劉 軻,周清波※,吳文斌,2,陳仲新,唐華俊(.農業(yè)部農業(yè)信息技術重點實驗室/中國農業(yè)科學院農業(yè)資源與農業(yè)區(qū)劃研究所,北京 0008;2.華中師范大學城市與環(huán)境科學學院,武漢 430079)
?
基于多光譜與高光譜遙感數(shù)據(jù)的冬小麥葉面積指數(shù)反演比較
劉軻1,周清波1※,吳文斌1,2,陳仲新1,唐華俊1
(1.農業(yè)部農業(yè)信息技術重點實驗室/中國農業(yè)科學院農業(yè)資源與農業(yè)區(qū)劃研究所,北京 100081;2.華中師范大學城市與環(huán)境科學學院,武漢 430079)
摘要:近年來,高光譜遙感數(shù)據(jù)廣泛應用于農作物葉面積指數(shù)(LAI)反演。與常用的多光譜遙感數(shù)據(jù)相比,高光譜數(shù)據(jù)能否提高農作物LAI反演的精度和穩(wěn)定性還存在爭議。針對這一問題,該研究利用實測冬小麥冠層高光譜反射率數(shù)據(jù),構造了不同光譜分辨率和波段組合的5種光譜數(shù)據(jù)?;贏CRM(a two-layer canopy reflectance model)模型、2套參數(shù)化方案及上述5種光譜數(shù)據(jù),對冬小麥LAI進行反演,分析光譜分辨率、高光譜數(shù)據(jù)波段選擇、模型參數(shù)不確定性3方面因素對LAI反演精度與穩(wěn)定性的影響。研究結果表明:當波段選擇適宜、模型參數(shù)不確定性較小且光譜數(shù)據(jù)分辨率較高時,LAI反演精度與穩(wěn)定性更高,提高光譜分辨率對LAI反演精度的改進作用隨光譜分辨率的升高而降低;反之,當高光譜數(shù)據(jù)波段選擇不當或者模型參數(shù)不確定性較大時,提高光譜數(shù)據(jù)的分辨率并未提高LAI反演精度。該研究解釋了“高光譜遙感數(shù)據(jù)能否提高植被參數(shù)反演精度”問題,為進一步發(fā)揮高光譜數(shù)據(jù)在農作物LAI反演中的潛力提供了科學參考。
關鍵詞:植被;遙感;光譜分析;葉面積指數(shù);高光譜;反演;波段選擇
劉軻,周清波,吳文斌,陳仲新,唐華俊. 基于多光譜與高光譜遙感數(shù)據(jù)的冬小麥葉面積指數(shù)反演比較[J]. 農業(yè)工程學報,2016,32(3):155-162.doi:10.11975/j.issn.1002-6819.2016.03.022http://www.tcsae.org
Liu Ke, Zhou Qingbo, Wu Wenbin, Chen Zhongxin, Tang Huajun. Comparison between multispectral and hyperspectral remote sensing for LAI estimation[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(3): 155-162. (in Chinese with English abstract)doi:10.11975/j.issn.1002-6819.2016.03.022http://www.tcsae.org
葉面積指數(shù)(LAI)反映了農作物生理生化過程和生產力狀況。獲取農作物LAI及其動態(tài)變化信息對農作物長勢監(jiān)測和產量估測等應用具有重要意義[1-3]。遙感技術能夠經濟、高效、無損地估測農作物LAI,成為了目前獲取大尺度農作物LAI的主要技術手段。高光譜數(shù)據(jù)尤其能夠刻畫地物光譜細節(jié),支持波形分析技術,因此,在農作物LAI遙感反演研究中得到了廣泛的應用[4-6]。較之多光譜數(shù)據(jù),高光譜數(shù)據(jù)能否提高農作物LAI反演的精度與穩(wěn)定性仍存在學術爭議。部分研究表明高光譜數(shù)據(jù)有助于提高LAI反演精度[7-10],也有研究得出與之相反的結論[11-13],例如李小文等[13]通過Li-Strahler模型參數(shù)敏感性分析認為:結構參數(shù)與波譜無關,波段數(shù)量增加并不能提供更多關于結構參數(shù)的信息??偨Y前人研究可以發(fā)現(xiàn),因為當時的客觀條件和主流LAI反演方法,這些高光譜-寬波段的對比研究存在很多不足[7],例如:早期常用的AVIRIS(airborne visible infrared imaging spectrometer)傳感器信噪比低[12];僅依賴冠層反射率模型模擬光譜而非實測光譜[11-12];基于植被指數(shù)(vegetation indices,VIs)開展LAI反演而未采用更能體現(xiàn)高光譜數(shù)據(jù)優(yōu)勢的光譜特征參量,如紅邊參量等[7,11,14]。因此,如何合理利用高光譜數(shù)據(jù)提高LAI反演精度與穩(wěn)定性仍是亟待研究的重要科學問題。
