伏數(shù)據(jù)進(jìn)行分析后發(fā)現(xiàn),與正常玉米反射率相比,倒伏玉米反射率在可見光波段減小,而在近紅外波段增大,具有明顯的各向異性特征。在衛(wèi)星遙感監(jiān)測(cè)方面,李宗南等[13]利用Worldview-2多光譜影像,對(duì)灌漿期倒伏玉米地塊的光譜和紋理特征進(jìn)行分析估算了玉米倒伏面積;Zhou等[14]通過高分1號(hào)光學(xué)衛(wèi)星獲取區(qū)域尺度玉米倒伏前后影像,計(jì)算倒伏前后植被指數(shù)與光譜反射率變化,采用競(jìng)爭(zhēng)自適應(yīng)加權(quán)算法進(jìn)行特征篩選,利用隨機(jī)森林和最小偏二乘法建立玉米倒伏監(jiān)測(cè)模型,結(jié)果表明隨機(jī)森林建立的模型優(yōu)于最小偏二乘法。王立志等[15]基于HJ-1B多光譜衛(wèi)星數(shù)據(jù)構(gòu)建了比值植被指數(shù)Ratio Vegetation Index,RVI差值的農(nóng)作物倒伏監(jiān)測(cè)模型,對(duì)區(qū)域尺度玉米倒伏進(jìn)行監(jiān)測(cè)與受災(zāi)情況評(píng)估;韓東等[16]基于哨兵-1A衛(wèi)星雷達(dá)后向散射系數(shù)構(gòu)建了區(qū)域尺度下的玉米倒伏監(jiān)測(cè)模型,并對(duì)玉米倒伏程度進(jìn)行了分級(jí)監(jiān)測(cè)。由于倒伏災(zāi)害一般由極端天氣造成,對(duì)衛(wèi)星數(shù)據(jù)質(zhì)量影響很大,待衛(wèi)星再次過境時(shí)獲取的倒伏數(shù)據(jù)對(duì)災(zāi)害評(píng)估已缺乏準(zhǔn)確性和及時(shí)性了。在無人機(jī)遙感監(jiān)測(cè)方面,張新樂等[17]對(duì)倒伏玉米構(gòu)建了5種典型特征組合,結(jié)果顯示多類紋理特征法得到結(jié)果最優(yōu);戴建國(guó)等[18]通過分析倒伏與未倒伏棉花的光譜反射率差異提取了主成分紋理特征與多種植被指數(shù),建立了Logistic模型并進(jìn)行了精度評(píng)價(jià)及驗(yàn)證,測(cè)試集分類結(jié)果準(zhǔn)確率為91.30%;毛智慧等[19]利用無人機(jī)獲取玉米研究區(qū)數(shù)字表面模型,結(jié)合R、G、B色彩特征進(jìn)行分類提取玉米倒伏信息,結(jié)果表明無人機(jī)遙感在小區(qū)尺度上提取玉米倒伏信息是可行的;鄭二功等[20]提出了基于深度學(xué)習(xí)的田間玉米倒伏區(qū)域提取方法,利用無人機(jī)獲取的玉米倒伏圖像制作數(shù)據(jù)集訓(xùn)練分割網(wǎng)絡(luò),結(jié)果表明能夠較準(zhǔn)確地提取玉米倒伏區(qū)域。趙靜等[21]利用ArcGIS中的鑲嵌工具將不同圖像特征進(jìn)行融合,得到數(shù)字地表模型(Digital Surface Model,DSM)+RGB與DSM+過綠指數(shù)(Excess Green,EXG) 2種融合圖像,利用最大似然法和隨機(jī)森林法對(duì)2種特征融合圖像進(jìn)行監(jiān)督分類提取小麥倒伏面積,研究表明通過圖像特征融合的方法能有效提取倒伏小麥信息,為快速提取小麥倒伏面積提供參考。以上研究未充分使用反射率、紋理及植被指數(shù)等多種特征參與分類,特征篩選方法較為單一,也未對(duì)不同特征篩選方法與不同分類方法組合優(yōu)選,進(jìn)行玉米倒伏信息提取。

本研究以山東理工大學(xué)生態(tài)無人農(nóng)場(chǎng)的倒伏玉米為研究對(duì)象,利用四旋翼無人機(jī)獲取的玉米地塊尺度多光譜影像,提取影像的多光譜植被指數(shù)、反射率和紋理等特征。將提取的所有特征作為全特征集,將經(jīng)過3種典型算法篩選出的特征作為不同子集,采用5種監(jiān)督分類模型對(duì)4種特征集進(jìn)行分類提取研究區(qū)內(nèi)玉米倒伏信息,以期得到最佳特征集和分類模型的組合,為精準(zhǔn)、快速掌握玉米倒伏災(zāi)損提供參考。

1 材料與方法

1.1 研究區(qū)概況

研究區(qū)地點(diǎn)為山東理工大學(xué)生態(tài)無人農(nóng)場(chǎng)試驗(yàn)田,位于山東省淄博市朱臺(tái)鎮(zhèn)(36°57′15.30″N,118°13′1.00″E),該地區(qū)屬溫帶季風(fēng)氣候,降水季分布不均衡,全年降水量有60%~70%集中于夏季,多年平均年降水量為679.5 mm,無霜期一般為174~260 d。研究區(qū)總面積約5 000 m2,試驗(yàn)田內(nèi)玉米小麥輪作,玉米種植時(shí)間為2019年6月中旬。

1.2 數(shù)據(jù)獲取

2019年8月11—13日,受第九號(hào)臺(tái)風(fēng)“利奇馬”影響,研究區(qū)發(fā)生了大面積玉米倒伏(玉米正處于抽雄-吐絲期)。本試驗(yàn)采用深圳市大疆創(chuàng)新科技有限公司生產(chǎn)的經(jīng)緯M210 V2無人機(jī),機(jī)體最大起飛質(zhì)量為6.14 kg,機(jī)身軸距643 mm,最大水平飛行速度為20.5 m/s,最大承受風(fēng)速12 m/s,續(xù)航時(shí)間為34 min。

