Estimating nitrogen status of rice using the image segmentation of G-R thresholding method

Field Crops Research 149(2013)33–39Contents lists available at SciVerse ScienceDirectField CropsResearchj o u r n a l h o m e p a g e :w w w.e l s e v i e r.c o m /l o c a t e /f crEstimating nitrogen status of rice using the image segmentation of G-R thresholding methodYuan Wang,Dejian Wang ∗,Gang Zhang,Jun WangInstitute of Soil Science,Chinese Academy of Sciences,71East Beijing Road,Nanjing 210008,Chinaa r t i c l ei n f oArticle history:Received 15October 2012Received in revised form 7April 2013Accepted 7April 2013Keywords:Digital camera Visible spectrumImage processing technology Nitrogen nutrition Ricea b s t r a c tA camera can record spectral information of visible bands.In this study,a digital camera was used to take pictures of the canopies of 3rice (Oryza sativa L.)cultivars with 6different nitrogen (N)application rates.Canopy images were segmented by setting threshold values based on the magnitude and distribution of the green channel minus red channel (GMR)value,and then correlations were established between image feature parameters and the 3plant indices (i.e.,above-ground biomass,N content and leaf area index)before and after image segmentation.Results showed significant exponential relationships between the image parameters and the plant indices.Before the segmentation,the GMR values were closely related to the 3plant indices,with correlation coefficient of 0.93**,0.93**and 0.94**,respectively;while after the segmentation,the correlation coefficients between canopy cover (CC)and plant indices were 0.90**,0.91**and 0.95**,respectively.We conclude that GMR and CC will be valid indicators in the application of N diagnosis both for japonica and indica rice.And the canopy image segmentation method is fast in data processing and easily adaptable.©2013Elsevier B.V.All rights reserved.1.IntroductionAs nitrogen (N)is the most important nutrient essential for the growth of crops,raised N application rate can effectively increase crop yields.But when it comes to over application,it would also cause a series of environmental problems and even yield decrease (Wang et al.,2003;Zhu,2000).Obtaining crop N status timely not only enable farmers to increase nitrogen use efficency (NUE),but also reduce the water and atmospheric pollution caused by exces-sive N application (Miao et al.,2011;Peng et al.,2002).Chlorophyll meter (SPAD-502)has been extensively used in studies on nutrition diagnosis of various crops by rapid determi-nation of relative chlorophyll contents.Peng et al.(2006)showed that the real-time nitrogen management (RTNM)of rice depending on the SPAD threshold value improved NUE by 30%or more with-out affecting the yield.However,the measuring area of SPAD-502is only 6mm 2,and it has to do a large numbers of repeated deter-mination before a reliable result can be obtained (Blackmer and Schepers,1995).Moreover,SPAD readings vary sharply between crops,varieties and growth stages.Therefore,only after calibrationAbbreviations:CC,canopy cover;NRI,normalized redness intensity;NGI,nor-malized greenness intensity;GMR,green channel minus red channel;LNC,leaf nitrogen concentration;LNA,leaf nitrogen accumulation.∗Corresponding author.Tel.:+862586881253;fax:+862586881253.E-mail address:djwang@ (D.Wang).of its readings,can SPAD improve its applicability (Lin et al.,2010;Peng et al.,1993).Leaf nitrogen concentration (LNC)reflects in the leaf color intu-itively and is easy to obtain.Therefore,the indicators used by most of the researches on fertilization recommendation are LNC or rela-tive LNC,such as the application of portable chlorophyll meter and hyperspectral sensors.However,there are significant changes for LNC during the rice growing period,even in the case of excessive N (Zhao et al.,2006),so it is difficult to determine the N sta-tus without the ‘sufficiency indices’,which was calculated from readings of chlorophyll meter or other devices relative to well-fertilized reference plots (Bausch and Brodahl,2012;Hussain et al.,2000;Samborski et al.,2009).Leaf nitrogen accumulation (LNA)and above-ground N accumulation have also been used for fertil-ization recommendation in many researches besides the LNC.Li et al.(2010)also pointed out that N concentration varied with the amount of N as well as the amount of biomass,so N taken up early in crop development led to increased growth but lower N concen-tration than a similar amount of N taken up later,and he concluded that “the most successful indicators for crop N diagnosis are those that measure,or are correlated with,the mass of N per unit ground area”(Flowers et al.,2003;Ju and Christie,2011;Lukina et al.,1999).With the development of the remote-sensing technology in recent years,the technologies of satellite imagery,aero-photographing (Williams et al.,2010)and hyperspectral remote sensing (Hansen and Schjoerring,2003)are widely used in the stud-ies on nutrition diagnosis of crops.Being one of the most convenient0378-4290/$–see front matter ©2013Elsevier B.V.All rights reserved./10.1016/j.fcr.2013.04.00734Y.Wang et al./Field Crops Research149(2013)33–39tools for remote sensing of visible spectrum,digital cameras are also being widely used.The application of cameras in agricultural monitoring began in the1990s,mainly in automatic quality grad-ing of agricultural products(Zhang et al.,2011)and in detecting weeds(Gerhards and Oebel,2006),pests and diseases(Dammer et al.,2011).Kawashima and Nakatani(1998)obtained the nutri-ent status of plants based on the estimation of chlorophyll content using a portable color video camera,but they used manul image extraction method and it was highly subjective to human prefer-ence.Most of the researches on crop N diagnosis are based on taking vertical pictures using digital camera,and segment the images and extract various kinds of parameters,such as green/red(Adamsen et al.,1999)and canopy cover(Li et al.,2010).Rorie et al.