leaf nitrogen content indirectly estimated by leaf traits derived from the prospect model
TRANSCRIPT
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1
Leaf Nitrogen Content Indirectly Estimated by LeafTraits Derived From the PROSPECT Model
Zhihui Wang, Andrew K. Skidmore, Roshanak Darvishzadeh, Uta Heiden, Marco Heurich, and Tiejun Wang
Abstract—Leaf nitrogen content has so far been quantifiedthrough empirical techniques using hyperspectral remote sensing.However, it remains a challenge to estimate the nitrogen contentin fresh leaves through inversion of physically based models. Leafnitrogen has been found to correlate with leaf traits (e.g., leafchlorophyll, dry matter, and water) well through links to the pho-tosynthetic process, which provides potential to estimate nitrogenindirectly. We therefore set out to estimate leaf nitrogen con-tent by using its links to leaf traits that could be retrieved froma physically based model (PROSPECT) inversion. Leaf optical(directional-hemispherical reflectance and transmittance between350 and 2500 nm) and leaf biochemical (nitrogen, chlorophyll, drymatter, and water) properties were measured. Correlation anal-ysis showed that the area-based nitrogen correlations with leaftraits were higher than mass-based correlations. Hence, simpleand multiple linear regression models were established for area-based nitrogen using three leaf traits (leaf chlorophyll content,leaf mass per area, and equivalent water thickness). In addi-tion, the traits were retrieved by the inversion of PROSPECTusing an iterative optimization algorithm. The established empir-ical models and the leaf traits retrieved from PROSPECT wereused to estimate leaf nitrogen content. A simple linear regressionmodel using only retrieved equivalent water thickness as a pre-dictor produced the most accurate estimation of nitrogen (R2 =0.58, normalized RMSE = 0.11). The combination of empir-ical and physically based models provides a moderately accurateestimation of leaf nitrogen content, which can be transferred toother datasets in a robust and upscalable manner.
Index Terms—Hyperspectral remote sensing, leaf nitrogen, leaftraits, PROSPECT model.
I. INTRODUCTION
L EAF nitrogen (N) is a primary regulator of physiologicalprocesses, such as photosynthesis, leaf respiration, and
transpiration [1]–[5], and it is related to canopy and stand-level
Manuscript received September 13, 2014; revised February 09, 2015;accepted April 01, 2015. The work of Z. Wang was supported in part bythe China Scholarship Council (CSC) under Grant 201204910232 and in partby ITC Research Fund under Grant 93003032, Enschede, The Netherlands.(Corresponding authors: Zhihui Wang and Tiejun Wang.)
Z. Wang is with the Faculty of Geo-Information Science and EarthObservation (ITC), University of Twente, 7500 AE Enschede, The Netherlands,and also with the School of Mathematical and Geospatial Sciences, RMITUniversity, Melbourne, Vic., 3001, Australia (e-mail: [email protected]).
A. K. Skidmore, R. Darvishzadeh, and T. Wang are with the Faculty ofGeo-Information Science and Earth Observation (ITC), University of Twente,7500 AE Enschede, The Netherlands (e-mail: [email protected];[email protected]; [email protected]).
U. Heiden is with the Department of Land Surface, Earth Observation Center,German Aerospace Center, Oberpfaffenhofen, 82234 Wessling, Germany(e-mail: [email protected]).
M. Heurich is with the Bavarian Forest National Park, 94481 Grafenau,Germany (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSTARS.2015.2422734
traits, such as light use efficiency, wood growth, and net primaryproduction [6]–[8]. Nitrogen is also a critical factor for plantgrowth, and it plays an important role in terrestrial ecosystemcarbon dynamics, which acts as potential climate feedback [9]–[12]. Nitrogen is also an important input variable of ecosystemprocess models [8], [13]–[15].
Although leaf nitrogen has been quantified for dry and freshleaves, from leaf to canopy level, using imaging spectroscopy[6], [16], [17], empirical models are still the dominant methodused to estimate nitrogen, such as the spectral indices model[18]–[23], stepwise multiple linear regression [24]–[26], partialleast squares regression [6], [27], [28], support vector regres-sion [29], and artificial neural network [30]–[32]. The keychallenge for empirical models relates to their transferabilityto different areas, since they are built on a site-specific basis.Besides, when hyperspectral data are used, the selected wave-bands from different studies are inconsistent, and often deviatefrom nitrogen absorption bands [26], [30], [33].
Physically based models offer advantages compared toempirical models in robustness and transferability [34]–[36],and have been widely applied to retrieve leaf biochemicalparameters (e.g., leaf chlorophyll, dry matter, and water con-tent) from remotely sensed data [22], [37]–[44]. However,limited research focused on the utility of physically based mod-els for estimating leaf nitrogen content [33], [34], [45], [46].Several studies attempted to incorporate leaf nitrogen into theabsorption and scattering processes in the PROSPECT leafoptical properties model [34] and found that nitrogen contentfor dry leaves could be moderately well estimated throughPROSPECT model inversion [33], [34], [47]. But so far, leafnitrogen content has not been successfully estimated for freshleaves using physically based models. The reason is that forfresh leaves, leaf reflectance and transmittance are insensitiveto protein because of the small percentage of nitrogen in theleaf mass [33], [45], [48]. The strong covariance with othercompounds, such as chlorophyll and water, also led to incon-sistencies in retrieving nitrogen via the physically based modelinversion [33], [34], [45], [46]. The idea of incorporating leafnitrogen into the PROSPECT model was therefore abandonedin the 1990s; in other words, the inversion of PROSPECT forretrieving leaf nitrogen content in fresh leaves remains as achallenge and unresolved problem.
However, the model parameters (leaf chlorophyll content,CHLarea, g/cm2; leaf mass per area, LMA, g/cm2; and equiv-alent water thickness, EWT, g/cm2), which are known ascommon leaf traits, can be retrieved via PROSPECT modelinversion with intermediate to good accuracy [22], [37]–[44]and have been identified as good indicators of nitrogen [4],
1939-1404 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
[5], [49]. For example, the moderately strong correlation (aver-age Pearson correlation coefficient r = 0.65± 0.15) betweennitrogen and chlorophyll within and across ecosystems [50] hasprovided a foundation for estimating nitrogen through chloro-phyll [4], [5], [50]–[52]. However, using chlorophyll alonehas been insufficient for estimating nitrogen, because theircorrelation is less strong in nitrogen-rich ecosystems [53]. Acorrelation between nitrogen and LMA has also been foundacross species in plant science studies, due to their close linkagewith the photosynthetic process [49], [54]–[57].
Leaf nitrogen content and leaf traits (i.e., chlorophyll, drymatter, and water content) can be expressed on a mass or areabasis [5], [58], [59]. The mass-based parameters are importantfor leaf economy, as leaf mass is considered as an investmentof biomass for carbon fixation [58]. The area-based param-eters are important from a physiological perspective becausephysicochemical processes relating to photosynthesis and car-bon acquisition, such as light interception, CO2 diffusion,and transpiration, occur as a flux per unit leaf surface area[58], [59]. LMA, or the inverse ratio of LMA—specific leafarea—(SLA), links the mass- and area-based expressions ofleaf traits (Let LTmass and LTarea denote mass- and area-based leaf traits, respectively, then LTmass = LTarea/LMA, orLTmass = LTarea ∗ SLA.) [5], [59]. The strength of correla-tions between leaf nitrogen content and these leaf traits varieswhen they are expressed on a mass or area basis [50], [56], [57],[60]. Accordingly, the correlation between both mass-based andarea-based leaf nitrogen and leaf traits needs to be exploredbefore applying the relationship to estimating nitrogen content.