利用長(zhǎng)光禹辰信息技術(shù)與裝備(青島)有限公司自主研發(fā)的MS600Pro多光譜相機(jī)(參數(shù)如表1所示)獲取圖像數(shù)據(jù),該相機(jī)包含藍(lán)光、綠光、紅光、紅邊和2個(gè)近紅外波段。獲取的圖像像素為120萬、分辨率為1 280×960,所有圖像儲(chǔ)存在相機(jī)自帶內(nèi)存卡中。由于臺(tái)風(fēng)及雨水持續(xù)時(shí)間較長(zhǎng),玉米植株生長(zhǎng)高度較高,采集試驗(yàn)數(shù)據(jù)時(shí)(2019年8月15日11:00—14:00),受影響較小的玉米植株?duì)顟B(tài)處于自行立起前期,受影響較大玉米植株需7~10 d才能基本恢復(fù)正常。采集當(dāng)日天氣晴朗無云,無人機(jī)飛行高度為50 m,飛行速度3 m/s,航向旁向重疊率均為75%,共獲取多光譜圖像828張。獲取的多光譜圖像,通過Pix4Dmapper軟件進(jìn)行輻射定標(biāo)、幾何校正和拼接等預(yù)處理,得到6張單波段tif圖像,通過ENVI 5.3軟件波段合成(Layer stacking)功能將6幅單波段圖像合成多光譜圖像。

表1 試驗(yàn)用MS600Pro相機(jī)參數(shù)

1.3 研究方案

首先計(jì)算研究區(qū)多光譜圖像的10種多光譜植被指數(shù),其次使用主成分分析法(Principal Component Analysis,PCA)對(duì)6波段多光譜圖像進(jìn)行降維,再使用最小噪聲分離變換(Minimum Noise Fraction Rotation,MNF)對(duì)6波段多光譜圖像的紋理濾波特征進(jìn)行降維。將試驗(yàn)區(qū)按2∶1的比例劃分為目標(biāo)區(qū)域和驗(yàn)證區(qū)域(圖 1)。

目標(biāo)區(qū)域圖像是由包含不同植被指數(shù)、高通濾波反射率、低通濾波反射率和紋理濾波特征合成的多波段圖像,在RGB彩色顯示模式下,倒伏區(qū)域玉米亮度較高且顏色與正常(未倒伏)玉米相較顏色淺,土壤背景顏色與倒伏玉米和未倒伏玉米相差較大,上述三類地物特征差異較明顯,通過目視解譯方式,使用ENVI感興趣區(qū)域工具(Region of Interest Tool)提取目標(biāo)區(qū)域內(nèi)土壤背景、正常玉米和倒伏玉米樣本,進(jìn)而獲取目標(biāo)區(qū)域內(nèi)三類地物的植被指數(shù)、高通濾波反射率、低通濾波反射率和紋理濾波特征。利用ReliefF、支持向量機(jī)遞歸(Support Vector Machines-Recursive Feature Elimination,SVM-RFE)和套索算法(Least Absolute Shrinkage and Selection Operator,Lasso)三種特征篩選方法篩選特征創(chuàng)建數(shù)據(jù)子集。將目標(biāo)區(qū)域內(nèi)樣本數(shù)據(jù)按照3∶1比例劃分為訓(xùn)練集樣本和測(cè)試集樣本,通過多種監(jiān)督分類模型進(jìn)行訓(xùn)練,選取最佳特征篩選方法及精度最高的監(jiān)督分類器,對(duì)目標(biāo)區(qū)域進(jìn)行玉米倒伏識(shí)別。玉米倒伏信息提取技術(shù)路線如圖2所示。

1.4 分類特征提取

1.4.1 反射率特征提取

高通濾波在保持圖像高頻信息時(shí)消除了圖像中的低頻成分,它可用來增強(qiáng)紋理、邊緣等信息;低頻濾波保存了圖像中的低頻成分,使圖像平滑。本研究選擇多光譜圖像6個(gè)波段的6個(gè)低通濾波反射率和6個(gè)高通濾波反射率。

1.4.2 植被指數(shù)提取

植被指數(shù)被大量用在遙感領(lǐng)域,用于評(píng)價(jià)植被覆蓋度、植被生長(zhǎng)狀況。本試驗(yàn)選擇常用的10種多光譜植被指數(shù),如表2所示。

表2 用于玉米倒伏信息提取的多光譜植被指數(shù)

注:NIR為近紅外波段反射率,RED為紅光波段反射率,BLE為藍(lán)光波段反射率,GRE為綠光波段反射率,REG為紅邊波段反射率。

Note: NIR is the reflectance of near-infrared band; RED is the reflectance of red band; BLE is the reflectance of blue band; GRE is the reflectance of green band; REG is the reflectance of red edge band.

1.4.3 紋理特征提取

紋理特征通過灰度空間變化及其重復(fù)性反映地物的視覺粗糙度,能充分反映圖像特征,不同物體表現(xiàn)出的紋理類型一般不同,可用于描述和識(shí)別地物。同一類別地物整體表征看似相似,但局部細(xì)節(jié)紋理特征有區(qū)別[31]。在各種紋理分析方法中,灰度共生矩陣分析方法是認(rèn)可度較高的方法之一,具有較強(qiáng)的魯棒性和適應(yīng)能力。本研究選用基于二階概率統(tǒng)計(jì)濾波的角二階矩(Angular Second Moment)、對(duì)比度(Contrast)、相關(guān)性(Correlation)、相異性(Dissimilarity)、熵(Entropy)、協(xié)同性(Homogeneity)、均值(Mean)、方差(Variance)8種紋理特征,通過基于最小噪聲分離變換和基于主成分分析的方法對(duì)原始6通道多光譜圖像產(chǎn)生的48項(xiàng)紋理特征進(jìn)一步篩選。