基于冠層反射率模型(物理模型)的LAI高光譜反演精度主要受“維數(shù)災難”和“病態(tài)反演”問題制約。1)高光譜數(shù)據(jù)眾多波段間的多重共線性導致“維數(shù)災難”,致使反演模型決定系數(shù)高但是預測精度差。因此,應進行數(shù)據(jù)降維,選擇盡可能正交的波段反演LAI[7,15]。2)物理模型不同的參數(shù)值組合易得到相似的光譜反射率,增加LAI反演的不確定性,導致“病態(tài)反演”問題[16],該問題與物理模型參數(shù)化的不確定性密切相關。因此,在基于高光譜和多光譜數(shù)據(jù)反演LAI的對比研究中,一方面,應基于相同波段選擇方案與物理模型參數(shù)不確定性水平,剝離除光譜分辨率以外的其他因素對反演的影響;另一方面,應驗證不同的波段選擇方案與物理模型不確定性水平對反演的影響,以便限制不利因素,充分發(fā)揮高光譜數(shù)據(jù)的優(yōu)勢。
本研究利用ACRM(a two-layer canopy reflectance model)冠層反射率模型和基于查找表(look-up table,LUT)的反演方法開展LAI反演研究?;?個時相的冬小麥冠層反射率高光譜實測數(shù)據(jù)構造了不同光譜分辨率和波段組合的5種反射率數(shù)據(jù),制訂了ACRM模型的2套參數(shù)化方案?;诖?,開展LAI反演試驗,探索了遙感數(shù)據(jù)光譜分辨率、波段選擇方案、物理模型參數(shù)不確定性3方面因素對LAI反演的影響,揭示了發(fā)揮高光譜數(shù)據(jù)優(yōu)勢的必要條件,以期為合理利用高光譜數(shù)據(jù)反演作物LAI提供科學參考。
1.1試驗區(qū)與田間測量
研究區(qū)域位于河北省衡水深州市,地處華北平原,是中國典型的冬小麥種植區(qū)。田間測量分別于2014年4 月29日和5月21日進行,此時正值冬小麥拔節(jié)期和抽穗期,在反演試驗中分別記為T1和T2,共測量了5塊冬小麥田塊,為了體現(xiàn)測量的代表性,不同田塊種植的品種不同,各試驗田塊見表1。
表1 試驗樣地Table 1 Sample plots
大致沿田塊對角線均勻布設5~6個樣本點,研究區(qū)共設28個樣本點。每個樣本點為長50 cm、寬4壟(約150 cm)的樣方,在樣方中進行測量。利用FieldSpec 4 光譜儀(美國ASD 公司生產)測量冠層光譜,探頭視場角為10°,垂直向下,距小麥冠層頂部高度約1 m,光譜測量范圍350~2 500 nm,光譜分辨率1 nm。為減少計算量,將光譜分辨率重采樣為5 nm,并采用Gao[17]提出的基于移動窗內局部均值與局部方差的方法計算各波段的信噪比,結果表明感興趣波段445~1 200 nm信噪比均大于50,可直接用于反演。采用LAI-2200冠層分析儀(美國LI-COR公司生產)測量LAI。在每個樣本點分別選取5和10株冬小麥葉片樣本各一份,使用LI-3000便攜式葉面積儀(美國LI-COR公司生產)現(xiàn)場測量各樣本的總葉面積,并秤量葉鮮質量后冷藏送至實驗室。5株冬小麥的葉片樣本用于測量葉片葉綠素含量Cab;10株冬小麥的葉片樣本烘干至恒質量,秤量葉片干質量,根據(jù)葉片干、濕質量求得冬小麥葉片干物質含量Cm、含水量Cw和比葉重(leaf specific weight,SLW)。利用各樣本的總葉面積,將Cab、SLW、Cw單位換算為g/m2,以符合ACRM模型的預定義[18]。
1.2合成模擬TM多光譜數(shù)據(jù)
為探索遙感數(shù)據(jù)光譜分辨率對冬小麥LAI反演的影響,并排除不同數(shù)據(jù)源信噪比、傳感器特性等干擾因素,本研究基于實測冬小麥冠層高光譜反射率,利用Landsat 5 TM傳感器光譜響應函數(shù)合成TM多光譜數(shù)據(jù)紅(R)、綠(G)、藍(B)、近紅外(NIR)波段[7,12,19],公式如下
式中ρm為模擬TM多光譜反射率;ρh,λ為λ波長的高光譜反射率;RSRλ為對應波長的TM光譜響應函數(shù),從美國地質調查局(United States Geological Survey,USGS)網站下載[20];Lλ為太陽直射輻射亮度,由MODTRAN (moderate resolution atmospheric transmission)模型模擬得到。MODTRAN大氣模型選擇“中緯度夏季(MODEL=2,CARD 1)”,氣溶膠模式設為“鄉(xiāng)村氣溶膠模式(IHAZE=1,CARD 2)”,其他參數(shù)取值參考相關研究[21],限于篇幅不再贅述。
本研究LAI反演步驟如下:1)通過分析ACRM模型各參數(shù)的敏感性,選取自由變量,確定模型參數(shù)化方案。2)正向運行ACRM模型,構建查找表。3)反演波段選擇。4)比較模擬實測光譜,計算二者選定波段的均方根誤差(RMSE)作為代價函數(shù),得到LAI反演結果。
2.1ACRM冠層反射率模型及其反演方法