The correlations between leaf nitrogen content and leaf traits(i.e., chlorophyll, dry matter, and water content) offer potentialways to estimate nitrogen indirectly from spectral measure-ments. To our knowledge, there are only a few spectroscopicstudies focusing on using the nitrogen links to leaf traits (i.e.,LMA and EWT), although its correlation with chlorophyll hasbeen widely considered [50], [51]. Both correlations betweenmass-based versus area-based leaf nitrogen content and leaftraits need to be further understood, if they are to be extrapo-lated to regional or global scales. Moreover, the combinationof leaf traits may improve how we can explain the variance ofnitrogen. This paper aims to explore the relationship betweenmass-based versus area-based leaf nitrogen content and leaftraits and to apply it to retrieve leaf nitrogen content from freshleaf spectra combined with physically based models.
II. MATERIALS AND METHODS
A. Study Area and Field Data
The study area is located in the southern part of the BavarianForest National Park (49◦3′19′′N, 13◦12′9′′E), Germany(Fig. 1). The park has a total area of 24 218 ha. The bedrockof the region is primarily composed of gneiss and granite. Soilsweathered from these parent materials are naturally acid andlow in nutrients. The main soil types are brown soils, loosebrown soils, and podzol brown soils. Elevation ranges from600 to 1453 m. The climate is temperate with a total annualprecipitation between 1200 and 1800 mm and a mean annualtemperature of 5.1◦C in the valleys, 5.8◦C on hillsides, and
Fig. 1. Location of the study area in the Bavarian Forest National Park (BFNP),Germany. The satellite image was from the world basemap in ArcGIS software(ESRI, Inc., USA).
3.8◦C in the higher montane zones [61]. Dominant speciesof the forests are Norway spruce (Picea abies) (67%) andEuropean beech (Fagus sylvatica) (24.5%), with some white fir(Abies abies) (2.6%), sycamore maples (Acer psudoplatanus)(1.2%), and mountain ash (Sorbus aucuparia) (3.1%) [61].Since the mid-1990s, the forests of the National Park have beenaffected by massive proliferation of the spruce bark beetle (Ipstypographus). By 2012, this had resulted in the death of maturespruce stands over an area amounting to 6000 ha [62].
Fieldwork was conducted from mid-July to mid-August,2013, using a stratified random sampling strategy. Land usedata were obtained from the Department of Conservation andResearch, Bavarian Forest National Park. Based on plant func-tional types, the study region was stratified into broadleafdeciduous, evergreen coniferous, and mixed areas. Twenty-oneplots were randomly selected from the broadleaf deciduousareas and mixed areas. Each plot was 30m × 30m in size,and a Leica GPS 1200 (Leica Geosystems AG, Heerbrugg,Switzerland) was used to record the center location of eachplot (with an accuracy of approximately 1 m). Within each plot,depending on the homogeneity, one to three trees were selectedfor collecting leaf samples, resulting in overall 53 samplesincluding 44 European beech, 4 sycamore maple, 3 mountainash, 1 goat willow (Salix caprea), and 1 broad-leaved lime (Tiliaplatyphyllos). Each sample was composed of at least 20 leavestaken from the branches of an individual tree. The branches ofsunlit leaves were shot down from the top of each selected treeusing a crossbow. Leaves were immediately measured usinga portable SPAD-502 Leaf Chlorophyll Meter (Minolta, Inc.,Japan), and the averaged SPAD readings (M, unitless) for leavesin each sample were recorded.
Leaf samples were stored in zip-lock plastic bags with wetpaper towels and placed in a cooler with ice before transporta-tion to the laboratory for further measurement.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
WANG et al.: LEAF NITROGEN CONTENT ESTIMATED BY LEAF TRAITS 3
TABLE ITHE SELECTED LEAF TRAITS IN THIS STUDY
Wf , Wd, and A are the leaf fresh weight (in g), dry weight (in g), and leaf area (in cm2), respectively. CHLarea is the area-based leaf chlorophyll content (in μg/cm2). ρw is the physical constant representing the density of pure water (1 g cm−3).LMA and EWT are also known as area-based leaf dry matter content and leaf water content (LWC, 1 cm = 1 g/cm2),respectively, in PROSPECT [34], [85].
B. Lab Chemical Analysis
The averaged SPAD value for each sample was con-verted to area-based leaf chlorophyll content (CHLarea inμg/cm2) using an empirically calibrated equation (chlorophyll(μmol/m2) = 10∧(M∧0.265)) provided by [63]. Though theMarkwell equation was built for soybean and maize leaves, ithas been tested and used for other plant species in a number ofremote sensing studies [43], [63]–[65]. The fresh weight (Wf
in g) of leaves was measured using a digital scale, and theirleaf area (A in cm2) was measured with a LI-3000C PortableArea Meter (LI-COR, Inc., Lincoln, NE, USA). Then the spec-tral measurements were taken, and after that all the sampleswere oven-dried at 65◦C for 48 h, and their dry weights (Wd ing) were measured. Leaf fresh weight, dry weight, and leaf areawere used for deriving the water content, see details in Table I.Dried samples were stored in paper bags in a cool, dark placebefore chemical analysis. They were then ground with a mor-tar and pestle to pass through a 250-μm mesh screen. Theleaf nitrogen concentration (Nmass, % dry weight) was deter-mined using an AQ1 Discrete Analyzer (SEAL Analytical, Inc.,Mequon, WI, USA) following a modified Kjeldahl procedure,after decomposing samples with a mixture of sulfuric acid,selenium, and salicylic acid [66].
C. Spectral Measurements
Leaf directional hemispherical reflectance and transmittancefrom 350 to 2500 nm were measured on 10 leaves in each sam-ple, using an ASD FieldSpec-4 Pro FR spectrometer and anASD RTS-3ZC Integrating Sphere designed for the spectrom-eter (Analytical Spectral Devices, Inc., Boulder, CO, USA).The ASD provided measured spectra with an interval of 1 nm.Two hundred scans per leaf were averaged to a single spectrumto minimize noise. Raw radiance was converted to reflectanceusing a calibrated reference standard (with approximately 99%reflectance).
The spectra were first calibrated for dark current and straylight, according to the Integrating Sphere User Manual [67].Bands before 400 nm were removed due to noise. The spectra
from ten leaves were averaged to represent each sample.They were then smoothed to minimize noise using a movingSavitzky–Golay filter [68], [69] with a frame size of 13 datapoints (second-order polynomial). The parameters of the filterwere determined to minimize noise and to maintain spectralfeatures based on visual inspection. Only 47 of the 53 sam-ples were considered for further analysis, because six samplesbecame desiccated due to improper storage before the spectralmeasurement.
D. Retrieval of Leaf Traits
The measurements of leaf area (A), chlorophyll content(CHLarea), nitrogen content (Nmass), fresh weight (Wf ), anddry weight (Wd) were used to derive several common area-based and mass-based leaf traits (Table I).
E. Statistical Analysis
First, the Pearson correlation coefficient (r) was calculated inorder to investigate the association between leaf nitrogen (bothmass-based and area-based) and the selected leaf traits overall samples (see Table I). Species-specific correlation coeffi-cients were not calculated due to the limited number of samplesfor the species other than European beech. The mass-basedor area-based leaf nitrogen having in average higher correla-tion coefficients (r ≥ 0.6) with leaf traits was selected for laterestimation. Only those leaf traits having significant correla-tions (p < 0.01) with the selected leaf nitrogen were chosenfor further regression analysis. Then, both simple and mul-tiple linear regression models were developed to model leafnitrogen content as a function of leaf traits. Unlike simplelinear regression models, multiple regression models includetwo or three leaf traits as independent variables in order toexplore if extra variables can improve the accuracy in estimat-ing leaf nitrogen content. Nonlinear regression was also testedbut did not improve the estimation accuracy over linear regres-sion models, thus linear regression models were adopted forsimplicity.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
TABLE IISPECIFIC RANGES FOR PARAMETERS IN THE PROSPECT MODEL
aThe ranges were determined based on the prior knowledge from field measurements.
TABLE IIITHE STATISTICS OF LEAF TRAITS (SAMPLE SIZE = 47)
See Table I for the details of leaf traits.