1)基于最小噪聲分離變換的紋理濾波特征提取

本試驗(yàn)引入MNF對(duì)48項(xiàng)紋理特征進(jìn)行降維,降低特征數(shù)量,圖3為MNF權(quán)重計(jì)算結(jié)果,橫坐標(biāo)為波段數(shù),縱坐標(biāo)為貢獻(xiàn)值,自第7個(gè)波段開始,每個(gè)波段的貢獻(xiàn)值趨于一致,且對(duì)分類影響較小,故選擇前6個(gè)波段作為最終紋理濾波特征。

2)基于主成分分析的紋理濾波特征提取

多光譜圖像各波段間具有很高相關(guān)性,主成分分析能去除波段間冗余信息,保留對(duì)分類更有利的信息,通過主成分分析發(fā)現(xiàn)6個(gè)波段中前3個(gè)波段的信息量已超過95%,對(duì)前3個(gè)主成分波段進(jìn)行紋理濾波計(jì)算,共產(chǎn)生24項(xiàng)紋理濾波特征。

1.5 特征算法及監(jiān)督分類模型選取

全特征集包含20種植被指數(shù)特征、6項(xiàng)高通濾波特征、6項(xiàng)低通濾波特征、經(jīng)過PCA降維后的24項(xiàng)紋理特征和經(jīng)過MNF降維后的6項(xiàng)特征。

特征篩選的目的是減輕數(shù)據(jù)維度災(zāi)難,去除無關(guān)特征或相關(guān)性較低特征,提高模型運(yùn)算效率。常用的特征篩選方法大致分為過濾式(Filter)、包裹式(Wrapper)和嵌入式(Embedding)[32-33]三類,過濾式特征篩選法先選定特征再進(jìn)行學(xué)習(xí),具有較強(qiáng)通用性;包裹式特征篩選方法利用學(xué)習(xí)算法的性能來評(píng)價(jià)自身優(yōu)劣,篩選得到特征子集分類性能較好;嵌入式特征篩選方法將特征選擇過程作為學(xué)習(xí)過程的一部分,在學(xué)習(xí)過程中自動(dòng)進(jìn)行特征篩選,優(yōu)點(diǎn)是效果最好,速度最快,模式單調(diào)。

ReliefF是Relief算法的拓展,用于兩類以上樣本特征篩選,屬于典型的過濾式特征篩選方法[34-35]。該算法每次從訓(xùn)練集樣本中集中隨機(jī)取出一個(gè)樣本,然后從的同類樣本中找出的個(gè)近鄰樣本,從每個(gè)的不同類別樣本集中均找出個(gè)近鄰樣本,然后更新每個(gè)特征的權(quán)重。ReliefF特征篩選算法通過Python程序?qū)崿F(xiàn),對(duì)全特征集62項(xiàng)特征進(jìn)行篩選,選擇權(quán)重大于0.04的10項(xiàng)特征作為ReliefF特征子集,分別為增強(qiáng)型植被指數(shù)EVI840、歸一化植被指數(shù)NDVI840、歸一化植被指數(shù)NDVI940、優(yōu)化土壤調(diào)節(jié)植被指數(shù)OSAVI840、重歸一化植被指數(shù)RDVI840、土壤調(diào)整植被指數(shù)SAVI840、MNF第二紋理波段、MNF第四紋理波段、MNF第五紋理波段和藍(lán)波段低通反射率。

SVM-RFE是一種包裹式特征篩選方法,該算法使用一個(gè)基模型進(jìn)行多次迭代訓(xùn)練,每次訓(xùn)練結(jié)束均會(huì)根據(jù)每個(gè)特征的系數(shù)對(duì)特征進(jìn)行打分,去掉得分最小特征,利用其余特征構(gòu)建新特征集繼續(xù)迭代計(jì)算,直到篩選出合適特征[36]。該算法通過Python程序篩選出13項(xiàng)特征,分別為增強(qiáng)型EVI940、綠色歸一化植被指數(shù)GNDVI940、優(yōu)化土壤調(diào)節(jié)植被指數(shù)OSAVI840、比值植被指數(shù)RVI840、比值植被指數(shù)RVI840、土壤調(diào)節(jié)植被指數(shù)SAVI840、土壤調(diào)節(jié)植被指數(shù)SAVI940、轉(zhuǎn)換型指數(shù)TVI840、MNF第三紋理波段、MNF第四紋理波段、藍(lán)波段低通反射率、綠波段低通反射率和第三主成分方差。

Lasso屬于嵌入式特征篩選方法,采用L1正則化(L1-regularization)線性回歸將對(duì)分類貢獻(xiàn)小或無貢獻(xiàn)的特征權(quán)值降為0,達(dá)到稀疏化和特征選擇的目的[37-38]。該算法同樣通過Python程序?qū)崿F(xiàn),選擇權(quán)重大于0.01的12項(xiàng)特征作為L(zhǎng)asso特征子集,分別為比值植被指數(shù)RVI940、轉(zhuǎn)換型指數(shù)TVI840、MNF第二紋理波段、MNF第三紋理波段、MNF第五紋理波段、第一主成分熵值、第一主成分相關(guān)性、第二主成分均值、第二主成分相異性、第二主成分相關(guān)性、第三主成分均值和第三主成分相異性。

在目標(biāo)區(qū)域用ENVI 5.3感興趣工具選擇690個(gè)樣本組成樣本數(shù)據(jù)集,每個(gè)樣本包含20項(xiàng)植被指數(shù)特征、PCA降維后前3個(gè)主成分波段的24項(xiàng)紋理濾波特征、MNF篩選后的6項(xiàng)紋理濾波特征、6項(xiàng)低通濾波反射率和6項(xiàng)高通濾波反射率特征。隨機(jī)將樣本數(shù)據(jù)集中515個(gè)樣本作為訓(xùn)練集,175個(gè)樣本作為測(cè)試集,訓(xùn)練集每個(gè)類別具體樣本數(shù)為:土壤及背景樣本訓(xùn)練樣本112個(gè)、正常玉米訓(xùn)練樣本177個(gè)、倒伏玉米訓(xùn)練樣本226個(gè)。測(cè)試集每個(gè)類別具體樣本數(shù)為:土壤及背景樣本測(cè)試樣本40個(gè)、正常玉米測(cè)試樣本60個(gè)、倒伏玉米測(cè)試樣本75個(gè)。利用常用的樸素貝葉斯(Naive Bayes,NB)、K近最鄰(K-Nearest Neighbor,KNN)、決策樹(Decision Tree,DT)、支持向量機(jī)(Support Vector Machine,SVM)和人工神經(jīng)網(wǎng)絡(luò)(Artificial Neural Network,ANN)模型對(duì)訓(xùn)練集進(jìn)行模型構(gòu)建,通過10折交叉驗(yàn)證法及網(wǎng)格搜索對(duì)模型進(jìn)行優(yōu)化,利用測(cè)試集獲得5種模型的分類精度。