本研究選用ACRM模型原因如下:1)ACRM模型考慮了熱點效應、植被葉片群聚效應和雙層冠層對植被二向反射的影響。針對具有顯著直立結構的作物冠層,ACRM模型利用馬爾可夫鏈原理改進了葉傾角分布函數(shù)。2)ACRM模型耦合了PROSPECT模型描述葉片組分光譜及Price[22]提出的4個基函數(shù)(rsl1~rsl4)描述土壤光譜,并以?ngstr?m濁度系數(shù)模擬天空光比,解決了輻射傳輸模型中相關輸入?yún)?shù)難以獲取的問題[23]。3)前人研究同時表明,ACRM模型模擬精度良好,是較為成熟、完善、可靠的均勻冠層反射率模型[24-27]。
冠層反射率模型反演方法主要有迭代優(yōu)化[12]、基于查找表的反演[28]以及基于人工神經網絡[29]、支持向量回歸[16]等統(tǒng)計方法的反演。本研究選擇基于查找表的反演方法,主要原因:一方面,該方法在近年來冠層反射率模型反演研究中應用非常廣泛[9,28,30-32],針對該方法的驗證和改進有廣泛的指導意義;另一方面,基于查找表的反演方法原理簡單、計算快捷、反演精度較高[31-32]。
2.2ACRM模型輸入?yún)?shù)敏感性分析
合理確定自由變量對于查找表反演精度有很明顯的影響。本研究采用EFAST(extended Fourier amplitude sensitivity test)方法[33]分析了ACRM模型11個主要參數(shù)[26,34]在各波段的敏感性。EFAST方法考慮了參數(shù)間的交互作用,因而比常用的局部敏感性分析方法更全面、客觀[34]。設待分析參數(shù)的取值在其允許范圍內均勻分布,隨機抽取5 000組參數(shù)取值組合,代入ACRM模型,生成模擬反射率數(shù)據(jù)集。將參數(shù)取值組合與模擬反射率數(shù)據(jù)集輸入SimLab軟件,現(xiàn)EFAST敏感性分析。由分析結果(圖1)可知,在可見光-近紅外(VNIR)范圍內,有8個參數(shù)全局敏感性指數(shù)大于0.1,分別是LAI、葉綠素含量Cab、SLW、平均葉傾角θm、土壤反射率參數(shù)rsl1、葉肉結構參數(shù)N、?ngstrom濁度系數(shù)β、馬爾可夫群聚參數(shù)Sz。
圖1 ACRM模型輸入?yún)?shù)敏感性分析Fig.1 Sensitivity analysis of variables in ACRM model
2.3ACRM模型參數(shù)化與查找表的構建
為研究ACRM模型參數(shù)不確定性對LAI反演的影響,本研究測試了2套參數(shù)化方案,分別記為S1、S2,見表2。β在較大空間尺度上相對均一,可固定為期望值。在缺乏先驗知識的情況下,設全局敏感性指數(shù)大于0.1的其余7個變量為自由變量,即方案S1。小麥葉肉結構參數(shù)N的取值相對固定[33],預試驗表明N=0.5適用于本研究大多數(shù)田塊;參數(shù)Sz取值范圍為0.4~1.0,本研究封壟后小麥的幾何形態(tài)為均勻冠層,Sz取值1,鑒于此,方案S2模擬了實際研究中作物種類及其時相已知的情況,即封壟后冬小麥,將N與Sz固定為0.5和1,其余自由變量的取值及變化范圍與方案S1相同。
表2 ACRM模型參數(shù)取值范圍Table 2 Ranges of input variables for ACRM model
本研究中自由變量取值范圍較寬(見表2),模擬了實際應用中缺乏先驗知識的通常情況。熱點參數(shù)SL參數(shù)化為LAI的函數(shù)[35]。葉片表面蠟質折射指數(shù)n、葉片干物質含量Cm、橢圓葉片角分布參數(shù)eL對VNIR范圍內反射率貢獻低,固定為各自期望值[26]。葉片棕色素含量Cbp僅對衰老葉片的反射率影響顯著。本研究中,綠色葉片占絕對優(yōu)勢,故Cbp設為0[36]。β表征大氣總體渾濁狀況,與波長無關。本研究依據(jù)MODIS氣溶膠產品(MOD/MYD04)獲取470、660 nm處的氣溶膠光學厚度τ,根據(jù)Iqbal[37]的算法求得β。太陽天頂角θsza根據(jù)各樣本點田間光譜觀測時刻和田塊經緯度,基于美國俄勒岡大學太陽輻射監(jiān)測實驗室提供的Solar Position Calculator計算生成。
2.4光譜分辨率和波段選擇
參照相關研究[9,28,38],本研究以VNIR波長范圍(445~1 065 nm)的冠層反射率反演LAI,嘗試了5種不同光譜分辨率和波段組合的光譜數(shù)據(jù),分別記為B1~B5,見表3。其中B1為模擬TM多光譜數(shù)據(jù);B2、B3分別為VNIR范圍內所有波段、TM敏感波段的高光譜數(shù)據(jù);B4為用于LAI反演的高光譜優(yōu)選波段數(shù)據(jù)。為進一步考察光譜分辨率對LAI高光譜反演的影響,以B4選定波段為中心波長,求其附近±10 nm范圍內各波段的平均反射率,合成光譜分辨率20 nm的多光譜數(shù)據(jù),記為B5。
表3 LAI反演各光譜數(shù)據(jù)Table 3 Spectral data applied for LAI retrieval