Statistical tests were applied to assess the validity of theregression models, as described in [70]. The model residu-als (observed minus estimated value) were tested if they areindependent, homoscedastic, and normally distributed. F-testwas performed to test the significance of the linear regres-sion model, and t-test was performed to test the significanceof individual regression coefficients. Besides, the varianceinflation factor (VIF) was assessed for correlated predictorvariables in the multiple regression models. Independent vari-ables with VIF under 10 were considered as indication of lowmulticollinearity [71].
All regression models were validated using a “leave-one-out”cross-validation, with the final model developed using all 47samples. For each dependent variable, a model is built usingindependent variables from 46 samples, which is used to esti-mate the value of the left-out sample. The procedure is repeatedfor 47 times in order to obtain estimates for all samples. Thecross-validated coefficient of determination (R2
CV), root-mean-square error (RMSECV), and normalized RMSE (NRMSECV)between predicted and measured values were calculated toevaluate the performance of the regression models [72]. Thecross-validated RMSE is a good indicator of the accuracy of themodel in predicting unknown samples, because the predictedsamples are different from the samples used to build the model[72]. For multiple regression models, standardized regressioncoefficients (beta coefficients) were calculated to compare therelative contribution of each independent variable to estimateleaf nitrogen content [73], [74]. Statistical analysis was con-ducted using IBM SPSS Statistics 20 (IBM, Inc.) and MATLAB(The MathWorks, Inc.).
F. Estimation of CHLarea, LMA, and EWT Using thePROSPECT Model
The PROSPECT leaf optical properties model was devel-oped to simulate leaf directional-hemispherical reflectance and
transmittance over the optical domain from 400 to 2500 nm[34]. It only needs four input parameters: leaf structure index(Nstruc), leaf chlorophyll content (CHLarea, μg/cm2), leaf drymatter content (known as leaf mass per area, LMA, g/cm2), andleaf water content (known as equivalent water thickness, EWT,cm, in Table I). The improved (1-nm resolution) and recali-brated version, PROSPECT-4, was chosen in this study [41].
Using an iterative optimization inversion algorithm, thefour input parameters of PROSPECT were estimated usingmeasured reflectance and transmittance spectra. The inver-sion was performed using a bounded optimization package,FMINSEARCHBND.M in MATLAB. (The MATLAB code isavailable in [75]). The inversion process was to find the parame-ter vector θ = [Nstruc, CHLarea, LMA, EWT]T, which min-imizes the merit function
J(θ)=∑λ2
λ1
(Rmes(λ)−Rmod(λ, θ))2+(Tmes(λ)−Tmod(λ, θ))
2
(1)
where λ is the wavelength, Rmes and Tmes are, respectively, themeasured reflectance and transmittance, and Rmod and Tmod
are the modeled values.The range of the input parameters, CHLarea, LMA, and EWT
(Table II), specified in the bounded optimization function wasdetermined based on the prior field information (Table III). Therange for the leaf structural index Nstruc was chosen basedon an earlier study by [34], which reported that dicotyledonshave Nstruc values between 1.5 and 2.5. The model inver-sions included three steps, and in each step, an optimizationwas performed for estimating different parameters over a dif-ferent wavelength range. For each range, the parameters thatwere estimated have the greatest influence on the reflectanceand transmittance. Details were as follows: 1) leaf structureindex was determined over 760–1300 nm following the threewavelengths method described in [33] and [41]; the accuracy ofestimation for the other three parameters can be improved if the
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
WANG et al.: LEAF NITROGEN CONTENT ESTIMATED BY LEAF TRAITS 5
leaf structure index is computed first and used as a known value[76]. 2) With the leaf structure index fixed to the value retrievedin the first step, along with prior information for ranges ofLMA and EWT, the leaf chlorophyll content was estimatedusing 450–690 nm by inversion of PROSPECT [77]. 3) Withthe estimated leaf structure index and estimated leaf chloro-phyll content, LMA and EWT were concurrently estimated overthe 900–2500 nm range, where water and dry matter are mostabsorbent [37], [41], [47].
The accuracies of the retrieved parameters (CHLarea, LMA,and EWT) via model inversion were evaluated by the coeffi-cient of determination (R2), the RMSE, and the normalizedRMSE (NRMSE = RMSE/range) between the estimated andmeasured values.
G. Estimation of Leaf Nitrogen Content
The regression models passing all statistical tests were cou-pled with their corresponding retrieved leaf traits from thePROSPECT model inversion (Section II-F) to estimate mass-based or area-based leaf nitrogen content (see Section II-E).The coefficient of determination (R2), the RMSE, and NRMSEbetween the measured leaf nitrogen content and predictedvalues were used to evaluate the performance of each model.
III. RESULTS
A. Characteristics of Leaf Properties
Table III summarizes the statistical characteristics of leafproperties in our dataset. It shows the area-based leaf chloro-phyll content ranged from 19.47 to 58.91 μg/cm2, whereas themass-based leaf chlorophyll content varied from 2.00 to 14.09μg/g. The area-based leaf nitrogen content spanned an eight-fold range of values, whereas the variation of mass-based leafnitrogen content was limited (1.63–3.97%). LMA had a five-fold range of values. EWT ranged from 0.0043 to 0.0156 cm.While area-based plant traits had a wide range of values, themass-based plant traits, GWCf , and LDMC showed limitedvariation.
B. Correlations of Leaf Traits and Linear Regression Models
Table IV illustrates the relationships between leaf nitro-gen and selected leaf traits for pooled samples (details inTable I). Species-specific correlation coefficients calculated forthe European beech were similar but slightly lower (results notshown). To involve the variations across species, the resultsof the pooled samples were utilized. For Nmass, the high-est positive relationship were with GWCf (r = 0.663, p <0.01), followed by area-based leaf chlorophyll content (r =0.637, p < 0.01) and mass-based chlorophyll content (r =0.555, p < 0.01). A weaker correlation was observed betweenNmass and LMA (r = −0.346, p < 0.05). In terms of Narea,we found the highest correlation with EWT (r = 0.841, p <0.01), followed by LMA (r = 0.686, p < 0.01). Moderatecorrelation was observed between Narea and area-based leafchlorophyll content (r = 0.597, p < 0.01), and no signifi-cant correlation was found between Narea and mass-based
TABLE IVCORRELATIONS BETWEEN LEAF NITROGEN CONTENT
AND OTHER LEAF TRAITS
∗∗Correlations significant at p < 0.01, and ∗correlations significant at p <0.05. See Table I for the definitions of leaf traits.
leaf chlorophyll content. Since Narea had a higher correla-tion with leaf traits than Nmass, we only considered area-basednitrogen and its highly correlated leaf traits for later anal-ysis. Furthermore, the parameters that are incorporated inPROSPECT are also area-based, which enables the use ofwell-fitting realtionships on the leaf traits derived from themodel.
All regression models passed the statistical tests except threemultiple regression models in the t-test (Table V). The regres-sion coefficient of LMA in (4), that of CHLarea in (6), and thatof EWT in (7) (Table V) were not significant. Multicollinearityamong independent variables for multiple regression models(VIF < 10) were not observed (Table V).
The equations of simple and multiple linear regressionmodels are listed in Table VI. The performance of eachmodel in predicting the leaf nitrogen content (Narea) is illus-trated in Table VII. Among the simple linear regressionmodels, the highest R2 values (R2
CV = 0.660) and lowestRMSE (RMSECV = 3.29E-05 g/cm2) occurred when usingEWT as the independent vairable. In comparison, low R2
were obtained when regressing with LMA (R2CV = 0.376) or
CHLarea (R2CV = 0.243).
The multiple linear regression model using LMA andCHLarea as predicting variables greatly improved the accuracyof estimation (R2
CV = 0.695) compared to simple regressionmodels using only one of the two variables. The two variableshad almost equal effects on estimating Narea, as seen from thestandardized regression coefficients (Table VII, coefficients =0.631 and 0.532 for LMA and CHLarea, respectively). Nocollinearity was found between LMA and CHLarea.