1.6 模型分類精度評(píng)價(jià)

使用準(zhǔn)確率()、精確率()、召回率()和精確率和召回率的調(diào)和平均數(shù)(1)對(duì)5種模型的分類結(jié)果進(jìn)行精度評(píng)價(jià)。4種評(píng)價(jià)指標(biāo)計(jì)算公式[39]為

式中TP、FP、TN和FN分別為真正例、假正例、真反例和假反例,是預(yù)測(cè)正確的樣本數(shù)量占總體樣本的比值,是真正例占所有預(yù)測(cè)正比例樣本的比值,表示預(yù)測(cè)正例的樣本數(shù)占所有正例樣本的比值,1是精確率與召回率的調(diào)和平均值。

1.7 驗(yàn)證區(qū)域精度評(píng)價(jià)

混淆矩陣可清楚列出每類地物正確分類個(gè)數(shù)、錯(cuò)分類別和錯(cuò)分個(gè)數(shù)。單純依據(jù)混淆矩陣無法評(píng)價(jià)分類精度的優(yōu)劣,因此由混淆矩陣衍生出了多種分類精度指標(biāo)用于分類模型精度評(píng)價(jià),其中應(yīng)用最廣泛的指標(biāo)有總體分類精度(Overall accuracy)和Kappa系數(shù)[40]。

本研究選擇總體分類精度和Kappa系數(shù)作為驗(yàn)證區(qū)域精度主要評(píng)價(jià)指標(biāo)。

2 結(jié)果與分析

2.1 模型分類結(jié)果對(duì)比

如表3所示,采用5種監(jiān)督分類模型分別對(duì)ReliefF、SVM-RFE和Lasso篩選出的特征子集進(jìn)行訓(xùn)練和測(cè)試與全特征數(shù)據(jù)集監(jiān)督分類結(jié)果相比,在全特征數(shù)據(jù)集分類結(jié)果中,SVM模型分類準(zhǔn)確率最高(97.69%);ReliefF篩選的特征子集分類結(jié)果中,KNN模型分類準(zhǔn)確率最高(98.84%);SVM-RFE特征子集分類結(jié)果中,KNN模型分類準(zhǔn)確率最高(98.84%);Lasso特征子集分類結(jié)果中,ANN模型分類準(zhǔn)確率最高(98.27%)。ReliefF特征子集訓(xùn)練的5種監(jiān)督分類模型訓(xùn)練集最低分類準(zhǔn)確率為89.02%,SVM-RFE和Lasso特征子集訓(xùn)練的5種監(jiān)督監(jiān)督分類模型訓(xùn)練集最低分類準(zhǔn)確率均為95.38%,與全特征集最低分類精度94.80%僅相差0.58%。

2.2 驗(yàn)證區(qū)域分類效果的對(duì)比與分析

分別利用對(duì)測(cè)試集分類準(zhǔn)確率較高的SVM、KNN和ANN模型,對(duì)驗(yàn)證區(qū)采用全特征、ReliefF特征子集、SVM-RFE特征子集和Lasso特征子集進(jìn)行玉米倒伏識(shí)別提取,通過人工目視解譯對(duì)驗(yàn)證區(qū)域地物劃分感興趣區(qū)域作為地面真值,驗(yàn)證區(qū)結(jié)果如圖4所示。驗(yàn)證區(qū)域不同特征子集的混淆矩陣、總體分類精度及Kappa系數(shù)如表4所示。

表3 5種模型分類精度的評(píng)價(jià)與比較

注:NB為樸素貝葉斯,KNN為K最近鄰,SVM為支持向量機(jī),DT為決策樹,ANN為人工神經(jīng)網(wǎng)絡(luò)。1是精確率和召回率的調(diào)和平均數(shù)。下同。

Note: NB is Naive Bayes; KNN is K Nearest Neighbor; SVM is Support Vector Machine; DT is Decision Tree; ANN is Artificial Neural Network.1is the harmonic mean of accuracy and precision. The same as below.

比較不同特征數(shù)據(jù)集的總體分類精度可知,在全部特征數(shù)據(jù)集分類結(jié)果中,SVM模型達(dá)到最高分類精度所需分類特征也最多,其分類精度最高,Kappa系數(shù)為0.88;ANN分類模型最適合與ReliefF特征篩選方法所篩選特征數(shù)據(jù)集進(jìn)行組合,分類精度最高,Kappa系數(shù)為0.84;對(duì)于SVM-RFE特征篩選方法所選的特征數(shù)據(jù)集,結(jié)合KNN模型進(jìn)行分類的總體精度最高,達(dá)93.49%,Kappa系數(shù)為0.90;對(duì)于Lasso特征篩選方法,參與特征數(shù)量為12個(gè),結(jié)合ANN模型進(jìn)行分類更具優(yōu)勢(shì),總體精度最高,達(dá)91.77% Kappa系數(shù)為0.88。對(duì)不同特征篩選方法,不同分類模型也具有不同優(yōu)勢(shì),在4 種不同特征數(shù)據(jù)集分類結(jié)果中,當(dāng)參與的分類特征數(shù)量急劇減少時(shí),ANN模型相對(duì)其他2種模型總體分類精度波動(dòng)非常??;當(dāng)參與分類特征數(shù)量急劇減少時(shí),SVM模型總體分類精度波動(dòng)最大,最高分類精度與最低分類精度的差值為43.59。KNN模型結(jié)合SVM-RFE特征篩選方法既減少了參與分類的特征數(shù)量,還能得到較優(yōu)分類效果。