高光譜波段選擇的步驟如下:1)基于逐步回歸的數(shù)據(jù)降維。為避免“維數(shù)災難”,參與反演的波段應盡量正交,且充分保持原始數(shù)據(jù)中有用的信息。由圖1可知,LAI 是NIR、紅谷范圍內最敏感參數(shù);Cab是除紅谷外的可見光范圍內的最敏感參數(shù)。因此,在全部28個實測樣本點中隨機選取6個樣本點,分別以選定樣本點的LAI、Cab實測值為自變量,各波段高光譜反射率為因變量進行逐步回歸,逐步回歸結果即為高光譜數(shù)據(jù)中反映LAI或Cab變化的相對獨立的波段,見表4。從中選擇NIR、紅谷范圍內與LAI最相關,以及除紅谷外的可見光范圍內與Cab最相關的波段,實現(xiàn)數(shù)據(jù)降維[15]。2)選取光譜模擬誤差最小的波段。當模型輸入?yún)?shù)取值盡可能準確(代入?yún)?shù)實測值或最優(yōu)估計值)時,模擬光譜與實測光譜間仍然存在誤差,本研究稱之為“光譜模擬誤差”。該誤差與波段相關,包含物理模型的模擬誤差和不可避免的模型參數(shù)取值不當導致的誤差兩部分,反映了客觀因素造成的反演誤差。為了估計并排除光譜模擬誤差對LAI高光譜反演的影響,本研究隨機抽取6個樣本點,將其參數(shù)實測值或最優(yōu)估計值代入ACRM模型,得到各樣本點的最優(yōu)模擬光譜。逐樣本點、逐波段計算最優(yōu)模擬光譜與實測光譜的距離,得到該樣本點的光譜模擬誤差。然后,按波段求所選6個樣本點光譜模擬誤差的均值,得到各波段的平均光譜模擬誤差,見圖2。以此為基礎,排除逐步回歸結果中光譜模擬誤差較大的波段,按照光譜模擬誤差最小的原則修正余下的波段,最終確定LAI高光譜反演較優(yōu)波段為B4。
表4 基于逐步回歸法的LAI、葉綠素含量Cab敏感波段選擇Table 4 Bands selected for LAI and Chlorophyll content Cabby stepwise regression
圖2 光譜模擬誤差Fig.2 Errors of spectral simulation
2.5基于查找表的ACRM模型反演
本研究以代價函數(shù)χRMSE衡量查找表中模擬光譜與實測光譜的差異,見式(2)。代價函數(shù)χRMSE應用廣泛、反演精度較高[28],適用于本研究。
式中χRMSE為代價函數(shù)值;nb為參與反演的波段數(shù);Ri,measured、Ri,simulated分別為波段i實測和模擬光譜反射率。
理論上,查找表中代價函數(shù)最小的參數(shù)組合即為反演結果。實際上,病態(tài)現(xiàn)象的存在使得最優(yōu)解不唯一。因此,本研究選取所有參數(shù)組合中χRMSE最小的前20%,分別對各參數(shù)求均值,以其作為反演結果[30,32]。
2.6LAI反演精度的評價
以線性回歸決定系數(shù)(R2)、反演值均方根誤差(RMSE)與平均相對誤差(MRE)3個統(tǒng)計量評估LAI反演的精度與穩(wěn)定性。
本研究使用ACRM模型2套參數(shù)化方案,基于不同分辨率和波段組合的5種光譜數(shù)據(jù)來反演2個時相的冬小麥LAI,比較全部(28個)實測樣本點上LAI實測值與反演值。見圖3。
3.1高光譜數(shù)據(jù)波段選擇對冬小麥LAI反演的影響
從圖3可見,不同波段選擇方案(B2、B3、B4)對冬小麥LAI高光譜反演精度與穩(wěn)定性有明顯影響。1)VNIR范圍全波段高光譜數(shù)據(jù)(B2,見圖3b)并未比TM敏感波段高光譜數(shù)據(jù)(B3,見圖3c)得到更好的反演結果??梢姡瑢崪y高光譜數(shù)據(jù)更寬的光譜覆蓋無助于提高LAI反演精度。因此,本研究中數(shù)據(jù)B2僅適用于T1-S2的反演試驗。2)比較基于B3與B4(高光譜優(yōu)選波段)的反演結果(分別對比圖3c與圖3d;圖3g與圖3h;圖3k與圖3l;圖3o與圖3p)可見,基于B4反演結果的RMSE和MRE較低,R2多數(shù)較高(除反演試驗T1-S2-B3/B4外)。該結果表明合理的波段選擇提高了LAI高光譜反演的精度與穩(wěn)定性。波段選擇對LAI反演的提高作用在模型參數(shù)不確定性較大(基于參數(shù)化方案S1)時更明顯。
3.2光譜分辨率和ACRM模型輸入?yún)?shù)不確定性對LAI反演的影響
對比基于多光譜數(shù)據(jù)(B1和B5)與高光譜數(shù)據(jù)(B3 和B4)的反演結果可知:1)ACRM模型參數(shù)不確定性較?。▍?shù)化方案S2)時,基于高光譜優(yōu)選波段數(shù)據(jù)的反演結果(T1/T2-S2-B4)最優(yōu),基于光譜分辨率20 nm的模擬多光譜數(shù)據(jù)的反演結果(T1/T2-S2-B5)次之,基于模擬TM多光譜數(shù)據(jù)的反演結果(T1/T2-S2-B1)再次之。此時,提高數(shù)據(jù)光譜分辨率對反演精度的改善作用隨光譜分辨率的升高而逐漸減小。以拔節(jié)期為例,試驗T1-S2-B4(圖3d)的RMSE比T1-S2-B5(圖3e)僅減小0.037;而T1-S2-B5的RMSE較之T1-S2-B1(圖3a)減小了0.207,改進更顯著。LAI高光譜反演最優(yōu)結果T1-S2-B4與模擬TM多光譜數(shù)據(jù)反演結果T1-S2-B1相比,表征反演穩(wěn)定性的R2雖無增加,但RMSE、MRE明顯減小,且散點圖的線性回歸線更接近于1:1線,表明系統(tǒng)誤差明顯減小?;诔樗肫跀?shù)據(jù)(T2)的反演試驗也反映出大致相同的趨勢。2)ACRM模型參數(shù)不確定性較大(參數(shù)化方案S1)時,基于模擬TM多光譜數(shù)據(jù)(B1)、光譜分辨率20 nm的多光譜數(shù)據(jù)(B5)與高光譜優(yōu)選波段數(shù)據(jù)(B4)三者的反演結果差異很小。例如,反演試驗T2-S1-B4(圖3p)較之T2-S1-B1(圖3n),前者R2雖然從0.0757升至0.2162,但反映誤差水平的RMSE反而增加了0.02。