Other three multiple linear regression models [(4), (6), and(7) in Table VI] with extra variables also provided accurate pre-diction of Narea (R2
CV = 0.648–0.710, Table VII). In terms ofthe model using EWT and LMA as predictors, EWT played thedominant role in explaining the variance of Narea (Table VII,standardized regression coefficients = 0.710 and 0.186 forEWT and LMA, respectively). This can be attributed to thecorrelation between EWT and LMA (r = 0.704, p < 0.01),which is also the reason that the regression coefficient of LMAin the model was not significant (Table V). The same phe-nomenon was observed in the model developed using EWTand CHLarea. The multiple regression model incorporating allthree leaf traits—LMA, EWT, and CHLarea—gave the high-est R2 values (R2
CV = 0.710) and lowest RMSE (RMSECV =3.04E-05 g/cm2). All three independent vairables had a sim-ilar effect on predicting Narea, according to their respectivecoefficients of 0.422, 0.324, and 0.350. These three multipleregression models [(4), (6), and (7) in Table VI] were not
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
TABLE VSTATISTICAL TESTS FOR VALIDITY OF REGRESSION MODELS FOR ESTIMATING LEAF NITROGEN
CONTENT (Narea) USING DIFFERENT COMBINATIONS OF INDEPENDENT VARIABLES
TABLE VIREGRESSION EQUATIONS BETWEEN LEAF NITROGEN CONTENT (Narea) AND DIFFERENT
COMBINATIONS OF INDEPENDENT VARIABLES
TABLE VIIPERFORMANCE OF REGRESSION MODELS FOR ESTIMATING LEAF NITROGEN CONTENT (Narea)
USING DIFFERENT COMBINATIONS OF INDEPENDENT VARIABLES
involved in nitrogen estimation for later analysis, due to theirnonsignificant coefficients.
C. PROSPECT Model Inversion for CHLarea, EWT, and LMA
Leaf traits serving as independent variables in regressionmodels were predicted from the PROSPECT model inversion.The inversion provided accurate estimates of CHLarea (R2 =0.54, RMSE = 7.72 g/cm2), EWT (R2 = 0.66, RMSE =0.0014 cm), and LMA (R2 = 0.64, RMSE = 0.0022 g/cm2)(Fig. 2). The inversion accuracy obtained here is compara-ble to previous studies [37], [39], [41], [78]. Based on thevalues of normalized RMSE (Fig. 2), it can be noted that EWT
(NRMSE = 0.13) was estimated with a higher accuracy thanLMA (NRMSE = 0.22) or CHLarea (NRMSE = 0.20).
D. Accuracy of Estimated Leaf Nitrogen Content
Narea was estimated with a combination of empirical regres-sion models and the retrieved leaf traits through the physicallybased model inversion using PROSPECT. Three simple lin-ear regression models [(1)–(3) in Table VI] and one multipleregression model [ (5) in Table VI] were used in the estimation.Of the four models, the most accurate estimation was achievedby the linear regression model using retrieved EWT, yieldingan R2 of 0.58, and an RMSE of 4.26E05 g/cm2 [Fig. 3(c) and
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
WANG et al.: LEAF NITROGEN CONTENT ESTIMATED BY LEAF TRAITS 7
Fig. 2. Measured versus estimated leaf traits obtained from PROSPECT model inversion: (a) CHLarea, leaf chlorophyll content; (b) EWT, equivalent waterthickness; and (c) LMA, leaf mass per area.
Fig. 3. Comparison between measured and estimated Narea (g/cm2) usingdifferent regression models, coupled with their independent variables retrievedfrom the PROSPECT model inversion. The independent variables of eachregression model were (a) CHLarea; (b) LMA; (c) EWT; (d) LMA; andCHLarea.
Table VIII]. The remaining three regression models offered lessaccurate estimations of leaf nitrogen, with the linear regres-sion model using LMA providing the least accurate estimation(R2 = 0.21, RMSE = 6.64E-05 g/cm2, Table VIII).
IV. DISCUSSION
This study confirms the feasibility of estimating leaf nitrogencontent by combining empirical and physically based mod-els; it has previously only been demonstrated using empirical
TABLE VIIIVALIDATION OF ESTIMATED Narea FROM THE COMBINATION OF
REGRESSION MODELS AND PREDICTED CHLarea , EWT, LMA FROM
THE PROSPECT INVERSION
methods [16], [17], [28], [46] and had been rejected as a suit-able approach for using physical model inversion with freshleaves [33], [37], [47]. Leaf nitrogen content (Narea) was mod-erately well estimated indirectly through the PROSPECT modelinversion using correlated leaf traits as the main driver.
Higher correlations were found between area-based leafnitrogen content and leaf traits (CHLarea, LMA, and EWT)than when nitrogen and leaf traits were expressed on a massbasis. This phenomenon can be explained by the role of nitro-gen and leaf traits in photosynthesis: most processes such aslight interception and carbon acquisition are expressed on aleaf surface area basis [58], [59]. The indirect estimation of leafnitrogen content through the PROSPECT model inversion relieson CHLarea, EWT, and LMA, which are expressed on an areabasis.
The role of leaf traits (LMA and EWT) in estimating leafnitrogen content as a medium has been largely ignored in the lit-erature, although chlorophyll is often considered as a proxy ofnitrogen [50], [51]. In this study, a higher correlation was foundbetween leaf nitrogen content and LMA/EWT than chlorophyll.The top-of-canopy sunlit leaves, exposed to illumination, areabove the saturation level for photosynthesis [79] and the frac-tion of leaf nitrogen allocated to chlorophyll becomes constant[80], while more nitrogen is invested in additional carbon-fixingcompounds [81], [82]. In addition, only 19% of leaf nitrogen inC3 plants is allocated to light-harvesting complexes [83], andonly 1.7% is directly in chlorophyll, whereas around 70% ofnitrogen is in molecules that are related to carbon fixation [84].LMA is a direct measure of leaf dry-mass investment per unit
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
Fig. 4. Indirect relationship between equivalent water thickness (EWT) andleaf nitrogen content (Narea).
of light-intercepting leaf area deployed, and it gathers carbon-related compounds, such as cellulose, lignin, hemicelluloses,and protein [33], [47]. The strong relationship between Narea
and LMA in our study confirms these facts and is also consistentwith earlier findings [49], [54]–[57], [60].
However, it is difficult to explain the correlation betweenNarea and EWT from a physiological perspective, becauseleaf water content is highly dynamic in the temporal dimen-sion compared with variation in Narea and LMA [85]. First,from statistical analysis, the high correlation between LMA andEWT (r = 0.704, p < 0.01) supports the notion of a relation-ship between Narea and EWT. A similar relationship betweenNarea and EWT (r = 0.814, p < 0.01) was also observed in thestudy of six tree species, including both deciduous and conifer-ous species by [60]. Second, EWT is related to Narea by way ofLMA or SLA (Fig. 4). EWT (leaf water) is a major determinantof SLA, since increased EWT results in decreasing leaf tissuedensity [86], [87]. SLA is determined by leaf tissue density andleaf thickness [87], [88]. While SLA or LMA is linked to Narea
through the photosynthetic process, a connection between EWTand Narea is also expected. The last reason is possibly dueto the role that leaf water plays in leaves. It provides inter-actions with nitrogen content in an indirect way, by trans-porting nutrients and acting as a regulator of photosynthesis[89]–[91].
Among the four regression models involved in nitrogen esti-mation, Narea was most accurately estimated with a simplelinear regression model using retrieved EWT than retrievedLMA or CHLarea. A possible explanation is that EWT gener-ates a model with a high explained variance of Narea comparedwith other leaf traits. Another reason is that EWT may havebeen retrieved with a higher accuracy than LMA, as confirmedin other studies [40], [92]. Though the multiple regressionmodel using measured LMA and CHLarea gave accurate esti-mation of Narea, its performance was deteriorated when itwas applied to the leaf traits retrieved from the PROSPECTinversion. This is because the error in model inversion waspropagated into the regression model and affected the accuracyof nitrogen estimation. Similar phenomenon was observed inthe simple linear regression models using LMA or CHLarea asan independent variable. Although error-propagation is true forall regression models, it is more severe for models using LMAor CHLarea than that with EWT, due to different accuracies ofthe retrieved parameters from model inversion. A higher accu-racy of predicted leaf nitrogen content is expected if LMA andCHLarea are retrieved more accurately from model inversion.