在驗(yàn)證區(qū)玉米倒伏分類結(jié)果中,3種特征子集的總體分類精度與全特征數(shù)據(jù)集總體分類精度相比有所下降,除ReliefF特征子集的總體分類精度最低外,其余2種特征子集精度下降幅度較小,且SVM-RFE特征子集中ANN與KNN模型分類精度相對(duì)全特征數(shù)據(jù)集的總體分類精度有所提高。結(jié)果表明,SVM-RFE算法不僅可以解決數(shù)據(jù)維度災(zāi)難,消除冗余的特征,還可以提高某些模型的總體分類精度。

3 結(jié) 論

本研究通過ReliefF、SVM-RFE及Lasso 特征篩選方法,對(duì)多光譜圖像的紋理特征、植被指數(shù)特征及反射率特征等62項(xiàng)特征進(jìn)行篩選,并通過NB、KNN、DT、SVM和ANN監(jiān)督分類模型對(duì)不同特征組成的子集進(jìn)行分類及驗(yàn)證,通過對(duì)比分類精度選出最優(yōu)特征與監(jiān)督分類模型組合,將最優(yōu)模型用于驗(yàn)證區(qū)域玉米倒伏信息提取,并使用混淆矩陣進(jìn)行驗(yàn)證,結(jié)果表明:

1)ReliefF、SVM-RFE與Lasso特征篩選算法均可在有效降低數(shù)據(jù)維度同時(shí)保持較高分類精度,SVM-RFE和Lasso特征算法測(cè)試集最低分類準(zhǔn)確率相同(均為95.38%)與全特征數(shù)據(jù)集最低分類準(zhǔn)確度(94.80%)更為接近,僅相差0.58%,表明通過特征篩選方法可大幅減少參與分類的特征數(shù)量,且可取得較高分類精度。

2)運(yùn)用不同特征篩選方法與5種監(jiān)督分類模型的最佳組合提取驗(yàn)證區(qū)域玉米倒伏信息,通過混淆矩陣驗(yàn)證結(jié)果可知,KNN和ANN模型能有效識(shí)別土壤背景、正常玉米和倒伏玉米,總體精度最高達(dá)93.49%和91.77%,Kappa系數(shù)分別為0.90和0.88,KNN模型結(jié)合SVM-RFE特征篩選方法分類結(jié)果最好。

3)根據(jù)驗(yàn)證區(qū)域總體分類精度可知,本研究所得最優(yōu)分類模型普適性較強(qiáng),能對(duì)其他同樣受到倒伏脅迫玉米地塊應(yīng)用。

本研究采用組合優(yōu)選、先復(fù)雜后簡(jiǎn)化的方法提取玉米倒伏信息,在減少特征參與且對(duì)分類精度影響較小前提下,完成玉米倒伏信息提取識(shí)別,但仍存在一些問題,有待進(jìn)一步研究:

1)多光譜相機(jī)價(jià)格較可見光相機(jī)價(jià)格高,在后續(xù)研究中,考慮使用可見光無人機(jī)獲取作物倒伏圖像,通過不同特征篩選及分類方法進(jìn)行組合,選出最優(yōu)分類組合方法提取作物倒伏信息。

2)本研究?jī)H識(shí)別提取了由臺(tái)風(fēng)災(zāi)害引起的玉米倒伏信息,并未對(duì)由其他原因且在不同生長(zhǎng)期受到倒伏脅迫的作物進(jìn)行識(shí)別提取。

3)本研究未對(duì)倒伏嚴(yán)重程度進(jìn)行詳細(xì)區(qū)分級(jí)。在后續(xù)研究中,考慮加入非監(jiān)督分類方法對(duì)倒伏玉米樣本進(jìn)行聚類分析,結(jié)合作物實(shí)際倒伏情況,對(duì)非監(jiān)督分類方法進(jìn)行精度評(píng)價(jià)。本研究試驗(yàn)是在像素級(jí)別上進(jìn)行的,后期試驗(yàn)也會(huì)考慮用監(jiān)督分類與面向?qū)ο蠓诸愃枷耄M(jìn)行作物倒伏識(shí)別和嚴(yán)重程度分級(jí)。

[1] 李闊,熊偉,潘婕,等. 未來升溫1.5 ℃與2.0 ℃背景下中國(guó)玉米產(chǎn)量變化趨勢(shì)評(píng)估[J]. 中國(guó)農(nóng)業(yè)氣象,2018,39(12):765-777.

Li Kuo, Xiong Wei, Pan Jie, et al. Trend evaluation on changes of maize yield in China under global warming by 1.5°C and 2.0°C[J]. Chinese Journal of Agrometeorology, 2018, 39(12): 765-777. (in Chinese with English abstract)

[2] 王杰,劉實(shí),蘭玉彬,等. 基于HJ-1A/BCCD數(shù)據(jù)的玉米倒伏識(shí)別方法[J]. 中國(guó)農(nóng)業(yè)氣象,2020,41(2):121-128.

Wang Jie, Liu Shi, Lan Yubin, et al. Method of maize lodging recognition based on HJ-1A/B CCD data[J]. Chinese Journal of Agrometeorology, 2020, 41(2): 121-128. (in Chinese with English abstract)

[3] Zhang H, Zhang L, Niu Y, et al. Comparison of water stress coefficient using three alternative canopy temperature-based indices[J]. International Journal of Precision Agricultural Aviation, 2020, 3(2): 28-34.

[4] Alexander Q S, Fletcher G, Jo H D, et al. An automatable, field camera track system for phenotyping crop lodging and crop movement[J]. HardwareX, 2018, 4: e00029.