綜上所述,1)當ACRM模型參數(shù)不確定性較?。ɑ趨?shù)化方案S2)時,光譜分辨率較高且經過合理波段選擇的數(shù)據(jù)表現(xiàn)出更優(yōu)的LAI反演精度與穩(wěn)定性。此時,提高光譜分辨率對反演精度的改善作用隨光譜分辨率的升高而逐漸減小。2)模型參數(shù)不確定性較大(參數(shù)化方案S1)時,一方面,合理的波段選擇對LAI高光譜反演精度與穩(wěn)定的改進程度比模型參數(shù)不確定性較小時更顯著;另一方面,提高數(shù)據(jù)的光譜分辨率難以明顯改善LAI反演精度。
圖3 基于不同光譜數(shù)據(jù)和不同參數(shù)方案的拔節(jié)期和抽穗期LAI反演結果Fig.3 LAI retrieval results at jointing stage and heading stage based on remote sensing data with different spectral resolution and different schemes of parameterization
3.3討論
1)本研究為高光譜數(shù)據(jù)能否提高LAI反演精度的學術爭論提供了一個可能的解釋。合理的波段選擇與冠層反射率模型參數(shù)不確定性足夠小是光譜分辨率較高的數(shù)據(jù)能夠提高LAI反演精度與穩(wěn)定性的必要條件。當上述條件滿足時,應優(yōu)先選擇光譜分辨率較高的數(shù)據(jù)。除高光譜數(shù)據(jù)外,光譜分辨率相對較高且能區(qū)分LAI特征波段(如紅谷、紅邊和NIR)的多光譜數(shù)據(jù)源(如MODIS、Landsat 8 OLI和WorldView 3)也有提高LAI反演精度的潛力,應在未來進一步加以探索。
2)為充分發(fā)揮高光譜數(shù)據(jù)在LAI反演方面的優(yōu)勢和潛力,未來研究應一方面深入研究高光譜數(shù)據(jù)波段選擇方法,尋找LAI反演最優(yōu)波段組合;另一方面提高模型參數(shù)化精度,例如,探索面向對象的參數(shù)優(yōu)化方法[38]。
3)為了合理選擇LAI高光譜反演波段,并兼顧波段選擇方法的普適性,本研究從全部28個樣本點中2次獨立地抽取6個樣本點,分別用于數(shù)據(jù)降維與光譜模擬誤差評估。雖然該波段選擇方法需要少量實測數(shù)據(jù),然而其最大限度確保了波段選擇的準確性。未來研究中,應從以下2方面完善波段選擇方法。一方面,實際應用中,冠層參數(shù)真值未知。因此,應深入探索光譜模擬誤差的成因及影響因素,發(fā)展普適的、較少依賴先驗知識的光譜模擬誤差評估方法和限制對策。另一方面,采用逐步回歸進行數(shù)據(jù)降維依賴LAI、Cab的實測值。未來研究中,應進一步探索基于高光譜數(shù)據(jù)自相關分析、主成分分析等無需實測數(shù)據(jù)的數(shù)據(jù)降維方法。
本研究使用ACRM模型2套參數(shù)化方案,基于不同光譜分辨率和波段組合的5種光譜數(shù)據(jù)來反演2個時相的冬小麥LAI,系統(tǒng)比較和分析了其反演精度與穩(wěn)定性。結果表明,基于ACRM模型的冬小麥LAI反演精度受遙感數(shù)據(jù)光譜分辨率、反演波段選擇與模型參數(shù)不確定性三方面因素影響。當波段選擇恰當,輸入?yún)?shù)不確定性較小時,光譜分辨率較高的數(shù)據(jù)表現(xiàn)出更優(yōu)的LAI反演精度與穩(wěn)定性。此時,提高數(shù)據(jù)光譜分辨率對反演精度的改善作用隨光譜分辨率的升高而逐漸減小。反之,未進行合理的波段選擇,或模型輸入?yún)?shù)不確定性較大時,提高數(shù)據(jù)的光譜分辨率難以顯著提升LAI反演精度。
[參考文獻]
[1] Sakamoto T, Shibayama M, Kimura A, et al. Assessment of digital camera-derived vegetation indices in quantitative monitoring of seasonal rice growth[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2011, 66(6): 872-882.