In the establishment of multiple linear regression models,e.g., (4), (6), and (7) in Table VI, the improvement in theaccuracy of predicting nitrogen was limited when adding extravariables to supplement EWT (Table VII), because EWT couldexplain a large part of the variance in nitrogen. These multiple
regression models generated nonsignificant coefficients, whichimpeded their further utility in nitrogen estimation by couplingwith inversion of physically based models. The reason of hav-ing nonsignificant coefficients could be the correlation betweenEWT and other leaf traits (LMA, CHLarea). However, the ideaof using combinations of leaf traits to enhance explaining thevariance of nitrogen should not be denied and needs to befurther investigated with larger datasets.
The combination of empirical and physical methods con-tributes to the moderate estimation of leaf nitrogen contentin fresh leaves. The advantage of using a physically basedmodel is the possibility of generating a large database of simu-lated spectra by varying the input parameters [51], which maybe applied to other sites. Though analysis was performed forarea-based leaf nitrogen content and leaf traits, our techniqueis also applicable to mass-based parameters. The area-basedleaf traits derived from PROSPECT model can be transformedto mass-based expressions by LMA or SLA, which can beused in empirical relationships established on mass-based leaftraits [5], [59]. It is important to estimate leaf nitrogen contenton a global scale because it is an essential biodiversity vari-able [53], [57], [93], and global maps could be generated ifhyperspectral satellites such as Sentinel [94], EnMAP [95] andHyspIRI (http://hyspiri.jpl.nasa.gov/) become operational. Ourmethod holds great value for exploring the potential of upscal-ing these results to canopy level by coupling with a vegetationcanopy reflectance model [22], [44], [96] using air-borne orspace-borne hyperspectral remote sensing.
V. CONCLUSION
Leaf nitrogen content in fresh leaves has not so far beenestimated by physically based methods [33], but here wedemonstrate an indirect estimation of leaf nitrogen by using leaftraits forcing retrieved from a physically based model inver-sion. In this study, the area-based nitrogen correlations withleaf traits were found to be higher than the mass-based cor-relations. Regression models were derived for area-based leafnitrogen content using highly correlated leaf traits (LMA, EWT,and CHLarea) as independent variables. The empirical modelsand retrieved leaf traits were combined to estimate leaf nitro-gen content. Our results indicated that EWT was retrieved witha higher accuracy than LMA or CHLarea through the inversionof PROSPECT. Area-based leaf nitrogen content was estimatedmore accurately by regression models using EWT as predic-tor than LMA or CHLarea. A linear regression model usingEWT as a predictor provided the most accurate estimation ofnitrogen. The combination of empirical and physically basedmodels serves as a reliable method for estimating leaf nitro-gen, although its transferability needs to be explored using otherdatasets. Further investigation is also needed to upscale thestudy to canopy level, coupled with a canopy radiative trans-fer model using air-borne or space-borne hyperspectral remotesensing. Regional and global mapping of leaf nitrogen will fur-ther improve our understanding of the photosynthesis process,net primary productivity, and carbon dynamics [2], [4], [8],[97]. Our study provides practical techniques for estimating
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
WANG et al.: LEAF NITROGEN CONTENT ESTIMATED BY LEAF TRAITS 9
leaf nitrogen, which will be beneficial to the assessment ofbiodiversity and ecosystem services.
ACKNOWLEDGMENT
The authors are grateful to the “Applied spectroscopy”team of the German Aerospace Center (DLR) and BavarianForest National Park for assistance in the fieldwork. They alsoacknowledge the support of the “Bavarian Data Pool” data-sharing initiative. They also thank Jackie Senior for editing themanuscript and the two anonymous reviewers for their valuablecomments.
REFERENCES
[1] P. B. Reich, “Key canopy traits drive forest productivity,” Proc. Roy. Soc.B, vol. 279, no. 1736, pp. 2128–2134, 2012.
[2] P. B. Reich et al., “Nitrogen limitation constrains sustainability of ecosys-tem response to CO2,” Nature, vol. 440, no. 7086, pp. 922–925, Apr.2006.
[3] P. B. Reich, D. S. Ellsworth, and M. B. Walters, “Leaf structure (specificleaf area) modulates photosynthesis–nitrogen relations: Evidence fromwithin and across species and functional groups,” Funct. Ecol., vol. 12,no. 6, pp. 948–958, 1998.
[4] J. Evans, “Photosynthesis and nitrogen relationships in leaves of C3plants,” Oecologia, vol. 78, no. 1, pp. 9–19, 1989.
[5] C. Field and H. A. Mooney, “The photosynthesis-nitrogen relation-ship in wild plants,” in On the Economy of Plant Form and Function,T. J. Givinsh, Ed. Cambridge, U.K.: Cambridge Univ. Press, 1986,pp. 25–55.
[6] M. L. Smith et al., “Direct estimation of aboveground forest productivitythrough hyperspectral remote sensing of canopy nitrogen,” Ecol. Appl.,vol. 12, no. 5, pp. 1286–1302, Oct. 2002.
[7] D. S. Green, J. E. Erickson, and E. L. Kruger, “Foliar morphologyand canopy nitrogen as predictors of light-use efficiency in terrestrialvegetation,” Agric. Forest Meteorol., vol. 115, no. 3–4, pp. 163–171,2003.
[8] S. V. Ollinger and M. L. Smith, “Net primary production and canopynitrogen in a temperate forest landscape: An analysis using imaging spec-troscopy, modeling and field data,” Ecosystems, vol. 8, no. 7, pp. 760–778,Nov. 2005.
[9] S. V. Ollinger et al., “Canopy nitrogen, carbon assimilation, and albedoin temperate and boreal forests: Functional relations and potential climatefeedbacks,” Proc. Nat. Acad. Sci. U.S.A., vol. 105, no. 49, pp. 19336–19341, Dec. 2008.
[10] J. F. Lamarque et al., “Assessing future nitrogen deposition and car-bon cycle feedback using a multimodel approach: Analysis of nitro-gen deposition,” J. Geophys. Res. Atmos., vol. 110, no. D19303,Oct. 2005 [Online]. Available: http://onlinelibrary.wiley.com/doi/10.1029/2005JD005825/full
[11] H. Sievering, I. Fernandez, J. Lee, J. Hom, and L. Rustad, “Forestcanopy uptake of atmospheric nitrogen deposition at eastern US conifersites: Carbon storage implications?,” Global Biogeochem. Cycles, vol. 14,no. 4, pp. 1153–1159, Dec. 2000.
[12] S. L. Morford, B. Z. Houlton, and R. A. Dahlgren, “Increased forestecosystem carbon and nitrogen storage from nitrogen rich bedrock,”Nature, vol. 477, no. 7362, pp. 78–81, Sep. 2011.
[13] S. E. Plummer, “Perspectives on combining ecological process modelsand remotely sensed data,” Ecol. Model., vol. 129, no. 2–3, pp. 169–186,May 2000.
[14] S. Zaehle et al., “Evaluation of 11 terrestrial carbon–nitrogen cycle mod-els against observations from two temperate free-air CO2 enrichmentstudies,” New Phytol., vol. 202, no. 3, pp. 803–822, 2014.
[15] C. Zhang et al., “A spatial-explicit dynamic vegetation model that couplescarbon, water, and nitrogen processes for arid and semiarid ecosystems,”J. Arid Land, vol. 5, no. 1, pp. 102–117, 2013.
[16] C. A. Wessman, J. D. Aber, D. L. Peterson, and J. M. Melillo, “Remotesensing of canopy chemistry and nitrogen cycling in temperate forestecosystems,” Nature, vol. 335, no. 6186, pp. 154–156, Sep. 1988.