[5] 李樹巖,王宇翔,胡程達(dá),等. 抽雄期前后大風(fēng)倒伏對(duì)夏玉米生長(zhǎng)及產(chǎn)量的影響[J]. 應(yīng)用生態(tài)學(xué)報(bào),2015,26(8):2405-2413.

Li Shuyan, Wang Yuxiang, Hu Chengda, et al. Effects of strong wind lodging at pre- and post-tasseling stages on growth and yield of summer maize[J]. Chinese Journal of Applied Ecology, 2015, 26(8): 2405-2413. (in Chinese with English abstract)

[6] 劉戰(zhàn)東,肖俊夫,南紀(jì)琴,等. 倒伏對(duì)夏玉米葉面積、產(chǎn)量及其構(gòu)成因素的影響[J]. 中國(guó)農(nóng)學(xué)通報(bào),2010,26(18):107-110.

Liu Zhandong, Xiao Junfu, Nan Jiqin, et al. Effect of different levels lodging on leaf area index, yield and its components of summer maize[J]. Chinese Agricultural Science Bulletin, 2010, 26(18): 107-110. (in Chinese with English abstract)

[7] Chauhan S, Darvishzadeh R, Boschetti M, et al. Remote sensing-based crop lodging assessment: Current status and perspectives[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 151: 124-140.

[8] Du X Y, Wan L, Cen H Y, et al. Multi-temporal monitoring of leaf area index of rice under different nitrogen treatments using UAV images. International Journal of Precision Agricultural Aviation, 2020, 3(1): 7-12.

[9] Geng Y, Dong Y Y, Zhao L L, et al. Desert locust plague monitoring using time series satellite data. International Journal of Precision Agricultural Aviation, 2020, 3(4): 24–30.

[10] 李明,黃愉淇,李緒孟,等. 基于無人機(jī)遙感影像的水稻種植信息提取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(4):108-114.

Li Ming, Huang Yuqi, Li Xumeng, et al. Extraction of rice planting information based on remote sensing image from UAV[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(4): 108-114. (in Chinese with English abstract)

[11] 王猛,隋學(xué)艷,梁守真,等. 夏玉米倒伏模擬試驗(yàn)及遙感監(jiān)測(cè)[J]. 測(cè)繪科學(xué),2017,42(8):137-141.

Wang Meng, Sui Xueyan, Liang Shouzhen, et al. Simulation testand remote sensing monitoring of summer corn lodging[J]. Science of Surveying and Mapping, 2017, 42(8): 137-141. (in Chinese with English abstract)

[12] 包玉龍,張繼權(quán),劉曉靜,等. 受風(fēng)災(zāi)影響后的農(nóng)作物冠層反射信息測(cè)量與分析[J]. 光譜學(xué)與光譜分析,2013,33(4):1057-1060.

Bao Yulong, Zhang Jiquan, Liu Xiaojing, et al. Measurement and analysis of reflected information from crops canopy suffering from wind disaster influence[J]. Spectroscapy and Spectral Analysis, 2013, 33(4): 1057-1060. (in Chinese with English abstract)

[13] 李宗南,陳仲新,任國(guó)業(yè),等. 基于Worldview-2影像的玉米倒伏面積估算[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(2):1-5.

Li Zongnan, Chen Zhongxin, Ren Guoye, et al. Estimation of maize lodging area based on worldview-2 image[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(2): 1-5. (in Chinese with English abstract)

[14] Zhou L F, Cheng S, Sun Q, et al. Remote sensing of regional-scale maize lodging using multitemporal GF-1 images[J]. Journal of Applied Remote Sensing, 2020, 14(1): 014514.

[15] 王立志,顧曉鶴,胡圣武,等. 基于多時(shí)相HJ-1B CCD影像的玉米倒伏災(zāi)情遙感監(jiān)測(cè)[J]. 中國(guó)農(nóng)業(yè)科學(xué),2016,49(21):4120-4129.

Wang Lizhi, Gu Xiaohe, Hu Shengwu, et al. Remote sensing monitoring of maize lodging disaster with multi-temporal HJ-1B CCD Image[J]. Scientia Agricultura Sinica, 2016, 49(21): 4120-4129. (in Chinese with English abstract)

[16] 韓東,楊浩,楊貴軍,等. 基于Sentinel-1雷達(dá)影像的玉米倒伏監(jiān)測(cè)模型[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(3):166-172.

Han Dong, Yang Hao, Yang Guijun, et al. Monitoring model of maize lodging based on Sentinel-1 radar image[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(3): 166-172. (in Chinese with English abstract)

[17] 張新樂,官海翔,劉煥軍,等. 基于無人機(jī)多光譜影像的完熟期玉米倒伏面積提取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(19):98-106.

Zhang Xinyue, Guan Haixiang, Liu Huanjun, et al. UAV crops multispectral lodging feature combination multi-texture features[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(19): 98-106. (in Chinese with English abstract)

[18] 戴建國(guó),張國(guó)順,郭鵬,等. 基于無人機(jī)遙感多光譜影像的棉花倒伏信息提取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(2):63-70.

Dai Jianguo, Zhang Guoshun, Guo Peng, et al. Information extraction of cotton lodging based on multi-spectral image from UAV remote sensing[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(19): 63-70. (in Chinese with English abstract)

[19] 毛智慧,鄧?yán)?,趙曉明,等. 利用無人機(jī)遙感提取育種小區(qū)玉米倒伏信息[J]. 中國(guó)農(nóng)學(xué)通報(bào),2019,35(3):62-68.

Mao Zhihui, Deng Lei, Zhao Xiaoming, et al. Extraction of maize lodging in breeding plot based on UAV remote sensing[J]. Chinese Agricultural Science Bulletin, 2019, 35(3): 62-68. (in Chinese with English abstract)

[20] 鄭二功,田迎芳,陳濤. 基于深度學(xué)習(xí)的無人機(jī)影像玉米倒伏區(qū)域提取[J]. 河南農(nóng)業(yè)科學(xué),2018,47(8):155-160.