[2] 李衛(wèi)國,王紀華,趙春江,等. 基于定量遙感反演與生長模型耦合的水稻產量估測研究[J]. 農業(yè)工程學報,2008,24(7):128-131. Li Weiguo, Wang Jihua, Zhao Chunjiang, et al. Estimating rice yield based on quantitative remote sensing inversion and growth model coupling[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2008, 24(7): 128-131. (in Chinese with English abstract)
[3] 黃健熙,馬鴻元,田麗燕,等. 基于時間序列LAI和ET同化的冬小麥遙感估產方法比較[J]. 農業(yè)工程學報,2015,31(4):197-203. Huang Jianxi, Ma Hongyuan, Tian Liyan, et al. Comparison of remote sensing yield estimation methods for winter wheat based on assimilating time-sequence LAI and ET[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(4): 197-203. (in Chinese with English abstract)
[4] Herrmann I, Pimstein A, Karnieli A, et al. LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands[J]. Remote Sensing of Environment, 2011, 115(8): 2141-2151.
[5] 夏天,吳文斌,周清波,等. 冬小麥葉面積指數(shù)高光譜遙感反演方法對比[J]. 農業(yè)工程學報,2013,29(3):139-147. Xia Tian, Wu Wenbin, Zhou Qingbo, et al. Comparison of two inversion methods for winter wheat leaf area index based on hyperspectral remote sensing[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2013, 29(3): 139-147. (in Chinese with English abstract)
[6] 林卉,梁亮,張連蓬,等. 基于支持向量機回歸算法的小麥葉面積指數(shù)高光譜遙感反演[J]. 農業(yè)工程學報,2013,29(11):139-146. Lin Hui, Liang Liang, Zhang Lianpeng, et al. Wheat leaf area index inversion with hyperspectral remote sensing based on support vector regression algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2013, 29(11): 139-146. (in Chinese with English abstract)
[7] Lee K-S, Cohen W B, Kennedy R E, et al. Hyperspectral versus multispectral data for estimating leaf area index in four different biomes[J]. Remote Sensing of Environment, 2004, 91(3/4): 508-520.
[8] Pu R, Gong P, Yu Q. Comparative analysis of EO-1 ALI and Hyperion, and Landsat ETM+ data for mapping forest crown closure and leaf area index[J]. Sensors, 2008, 8(6): 3744-3766.
[9] Verrelst J, Romijn E, Kooistra L. Mapping vegetation density in a heterogeneous river floodplain ecosystem using pointable CHRIS/PROBA data[J]. Remote Sensing, 2012, 4(12): 2866-2889.
[10] 范聞捷,閆彬彥,徐希孺. 尺度轉換規(guī)律與同步反演作物播種面積和葉面積指數(shù)[J]. 中國科學:地球科學,2010,40(12):1725-1732. Fan Wenjie, Yan Binyan, Xu Xiru. Crop area and leaf area index simultaneous retrieval based on spatial scaling transformation[J]. Science China Earth Sciences, 2010, 40(12): 1725-1732. (in Chinese with English abstract)
[11] 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, 2000, 76(2): 156-172.
[12] Jacquemoud S, Baret F, Andrieu B, et al. Extraction of vegetation biophysical parameters by inversion of the PROSPECT+ SAIL models on sugar beet canopy reflectance data: Application to TM and AVIRIS sensors[J]. Remote Sensing of Environment, 1995, 52(3): 163-172.
[13] 李小文,高峰,王錦地,等. 遙感反演中參數(shù)的不確定性與敏感性矩陣 [J]. 遙感學報,1997,1(1):5-14. Li Xiaowen, Gao Feng, Wang Jindi et al. Uncertainty and sensitivity matrix of parameters in inversion of physical BRDF model[J]. Journal of Remote Sensing, 1997, 1(1): 5-14. (in Chinese with English abstract)
[14] 梁亮,楊敏華,張連蓬,等. 基于SVR算法的小麥冠層葉綠素含量高光譜反演[J]. 農業(yè)工程學報,2012,28(20):162-171. Liang Liang, Yang Minhua, Zhang Lianpeng, et al. Chlorophyll content inversion with hyperspectral technology for wheat canopy based on support vector regression algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(20): 162-171. (in Chinese with English abstract)
[15] 宋開山,張柏,王宗明,等. 大豆葉綠素含量高光譜反演模型研究 [J]. 農業(yè)工程學報,2006,22(8):16-21.Song Kaishan, Zhang Bai, Wang Zongming, et al. Inverse model for estimating soybean chlorophyll concentration using in-situ collected canopy hyperspectral data[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2006, 22(8): 16-21. ( in Chinese w ith English abstract )
[16] Durbha S S, King R L, Younan N H. Support vector machines regression for retrieval of leaf area index from multiangle imaging spectroradiometer[J]. Remote Sensing of Environment, 2007, 107(1/2): 348-361.