[17] M. E. Martin and J. D. Aber, “High spectral resolution remote sensingof forest canopy lignin, nitrogen, and ecosystem processes,” Ecol. Appl.,vol. 7, no. 2, pp. 431–443, May 1997.
[18] P. Miphokasap, K. Honda, C. Vaiphasa, M. Souris, and M. Nagai,“Estimating canopy nitrogen concentration in sugarcane using field imag-ing spectroscopy,” Remote Sens., vol. 4, no. 6, pp. 1651–1670, Jun.2012.
[19] Y. Inoue, E. Sakaiya, Y. Zhu, and W. Takahashi, “Diagnostic mapping ofcanopy nitrogen content in rice based on hyperspectral measurements,”Remote Sens. Environ., vol. 126, pp. 210–221, Nov. 2012.
[20] P. M. Hansen and J. K. Schjoerring, “Reflectance measurement of canopybiomass and nitrogen status in wheat crops using normalized differencevegetation indices and partial least squares regression,” Remote Sens.Environ., vol. 86, no. 4, pp. 542–553, Aug. 2003.
[21] K. Yu et al., “Remotely detecting canopy nitrogen concentration anduptake of paddy rice in the Northeast China Plain,” ISPRS J. Photogramm.Remote Sens., vol. 78, pp. 102–115, Apr. 2013.
[22] J. Clevers and L. Kooistra, “Using hyperspectral remote sensing data forretrieving canopy chlorophyll and nitrogen content,” IEEE J. Sel. TopicsAppl. Earth Observ. Remote Sens., vol. 5, no. 2, pp. 574–583, Apr. 2012.
[23] P. F. Chen et al., “Critical nitrogen curve and remote detection of nitrogennutrition index for corn in the Northwestern Plain of Shandong Province,China,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 6,no. 2, pp. 682–689, Apr. 2013.
[24] D. L. Peterson et al., “Remote sensing of forest canopy and leaf bio-chemical contents,” Remote Sens. Environ., vol. 24, no. 1, pp. 85–108,Feb. 1988.
[25] R. F. Kokaly and R. N. Clark, “Spectroscopic determination of leafbiochemistry using band-depth analysis of absorption features and step-wise multiple linear regression,” Remote Sens. Environ., vol. 67, no. 3,pp. 267–287, Mar. 1999.
[26] P. J. Curran, J. L. Dungan, and D. L. Peterson, “Estimating the foliarbiochemical concentration of leaves with reflectance spectrometry testingthe Kokaly and Clark methodologies,” Remote Sens. Environ., vol. 76,no. 3, pp. 349–359, Jun. 2001.
[27] N. C. Coops, M. L. Smith, M. E. Martin, and S. V. Ollinger, “Predictionof eucalypt foliage nitrogen content from satellite-derived hyperspectraldata,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 6, pp. 1338–1346,Jun. 2003.
[28] M. E. Martin, L. C. Plourde, S. V. Ollinger, M. L. Smith, andB. E. McNeil, “A generalizable method for remote sensing of canopynitrogen across a wide range of forest ecosystems,” Remote Sens.Environ., vol. 112, no. 9, pp. 3511–3519, Sep. 2008.
[29] C. Axelsson, A. K. Skidmore, M. Schlerf, A. Fauzi, and W. Verhoef,“Hyperspectral analysis of mangrove foliar chemistry using PLSR andsupport vector regression,” Int. J. Remote Sens., vol. 34, no. 5, pp. 1724–1743, 2013.
[30] Z. Huang, B. J. Turner, S. J. Dury, I. R. Wallis, and W. J. Foley,“Estimating foliage nitrogen concentration from HYMAP data usingcontinuum removal analysis,” Remote Sens. Environ., vol. 93, no. 1–2,pp. 18–29, Oct. 2004.
[31] O. Mutanga and A. K. Skidmore, “Integrating imaging spectroscopy andneural networks to map grass quality in the Kruger National Park, SouthAfrica,” Remote Sens. Environ., vol. 90, no. 1, pp. 104–115, 2004.
[32] A. K. Skidmore et al., “Forage quality of savannas—Simultaneouslymapping foliar protein and polyphenols for trees and grass using hyper-spectral imagery,” Remote Sens. Environ., vol. 114, no. 1, pp. 64–72,2010.
[33] S. Jacquemoud et al., “Estimating leaf biochemistry using thePROSPECT leaf optical properties model,” Remote Sens. Environ.,vol. 56, no. 3, pp. 194–202, Jun. 1996.
[34] S. Jacquemoud and F. Baret, “PROSPECT: A model of leaf optical prop-erties spectra,” Remote Sens. Environ., vol. 34, no. 2, pp. 75–91, Nov.1990.
[35] R. Darvishzadeh, C. Atzberger, A. Skidmore, and M. Schlerf, “Mappinggrassland leaf area index with airborne hyperspectral imagery: A com-parison study of statistical approaches and inversion of radiative transfermodels,” ISPRS J. Photogramm. Remote Sens., vol. 66, no. 6, pp. 894–906, 2011.
[36] M. Schlerf and C. Atzberger, “Inversion of a forest reflectance modelto estimate structural canopy variables from hyperspectral remote sens-ing data,” Remote Sens. Environ., vol. 100, no. 3, pp. 281–294, Feb.2006.
[37] F. Baret and T. Fourty, “Estimation of leaf water content and specific leafweight from reflectance and transmittance measurements,” Agronomie,vol. 17, no. 9–10, pp. 455–464, 1997.
[38] G. le Maire, C. Francois, and E. Dufrene, “Towards universal broad leafchlorophyll indices using PROSPECT simulated database and hyperspec-tral reflectance measurements,” Remote Sens. Environ., vol. 89, no. 1,pp. 1–28, Jan. 2004.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
[39] D. Riano, P. Vaughan, E. Chuvieco, P. J. Zarco-Tejada, and S. L. Ustin,“Estimation of fuel moisture content by inversion of radiative transfermodels to simulate equivalent water thickness and dry matter content:Analysis at leaf and canopy level,” IEEE Trans. Geosci. Remote Sens.,vol. 43, no. 4, pp. 819–826, Apr. 2005.
[40] R. Colombo et al., “Estimation of leaf and canopy water content in poplarplantations by means of hyperspectral indices and inverse modeling,”Remote Sens. Environ., vol. 112, no. 4, pp. 1820–1834, Apr. 2008.
[41] J. B. Feret et al., “PROSPECT-4 and 5: Advances in the leaf optical prop-erties model separating photosynthetic pigments,” Remote Sens. Environ.,vol. 112, no. 6, pp. 3030–3043, Jun. 2008.
[42] Y. Q. Zhang, J. M. Chen, J. R. Miller, and T. L. Noland, “Leaf chlorophyllcontent retrieval from airborne hyperspectral remote sensing imagery,”Remote Sens. Environ., vol. 112, no. 7, pp. 3234–3247, Jul. 2008.
[43] R. Darvishzadeh, A. Skidmore, M. Schlerf, and C. Atzberger, “Inversionof a radiative transfer model for estimating vegetation LAI and chloro-phyll in a heterogeneous grassland,” Remote Sens. Environ., vol. 112,no. 5, pp. 2592–2604, 2008.
[44] K. Omari, H. P. White, K. Staenz, and D. J. King, “Retrieval of for-est canopy parameters by inversion of the PROFLAIR leaf-canopyreflectance model using the LUT approach,” IEEE J. Sel. TopicsAppl. Earth Observ. Remote Sens., vol. 6, no. 2, pp. 715–723, Apr.2013.
[45] T. P. Dawson, P. J. Curran, and S. E. Plummer, “LIBERTY—Modeling theeffects of leaf biochemical concentration on reflectance spectra,” RemoteSens. Environ., vol. 65, no. 1, pp. 50–60, Jul. 1998.