Zheng Ergong, Tian Yingfang, Chen Tao, et al. Region extraction of corn lodging in UAV images based on deep learning[J]. Journal of Henan Agricultural Sciences, 2018, 47(8): 155-160. (in Chinese with English abstract)

[21] 趙靜,潘方江,蘭玉彬,等. 無人機(jī)可見光遙感和特征融合的小麥倒伏面積提取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(3):73-80.

Zhao Jing, Pan Fangjiang, Lan Yubin, et al. Wheat lodging area extraction using UAV visible light remote sensing and feature fusionp[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(3): 73-80. (in Chinese with English abstract)

[22] Jordan C F. Derivation of leaf-area index from quality of light on the forest floor[J]. Ecology, 1969, 50(4): 663-666.

[23] 王正興,劉闖,Huete Alfredo. 植被指數(shù)研究進(jìn)展:從AVHRR-NDVI到MODIS-EVI[J]. 生態(tài)學(xué)報(bào),2003,23(5):979-987.

Wang Zhengxing, Liu Chuang, Huete Alfredo. From AVHRR-NDVI to MODIS-EVI: Advances in vegetation index research[J]. Acta Ecologica Sinica, 2003, 23(5): 979-987. (in Chinese with English abstract)

[24] Gitelson A A, Kaufman Y J, Merzlyak M N. Use of a green channel in remote sensing of global vegetation from EOS-MODIS[J]. Remote Sensing of Environment, 1996, 58(3): 289-298.

[25] Barnes E M, Clarke T R, Richards S E, et al. Coincident detection of crop water stress, nitrogen status and canopy density using ground-based multispectral data[C]//Proceedings of the 5th International Conference on Precision Agriculture and Other Resource Management. Bloomington, USA: University of Minnesota, 2000.

[26] Rouse J W. Monitoring the vernal advancement and retrogradation (green wave effect) of natural vegetation[J]. NASA, 1974: 1-12.

[27] Gitelson A A, Kaufman Y J, Merzlyak M N. Use of a green channel in remote sensing of global vegetation from EOS-MODIS[J]. Remote Sensing of Environment, 1996, 58(3): 289-298.

[28] Huete A R. A soil-adjusted vegetation index (SAVI)[J]. Remote Sensing of Environment, 1988, 25(3): 295-309.

[29] Roujean J L, Breon F M. Estimating PAR absorbed by vegetation from bidirectional reflectance measurements[J]. Remote Sensing of Environment, 1995, 51(3): 375-384.

[30] Broge N H, Leblanc E. Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density[J]. Remote Sensing of Environment, 2001, 76(2): 156-172.

[31] 朱秀芳,李石波,肖國(guó)峰. 基于無人機(jī)遙感影像的覆膜農(nóng)田面積及分布提取方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(4):106-113.

Zhu Xiufang, Li Shibo, Xiao Guofeng, et al. Method on extraction of area and distribution of plastic-mulched farmland based on UAV images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(4): 106-113. (in Chinese with English abstract)

[32] 季金勝. 高分辨率遙感影像典型地物目標(biāo)的特征選擇及其穩(wěn)定性研究[D]. 上海:上海交通大學(xué),2015.

Ji Jinsheng. Feature Selection and Its Stability for Typical Geoobjects of the High-Resolution Remote Sensing Image[D]. Shanghai: Shanghai Jiao Tong University, 2015. (in Chinese with English abstract)

[33] 肖艷,姜琦剛,王斌,等. 基于ReliefF和PSO混合特征選擇的面向?qū)ο笸恋乩梅诸怺J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(4):211-216.

Xiao Yan, Jiang Qigang, Wang Bin, et al. Object based land-use classification based on hybrid feature selection method of combining Relief F and PSO[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(4): 211-216. (in Chinese with English abstract)

[34] 黃林生,阮超,黃文江,等. 基于GF-1遙感影像和relief-mRMR-GASVM模型的小麥白粉病監(jiān)測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(15):167-175,314.

Huang Linsheng, Ruan Chao, Huang Wenjiang, et al. Wheat powdery mildew monitoring based on GF-1 remote sensing image and relief-mRMR-GASV M model[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(15): 167-175, 314. (in Chinese with English abstract)

[35] 劉吉超,王鋒. 基于Relief-F的半監(jiān)督特征選擇算法[J]. 鄭州大學(xué)學(xué)報(bào):理學(xué)版,2021,53(1):42-46,53.

Liu Jichao, Wang Feng. A Semi-supervised Feature Selection Algorithm Based on Relief-F[J]. Journal of Zhengzhou University: Natural Science Edition. 2021, 53(1): 42-46, 53. (in Chinese with English abstract)

[36] 王君. 基于SVM-RFE的特征選擇方法研究[D]. 大連:大連理工大學(xué),2015.

Wang Jun. The Research of Feature Selection Algorithm Based on SVM-RFE[D]. Dalian: Dalian University of Technology, 2015. (in Chinese with English abstract)

[37] 吳紅霞,吳悅,劉宗田,等. 基于Relief和SVM-RFE的組合式SNP特征選擇[J]. 計(jì)算機(jī)應(yīng)用研究,2012,29(6):2074-2077.

Wu Hongxia, Wu Yue, Liu Zongtian, et al. Combined SNP feature selection based on relief and SVM-RFE[J]. Application Research of Computers, 2012, 29(6): 2074-2077. (in Chinese with English abstract)

[38] 李紅,張凱,陳超,等. 基于高光譜成像技術(shù)的生菜冠層含水率檢測(cè)[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2021,52(2):211-217,274.

Li Hong, Zhang Kai, Chen Chao, et al. Detection of moisture content in lettuce canopy based on hyperspectral imaging technique[J]. Transactions of the Chinese Society for Agricultural Machinery, 2021, 52(2): 211-217, 274. (in Chinese with English abstract)

[39] 慕濤陽,趙偉,胡曉宇,等. 基于改進(jìn)的DeepLabV3+模型結(jié)合無人機(jī)遙感的水稻倒伏識(shí)別方法[J]. 中國(guó)農(nóng)業(yè)大學(xué)學(xué)報(bào),2022,27(2):143-154.