[17] Gao B-C. An operational method for estimating signal to noise ratios from data acquired with imaging spectrometers[J]. Remote Sensing of Environment, 1993, 43(1): 23-33.
[18] Kuusk A. Two-layer canopy reflectance model ACRM user Guide[Z]. Tartu: Tartu Observatory, 2009.
[19] 田國良. 熱紅外遙感[M]. 北京:電子工業(yè)出版社,2006.
[20] USGS. Spectral characteristics viewer[DB/OL]. (2013-12-03) [2015-03-22]. http://landsat.usgs.gov/tools_ spectral Viewer.php.
[21] 劉勇洪,權維俊,夏祥鰲,等. 基于MODTRAN模式與衛(wèi)星資料的晴空凈太陽輻射模擬[J]. 高原氣象,2008,27(6):1410-1415. Liu Yonghong, Quan Weijun, Xia Xiangao, et al. Net surface solar radiation in the clear sky based on MODTRAN model and satellite data [J]. Plateau Meteorology, 2008, 27(6): 1410-1415. (in Chinese with English abstract)
[22] Price J C. On the information content of soil reflectance spectra[J]. Remote Sensing of Environment, 1990, 33(2): 113-121.
[23] Kuusk A. A two-layer canopy reflectance model[J]. Journal of Quantitative Spectroscopy and Radiative Transfer, 2001, 71(1): 1-9.
[24] Delalieux S, Somers B, Hereijgers S, et al. A near-infrared narrow-waveband ratio to determine Leaf Area Index in orchards[J]. Remote Sensing of Environment, 2008, 112(10): 3762-3772.
[25] Houborg R, Boegh E. Mapping leaf chlorophyll and leaf area index using inverse and forward canopy reflectance modeling and SPOT reflectance data[J]. Remote Sensing of Environment, 2008, 112(1): 186-202.
[26] 姜志偉,陳仲新,任建強. 基于ACRM輻射傳輸模型的植被葉面積指數(shù)遙感反演[J]. 中國農業(yè)資源與區(qū)劃,2011,32(1):57-63. Jiang Zhiwei, Chen Zhongxin, Ren Jianqiang. The inversion of vegetation leaf area index based on the canopy radiative transfer model ACRM[J]. Chinese Journal of Agricultural Resources and Regional Planning, 2011, 32(1): 57-63. (in Chinese with English abstract)
[27] 王東偉,孟憲智,王錦地,等. 葉面積指數(shù)遙感反演方法進展[J]. 五邑大學學報:自然科學版,2009,23(4):47-52. Wang Dongwei, Meng Xianzhi, Wang Jindi, et al. The development of leaf area index inversion from remote sensing[J]. Journal of Wuyi University: Natural Science Edition, 2009, 23(4): 47-52. (in Chinese with English abstract)
[28] Duan S B, Li Z L, Wu H, et al. Inversion of the PROSAIL model to estimate leaf area index of maize, potato, and sunflower fields from unmanned aerial vehicle hyperspectral data[J]. International Journal of Applied Earth Observation and Geoinformation, 2014, 26(2): 12-20.
[29] Duveiller G, Weiss M, Baret F, et al. Retrieving wheat green area index during the growing season from optical time series measurements based on neural network radiative transfer inversion[J]. Remote Sensing of Environment, 2011, 115(3): 887-896.
[30] Dorigo W A. Improving the robustness of cotton status characterisation by radiative transfer model inversion of multi-angular CHRIS/PROBA data[J]. Selected Topics in Applied Earth Observations and Remote Sensing, IEEE Journal of, 2012, 5(1): 18-29.
[31] Richter K, Atzberger C, Vuolo F, et al. Experimental assessment of the Sentinel-2 band setting for RTM-based LAI retrieval of sugar beet and maize[J]. Canadian Journal of Remote Sensing, 2009, 35(3): 230-247.
[32] Vohland M, Mader S, Dorigo W. Applying different inversion techniques to retrieve stand variables of summer barley with PROSPECT+SAIL[J]. International Journal of Applied Earth Observation and Geoinformation, 2010, 12(2): 71-80.
[33] Saltelli A, Tarantola S, Campolongo F. Sensitivity Analysis in Practice: A Guide to Assessing Scientific Models[M]. Chichester: John Wiley & Sons Inc, 2004.
[34] 何維,楊華. 模型參數(shù)全局敏感性分析的EFAST方法[J].遙感技術與應用,2013,28(5):836-843. He Wei, Yang Hua. EFAST method for global sensitivity analysis of remote sensing model’s parameters[J]. Remote Sensing Technologyand Application, 2013, 28(5): 836-843. (in Chinese with English abstract)
[35] Verhoef W, Bach H. Simulation of hyperspectral and directional radiance images using coupled biophysical and atmospheric radiative transfer models[J]. Remote Sensing of Environment, 2003, 87(1): 23-41.
[36] Houborg R, Anderson M, Daughtry C. Utility of an image-based canopy reflectance modeling tool for remote estimation of LAI and leaf chlorophyll content at the field scale[J]. Remote Sensing of Environment, 2009, 113(1): 259-274.