[46] R. F. Kokaly, G. P. Asner, S. V. Ollinger, M. E. Martin, andC. A. Wessman, “Characterizing canopy biochemistry from imagingspectroscopy and its application to ecosystem studies,” Remote Sens.Environ., vol. 113, pp. S78–S91, Sep. 2009.
[47] T. Fourty, F. Baret, S. Jacquemoud, G. Schmuck, and J. Verdebout, “Leafoptical properties with explicit description of its biochemical composi-tion: Direct and inverse problems,” Remote Sens. Environ., vol. 56, no. 2,pp. 104–117, May 1996.
[48] F. Baret and T. H. Fourty, “Radiometric estimates of nitrogen statusof leaves and canopies,” in Diagnosis of the Nitrogen Status in Crops,G. Lemaire, Ed. New York, NY: Springer, 1997, pp. 201–227.
[49] D. S. Ellsworth and P. B. Reich, “Canopy structure and vertical patternsof photosynthesis and related leaf traits in a deciduous forest,” Oecologia,vol. 96, no. 2, pp. 169–178, 1993.
[50] L. Homolova, Z. Maenovsky, J. Clevers, G. Garcia-Santos, andM. E. Schaeprnan, “Review of optical-based remote sensing for plant traitmapping,” Ecol. Complexity, vol. 15, pp. 1–16, Sep. 2013.
[51] G. le Maire et al., “Calibration and validation of hyperspectral indices forthe estimation of broadleaved forest leaf chlorophyll content, leaf massper area, leaf area index and leaf canopy biomass,” Remote Sens. Environ.,vol. 112, no. 10, pp. 3846–3864, Oct. 2008.
[52] S. L. Ustin, S. Jacquemoud, P. J. Zarco-Tejada, and G. P. Asner, “Remotesensing of the environment: State of the science and new directions,”Manual Remote Sens., vol. 4, pp. 679–730, 2004.
[53] G. P. Asner and R. E. Martin, “Airborne spectranomics: Mappingcanopy chemical and taxonomic diversity in tropical forests,” Front. Ecol.Environ., vol. 7, no. 5, pp. 269–276, 2008.
[54] Ü. Niinemets, “Role of foliar nitrogen in light harvesting and shade toler-ance of four temperate deciduous woody species,” Funct. Ecol., vol. 11,no. 4, pp. 518–531, 1997.
[55] A. Rosati, K. R. Day, and T. M. DeJong, “Distribution of leaf mass perunit area and leaf nitrogen concentration determine partitioning of leafnitrogen within tree canopies,” Tree Physiol., vol. 20, no. 4, pp. 271–276,2000.
[56] K. Takahashi, T. Seino, and T. Kohyama, “Plastic changes of leaf mass perarea and leaf nitrogen content in response to canopy openings in saplingsof eight deciduous broad-leaved tree species,” Ecol. Res., vol. 20, no. 1,pp. 17–23, Jan. 2005.
[57] I. J. Wright et al., “The worldwide leaf economics spectrum,” Nature,vol. 428, no. 6985, pp. 821–827, 2004.
[58] K. Hikosaka, “Interspecific difference in the photosynthesis–nitrogenrelationship: Patterns, physiological causes, and ecological importance,”J. Plant Res., vol. 117, no. 6, pp. 481–494, 2004.
[59] J. Lloyd, K. Bloomfield, T. F. Domingues, and G. D. Farquhar,“Photosynthetically relevant foliar traits correlating better on a mass vsan area basis: of ecophysiological relevance or just a case of mathemati-cal imperatives and statistical quicksand?,” New Phytol., vol. 199, no. 2,pp. 311–321, 2013.
[60] F. B. Sullivan et al., “Foliar nitrogen in relation to plant traits andreflectance properties of New Hampshire forests,” Can. J. Forest Res.,vol. 43, no. 1, pp. 18–27, 2012.
[61] M. Heurich and M. Neufanger, “Die Wälder des NationalparksBayerischer Wald,” Ergebnisse der Waldinventur 2002/2003 imgeschichtlichen und waldökologischen Kontext, Grafenau, 178 S,Germany, 2005.
[62] A. Lausch, M. Heurich, and L. Fahse, “Spatio-temporal infestation pat-terns of Ips typographus (L.) in the Bavarian Forest National Park,Germany,” Ecol. Indic., vol. 31, pp. 73–81, Aug. 2013.
[63] J. Markwell, J. Osterman, and J. Mitchell, “Calibration of the MinoltaSPAD-502 leaf chlorophyll meter,” Photosynth. Res., vol. 46, no. 3,pp. 467–472, 1995.
[64] C. Atzberger, T. Jarmer, M. Schlerf, B. Kötz, and W. Werner,“Spectroradiometric determination of wheat bio-physical variables:Comparison of different empirical-statistical approaches,” in Proc. 23rdEARSeL Symp. Remote Sens. Transitions, Belgium, 2003, pp. 463–470.
[65] K. Yu, “Hyperspectral remote sensing of crop canopy chlorophyll andnitrogen: The relative importance of growth stages,” Ph.D. dissertation,Universität zu Köln, Cologne, Germany, 2014.
[66] I. Novozamsky, V. J. G. Houba, R. van Eck, and W. van Vark, “A noveldigestion technique for multi- element plant analysis,” Commun. Soil Sci.Plant Anal., vol. 14, no. 3, pp. 239–248, 1983.
[67] Integrating Sphere User Manual, Analytical Spectral Devices, Inc.(ASD), Boulder, CO, 2008.
[68] A. Savitzky and M. J. E. Golay, “Smoothing and differentiation of databy simplified least squares procedures,” Anal. Chem., vol. 36, no. 8,pp. 1627–1639, 1964.
[69] K. S. Schmidt and A. K. Skidmore, “Smoothing vegetation spectra withwavelets,” Int. J. Remote Sens., vol. 25, no. 6, pp. 1167–1184, 2004.
[70] J. Neter, W. Wasserman, and M. H. Kutner, Applied Linear StatisticalModels, 3rd ed. Homewood, IL, USA: Irwin, 1990.
[71] R. M. O’Brien, “A caution regarding rules of thumb for variance inflationfactors,” Qual. Quantity, vol. 41, no. 5, pp. 673–690, 2007.
[72] M. Schlerf, C. Atzberger, and J. Hill, “Remote sensing of forest biophys-ical variables using HyMap imaging spectrometer data,” Remote Sens.Environ., vol. 95, no. 2, pp. 177–194, 2005.
[73] J. Bring, “How to standardize regression coefficients,” Amer. Stat.,vol. 48, no. 3, pp. 209–213, 1994.
[74] R. S. Landis, “Standardized regression coefficients,” in Encyclopedia ofStatistics in Behavioral Science. Hoboken, NJ, USA: Wiley, 2005.
[75] J. D’Errico. (2012, Feb 06). Fminsearchbnd, Fminsearchcon [Online].Available: http://www.mathworks.com/matlabcentral/fileexchange/8277-fminsearchbnd–fminsearchcon/content/FMINSEARCHBND/fminsearchbnd.m
[76] P. H. Li and Q. Wang, “Retrieval of leaf biochemical parameters usingPROSPECT inversion: A new approach for alleviating ill-posed prob-lems,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 7, pp. 2499–2506,Jul. 2011.
[77] G. P. Asner and P. M. Vitousek, “Remote analysis of biological inva-sion and biogeochemical change,” Proc. Nat. Acad. Sci. U.S.A., vol. 102,no. 12, pp. 4383–4386, 2005.
[78] A. Romero, I. Aguado, and M. Yebra, “Estimation of dry matter contentin leaves using normalized indexes and PROSPECT model inversion,”Int. J. Remote Sens., vol. 33, no. 2, pp. 396–414, 2012.
[79] L. Poorter, S. F. Oberbauer, and D. B. Clark, “Leaf optical propertiesalong a vertical gradient in a tropical rain forest canopy in Costa Rica,”Amer. J. Bot., vol. 82, no. 10, pp. 1257–1263, 1995.