Mu Taoyang, Zhao Wei, Hu Xiaoyu, et al. Rice lodging recognition method based on UAV remote sensing combined with the improved DeepLabV3+ model[J]. Journal of China Agricultural University, 2022, 27(2): 143-154. (in Chinese with English abstract)

[40] 支俊俊,董婭,魯李燦,等. 基于無人機(jī)RGB影像的玉米種植信息高精度提取方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(18):48-54.

Zhi Junjun, Dong Ya, Lu Lican, et al. High-precision extraction method for maize planting information based on UAV RGB images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(18): 48-54. (in Chinese with English abstract)

Extraction of maize lodging information after typhoon based on UAV multispectral remote sensing

Zhao Jing1,2, Yan Chunyu1,2, Yang Dongjiang1,2, Wen Yuting1,2, Li Wenhua1,2, Lu Liqun2,3, Lan Yubin1,2※

(1.,,255000,; 2.,,255000,;3.,,255000,)

Bending of the lower part of the stalk (lodging) has posed a great threat to the yield, quality, and mechanical harvesting capacity in maize production. It is a high demand to quickly identify the lodging of maize subjected to the large wind load. In this study, an unmanned aerial vehicle (UAV)-based multispectral remote sensing was utilized to extract the maize lodging information after typhoon. A field test was conducted at the ecological unmanned farm of Shandong University of Technology of China. A quadrotor UAV carrying a 6-channel multispectral camera was also used to capture the image of the maize field block. A Pix4Dmapper software was selected to spline the multispectral images, and the band synthesis tool of ENVI software was used to process the six single-band gray images into one image with six bands. Firstly, ten kinds of commonly-used indices of multispectral vegetation were all selected to calculate, where 20 features of near-infrared bands were involved in the classification, due to the sensor included two near-infrared bands (840 and 940 nm). Secondly, a principal component analysis (PCA) was made to transform the original 6-band multispectral image, where the first three principal component bands with the most information were retained to extract texture features. Eight texture features were obtained in each band. The minimum noise fraction rotation (MNF) was applied to reduce the dimensionality of 48 texture features generated by the original 6-band multispectral image, further to screen the first 6 texture features with the most retention information. Finally, a low- and high-pass filtering was used to process the images, where the above 62 features were taken as the full feature set. The numbers of obtained subsets were 10, 13, and 12, respectively, using the support vector machines-recursive feature elimination (SVM-RFE), ReliefF and Least absolute shrinkage and selection operator (Lasso). Five supervised classification models were selected to train the feature subsets of the target region, including SVM, Naive Bayes, K-nearest neighbor (KNN), decision tree, and artificial neural network (ANN). The most suitable classification model for different data sets was selected to classify and evaluate the accuracy of the multi-spectrum of the validation region. The results show that ReliefF, SVM-RFE, and Lasso feature screening algorithms effectively reduce the dimension of the data while maintaining high classification accuracy. The lowest classification accuracy of ReliefF feature screening algorithm was 89.02%. The lowest classification accuracies of SVM-RFE and Lasso feature screening algorithms were both 95.38% that was closer to the lowest classification accuracy of the full-feature data set of 94.80%. There was only a 0.58% difference from the lowest accuracy of the full-feature data set, indicating a higher accuracy while a significant reduction in the number of features involved in classification. A confusion matrix verified that KNN and ANN models could effectively identify soil background, normal maize, and lodging maize, with the highest overall accuracies of 93.49% and 91.77%, respectively, where the Kappa coefficients were 0.90 and 0.88. KNN model combined with SVM-RFE feature screening method had the best classification results. Consequently, the fawer features had participated for the higher classification and recognition accuracy. The finding can provide technical support to the rapid and accurate extraction of maize lodging information after typhoon using the UAV multi-spectral remote sensing.

UAV; remote sensing; extraction; multispectral; maize; lodging information; typhoon disaster

2021-03-25

2021-11-23

山東省引進(jìn)頂尖人才“一事一議”專項(xiàng)經(jīng)費(fèi)資助項(xiàng)目(魯政辦字[2018]27號(hào))

趙靜,博士,副教授,研究方向?yàn)檗r(nóng)業(yè)遙感技術(shù)與智能檢測(cè)。Email:zbceozj@163.com

蘭玉彬,博士,教授,博士生導(dǎo)師,研究方向?yàn)榫珳?zhǔn)農(nóng)業(yè)航空。Email:ylan@sdut.edu.cn

10.11975/j.issn.1002-6819.2021.24.007

S127

A

1002-6819(2021)-24-0056-09

趙靜,閆春雨,楊東建,等. 基于無人機(jī)多光譜遙感的臺(tái)風(fēng)災(zāi)后玉米倒伏信息提取[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(24):56-64. doi:10.11975/j.issn.1002-6819.2021.24.007 http://www.tcsae.org

Zhao Jing, Yan Chunyu, Yang Dongjiang, et al. Extraction of maize lodging information after typhoon based on UAV multispectral remote sensing[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(24): 56-64. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2021.24.007 http://www.tcsae.org

猜你喜歡
植被指數(shù)反射率波段
利用鏡質(zhì)組反射率鑒定蘭炭與煤粉互混樣的方法解析
商品條碼印制質(zhì)量檢測(cè)參數(shù)
——缺陷度的算法研究
Ku波段高隔離度雙極化微帶陣列天線的設(shè)計(jì)
車燈反射腔真空鍍鋁反射率研究
最佳波段組合的典型地物信息提取
基于無人機(jī)圖像的草地植被蓋度估算方法比較
新型X波段多功能EPR譜儀的設(shè)計(jì)與性能
冬小麥SPAD值無人機(jī)可見光和多光譜植被指數(shù)結(jié)合估算
最佳波段選擇的遷西縣土地利用信息提取研究
基于地面邊緣反射率網(wǎng)格地圖的自動(dòng)駕駛車輛定位技術(shù)