[37] Iqbal M. An Introduction to Solar Radiation[M]. Toronto: Academic Press, 1983.
[38] Atzberger C, Richter K. Spatially constrained inversion of radiative transfer models for improved LAI mapping from future Sentinel-2 imagery[J]. Remote Sensing of Environment, 2012, 120(120): 208-218.
Comparison between multispectral and hyperspectral remote sensing for LAI estimation
Liu Ke1, Zhou Qingbo1※, Wu Wenbin1,2, Chen Zhongxin1, Tang Huajun1
(1. Key Laboratory of Agri-informatics, Ministry of Agriculture/Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China;2. College of Urban & Environmental Sciences, Central China Normal University, Wuhan 430079, China)
Abstract:Hyperspectral remote sensing has been commonly employed for crop LAI estimation in recent years. However, the advantages of hyperspectral data compared with multispectral data in LAI estimation remain debate. To compare multispectral and hyperspectral remote sensing for LAI estimation, five datasets with different spectral resolution, spectral coverage, and band selection were tested for retrieving LAI by inverting the ACRM (A Two-Layer Canopy Reflectance Model) model in this study. The study area is located in Shenzhou, Hebei Province, China. A field experiment was conducted during the jointing and heading stages of winter wheat (Triticum aestivum) in 2014. In situ measurements were performed in five winter wheat cultivars. The canopy spectra and the biophysical variables (LAI, leaf chlorophyll content, and leaf specific weight etc.) were measured. The inversion technique based on a look-up table (LUT) is adopted with the following procedure. Firstly, for determining the free variables of the LUT, sensitivities of the ACRM variables were evaluated using the EFAST algorithm. Two schemes of parameterizations were designed, separately denoted as “S1” and “S2”. The scheme S1 had 7 variables, whose EFAST global sensitivity index was larger than 0.1, as free variables. The scheme S2 further was used to fixe leaf mesophyll structure and Markov clumping parameter to their best estimation. Secondly, to select the optimum hyperspectral bands for LAI estimation, stepwise regression was adopted to eliminate the multicollinearity in hyperspectral data. The results of stepwise regression were further adjusted to avoid errors in spectral simulation. Thirdly, five datasets, separately denoted as B1 to B5, were composed based on the in situ measured hyperspectral spectra and the result of band selection, including B1: the synthetic Landsat 5 TM data; B2: hyperspectral data (5 nm spectral resolution) of visible light and near inferred (VNIR, 445-1 065 nm); B3: hyperspectral data covering the sensitive bands of TM within VNIR (445-945 nm); B4: the selected hyperspectral bands for LAI estimation; B5: multispectral data of 20 nm spectral resolution, with their center wavelengths located at the selected hyperspectral bands. The accuracy and stability between LAI retrieval based on the two schemes of ACRM parameterization and using the five datasets were compared. The experiments showed that: first, within the range of VNIR, LAI estimation did not benefit from the wider spectral coverage of in situ measured hyperspectral data than the synthetic TM data.Second, if the bands participating in the inversion were properly selected and the uncertainty in the parameterization of the ACRM model was fairly low, remote sensing data of higher spectral resolution would generally result in a more accurate LAI estimation. In this case, the effects of spectral resolution to the inversion accuracy were not linear. With the increase of spectral resolution, the benefit from higher spectral resolution could decrease. For instance, B5 yielded significantly more accurate LAI estimations than B1; however, B4 performed merely slightly better than B5. Third, if the bands for retrieving LAI were not properly selected (for instance, using dataset B3), or the parameterization of ACRM model was fairly uncertain (for instance, using the scheme S1), remotely sensed data with higher spectral resolution could not result in more accurate LAI estimation. In conclusion, remotely sensed data with higher spectral resolution generally yielded more accurate LAI estimation only when the band selection was properly performed and the uncertainty of the parameters was fairly low.Otherwise, there was no significant difference between multispectral and hyperspectral data for crop LAI retrieval. This study provides information for the advantages of using hyperspectral data to estimate LAI. Moreover, this study reveals the great potential to enhance the accuracy of LAI estimation by using multispectral data with relevantly high spectral resolution, for instance, MODIS, Landsat 8 OLI and WorldView 3.
Keywords:vegetation; remote sensing; spectrum analysis; leaf area index; hyperspectral data; inversion; band selection
通信作者:※周清波,男,湖南沅江人,博士,研究員,博士生導師。主要從事農情遙感領域的基礎研究和應用基礎研究。北京中國農業(yè)科學院農業(yè)資源與農業(yè)區(qū)劃研究所,100081;Email:zhouqingbo@caas.cn
作者簡介:劉軻,男,四川攀枝花人,博士生,主要從事作物參數(shù)遙感反演方法研究。北京中國農業(yè)科學院農業(yè)資源與農業(yè)區(qū)劃研究所,100081。Email:billc_st@163.com
基金項目:測繪地理信息公益性行業(yè)科研專項(201512028);國家自然科學基金項目(41271112)
收稿日期:2015-04-10
修訂日期:2015-12-16
中圖分類號:S126;TP79
文獻標志碼:A
文章編號:1002-6819(2016)-03-0155-06
doi:10.11975/j.issn.1002-6819.2016.03.022