[80] J. R. Evans and H. Poorter, “Photosynthetic acclimation of plants togrowth irradiance: The relative importance of specific leaf area and nitro-gen partitioning in maximizing carbon gain,” Plant Cell Environ., vol. 24,no. 8, pp. 755–767, 2001.
[81] U. Niinemets, “A review of light interception in plant stands from leafto canopy in different plant functional types and in species with varyingshade tolerance,” Ecol. Res., vol. 25, no. 4, pp. 693–714, Jul. 2010.
[82] M. T. Harrison, E. J. Edwards, G. D. Farquhar, A. B. Nicotra, andJ. R. Evans, “Nitrogen in cell walls of sclerophyllous leaves accountsfor little of the variation in photosynthetic nitrogen-use efficiency,” PlantCell Environ., vol. 32, no. 3, pp. 259–270, 2009.
[83] J. R. Evans, “Nitrogen and photosynthesis in the flag leaf of wheat(Triticum aestivum L.),” Plant Physiol., vol. 72, no. 2, pp. 297–302,1983.
[84] F. S. Chapin, III, A. J. Bloom, C. B. Field, and R. H. Waring, “Plantresponses to multiple environmental factors,” Bioscience, vol. 37, no. 1,pp. 49–57, Jan. 1987.
[85] S. L. Ustin, D. Riano, and E. R. Hunt, “Estimating canopy water contentfrom spectroscopy,” Isr. J. Plant Sci., vol. 60, no. 1–2, pp. 9–23, 2012.
[86] D. Meziane and B. Shipley, “Interacting determinants of specific leaf areain 22 herbaceous species: Effects of irradiance and nutrient availability,”Plant Cell Environ., vol. 22, no. 5, pp. 447–459, 1999.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
WANG et al.: LEAF NITROGEN CONTENT ESTIMATED BY LEAF TRAITS 11
[87] E. Garnier and G. Laurent, “Leaf anatomy, specific mass and water con-tent in congeneric annual and perennial grass species,” New Phytol.,vol. 128, no. 4, pp. 725–736, 1994.
[88] E. T. F. Witkowski and B. Lamont, “Leaf specific mass confounds leafdensity and thickness,” Oecologia, vol. 88, no. 4, pp. 486–493, 1991.
[89] P. J. Kramer and J. S. Boyer, Water Relations of Plants and Soils. NewYork, NY: Academic, 1995.
[90] H. Asbjornsen et al., “Ecohydrological advances and applications inplant–water relations research: A review,” J. Plant Ecol., vol. 4, no. 1–2,pp. 3–22, 2011.
[91] R. H. Waring and J. J. Landsberg, “Generalizing plant–water relations tolandscapes,” J. Plant Ecol., vol. 4, no. 1–2, pp. 101–113, 2011.
[92] J. Clevers, L. Kooistra, and M. E. Schalepman, “Using spectral informa-tion from the NIR water absorption features for the retrieval of canopywater content,” Int. J. Appl. Earth Observ. Geoinf., vol. 10, no. 3,pp. 388–397, Sep. 2008.
[93] H. M. Pereira et al., “Essential biodiversity variables,” Science, vol. 339,no. 6117, pp. 277–278, 2013.
[94] Z. Malenovsky et al., “Sentinels for science: Potential of sentinel-1, -2,and -3 missions for scientific observations of ocean, cryosphere, andland,” Remote Sens. Environ., vol. 120, pp. 91–101, May 2012.
[95] H. Kaufmann et al., “Environmental mapping and analysis program(EnMAP)–recent advances and status,” in Proc. IEEE Int. Geosci. RemoteSens. Symp. (IGARSS’08), Boston, MA, 2008, pp. IV-109–IV-112.
[96] W. Verhoef and H. Bach, “Coupled soil-leaf-canopy and atmosphereradiative transfier modeling to simulate hyperspectral multi-angular sur-face reflectance and TOA radiance data,” Remote Sens. Environ., vol. 109,no. 2, pp. 166–182, Jul. 2007.
[97] M. Heimann and M. Reichstein, “Terrestrial ecosystem carbon dynamicsand climate feedbacks,” Nature, vol. 451, no. 7176, pp. 289–292, 2008.
Zhihui Wang received the B.S. degree in geo-graphical information science from Nanjing NormalUniversity, Nanjing, China, in 2009, and the M.E.degree in signal and information processing from theInstitute of Remote Sensing Applications, ChineseAcademy of Sciences, Beijing, China, in 2012.She is currently pursuing the joint Ph.D. degree atthe Faculty of Geo-Information Science and EarthObservation (ITC), University of Twente, and at theSchool of Mathematical and Geospatial Sciences,RMIT University, Melbourne, Vic., Australia.
Her research interests include the retrieval of leaf biochemical parametersusing physically based models and hyperspectral remote sensing.
Andrew K. Skidmore received the Ph.D. degree inremote sensing and GIS from the Australian NationalUniversity, Canberra, Australia, in 1989.
He is the Professor of Spatial EnvironmentalResource Dynamics with the University of Twenteand Chairman of the Department of NaturalResources at the Faculty of ITC. He has authoredmore than 190 ISI journal articles and 17 bookchapters. His research interests include hyperspectralremote sensing, habitat monitoring under fragmenta-tion and climate change, as well as image processing
and more generally techniques for handling geo-information.
Roshanak Darvishzadeh received the Ph.D. degreein hyperspectral remote sensing of vegetation fromITC, Enschede, The Netherlands, and WageningenUniversity, Wageningen, The Netherlands, in 2008.
She is currently an Assistant Professor with theDepartment of Natural Resources, University ofTwente (Faculty ITC), Enschede, The Netherlands.Her research interests include quantitative remotesensing for mapping and modeling biophysical andbiochemical properties of vegetation with the use ofstatistical and radiative transfer models.
Uta Heiden received the Ph.D. degree in urbanimaging spectroscopy and urban ecology from theTechnical University of Berlin, Germany, in 2003.
She is the Head of the Research Group “AppliedSpectroscopy” at the German Aerospace Center inOberpfaffenhofen, Germany. Her research interestsfocus on developing spectroscopic methods in thebiodiversity context.
Dr. Heiden is the Application Support Manager ofEnMAP Ground Segment, member of the EnMAPScience Advisory Group (EnSAG), and Co-Chair of
the International Spaceborne Imaging Spectroscopy (ISIS) Working Group(Technical Committee of IEEE GRSS).
Marco Heurich received the B.S. degree inforestry from the University of Applied ScienceWeihenstephan-Triesdorf, Freising, Germany, in1994, the M.S. degree in geographical science andsystems from the University of Salzburg, Salzburg,Austria, in 2003, and the Ph.D. degree in forestryfrom TU München, München, Germany, in 2006.
He is a Senior Researcher with the Departmentof Conservation and Research of Bavarian ForestNational Park and a Postdoctoral Lecturer withthe Faculty of Environment and Natural Resources,
University of Freiburg, Germany, teaching wildlife management in nationalparks and research in wildlife ecology. His research interests include remotesensing and its applications in forest ecology and in wildlife ecology andmanagement.
Dr. Heurich received a scholarship for highly gifted students from Friedrich-Naumann-Stiftung and is the recipient of the Lennart-Bernadotte-Award forlandscape ecology.
Tiejun Wang received the M.S. degree in GISand remote sensing from the International Institutefor Geo-information Science and Earth Observation(ITC), Enschede, The Netherlands, in 2003 and thePh.D. degree in spatial ecology from WageningenUniversity, Wageningen, The Netherlands, in 2009.
He works as an Assistant Professor with theFaculty of Geo-information Science and EarthObservation, University of Twente, The Netherlands.His research interests include GIS, remote sensingand ecology, and its applications to biodiversity,
ecosystems, and conservation.Dr. Wang currently serves as an Associate Editor for the journal Wetlands. He
is a member of several professional associations such as The Wildlife Society,Society of Wetland Scientists, and GEOSS GEO-BON.