(2011) took pictures of corn leaf with a digital camera under artificial light. The picture was calibrated with reference color and turned into HSV (hue,saturation and value)color space.Its feature parameter was found to have good relationship with N content in the leaf and yield of the crop.To diagnose N status of a crop,the application of digital cameras and image processing techniques is less expensive than the use of other techniques,such as hyperspectral remote sensing and SPAD meter.Moreover,the sample image in the former covers an area much bigger than a SPAD could.On the basis of previous researches, this paper proposes a method for segmenting images of rice canopy by setting threshold value based on the magnitude and distribution of green channel minus red channel(GMR)value,and analyzes the relationships of image feature parameters before and after image segmentation with above-ground biomass,N content and leaf area index(LAI).In addition,the effectiveness of this method to diagnose N status of rice is also explored.2.Materials and methods2.1.General information of the experiment siteThe experiment was laid out in the Changshu Agricultural Ecology Experiment Station,Changshu,Jiangsu,China(120◦42 E, 31◦33 N).Located in the humid subtropical climate zone,the station enjoys annually493KJ cm−2of solar radiation,1800h of sunshine, 1200mm of precipitation and4933◦C of cumulative temperature ≥10◦C.The soil type for thefield experiment site is the gleyed paddy soil of the Taihu Lake region,containing1.79g kg−1of total nitrogen(TN),0.93g kg−1of total phosphorus(TP),18.7g kg−1of total potassium(TK),30.8g kg−1of organic matter,123mg kg−1of alkalytic N,13.1mg kg−1of plant available P and121mg kg−1of plant available K and pH of7.4(1:2,soil:water)in the0–15cm soil layer.2.2.Experiment designThis study used data from two independently designed N fertilizer gradient experiments.Experiment I was a long-term site-specific rice–wheat rotation experiment that started in1997and had six treatments,i.e.,CK,N0,N1,N2,N3and N4,designed to receive0,0,180,225,270and315kg N ha−1in rice season,respec-tively,and20kg P ha−1and90kg K ha−1for all the treatments except CK,and four replicates for each treatment.The data were selected from the2011rice growing season(May–November)with Nanjing46(NJ46)cultivar.Experiment II was the one carried out in 2011with paddyfields in a rice–wheat rotation.It was designed to have six N application rates,two rice cultivars and three repli-cates for each treatment.The six N application rates were0,120, 180,240,270and300kg N ha−1.In addition to the N application, each treatment received20kg P ha−1and90kg K ha−1.The two rice cultivars involved were japonica rice,Nanjing45(NJ45)and hybrid Table1Sampling dates of the three rice cultivars and numbers of samples.Cultivars First time Second time Third time Fourth timeNJ458July(18)a21July(18)12August(18)18August b(18) NJ464July(24)21July(24)10August(24)30August(24) LYP98July(18)21July(18)12August(18)26August(18)a Number in parentheses represents sampling number.b NJ45was short in life cycle,so the interval between the third and fourth sam-plings was only6days.indica rice,Liangyoupeijiu(LYP9).For both experiments the N was split into three applications,40%as basal fertilizer,20%as tillering fertilizer and40%as ear fertilizer,and the K application was split into50%as basal fertilizer and50%as ear fertilizer,and the P was applied once as basal fertilizer.Otherfield managements followed the practices as in ordinary farmlands.2.3.Sample collection and measurementThe above-ground part of rice was sampled every15days or so from transplanting to earing stage in2011.A total of4sets of samples were collected(Table1).The sample were oven-dried and then analyzed for biomass and TN.On the same day LAI was measured using a canopy analyzer(SunScan System-SS1,Delta-T Devices Ltd.).TN was measured using the Kjeldahl method(Lu, 2000).On the same day or the following day of sampling,pictures of the rice canopy were taken using a digital camera(EOS50D,Canon. Inc.).The camera was positioned1m above the top of the canopy using a tripod.Aperture priority mode was selected and the cam-era was set at5.6aperture,100ISO,4900K white balance and auto-focus with theflash turned off.The pictures were taken at 12:00–13:00of an overcast day without direct sunlight and the illuminance was about30–50thousand lux.Thus,when using auto exposure the actual exposure time of the camera was1/500to 1/320s.Pictures were stored in CR2raw format with a resolution of4752×3168.2.4.Image segmentation and data analysisThe pictures taken before the canopy reached100%in coverage contained some non-canopy elements,like soil and plant residues. Therefore,analyses were done separately of images before and after segmentation.Unsegmented or intact images contained some non-canopy elements but the segmented ones contained only canopy.Images were displayed in the RGB color model,each pixels in the image represented by an RGB triplet(red,green,blue value).The segmentation of images was based on the difference of reflectance spectrum between green vegetation and soil in the visible band. Green vegetation had an intensive reflection peak in the green band, whereas soil did not cause any apparent change in albedo in the vis-ible band.Therefore,after the processing of green channel minus red channel of an image(Matlab,The MathWorks,Inc.),the differ-ence between the canopy and non-canopy area became obvious in GMR value(Wang et al.,2012).Fig.1(b)is a color scale image plot-ted on the basis of the GMR values;and the gradual change in color in the image represented the change in GMR value.The color of the canopy section was sharply different from that of the non-canopy section,so it is feasible to set a GMR threshold for segmentation of a picture.Once a threshold was set,pixels with GMR value higher than the threshold were sorted as the rice canopy and the rest as the background(soil or plant residues).A raw imagefile contains minimally processed data from the image sensor of a digital camera.Thefile saves settings of white balance,color saturation,contrast,and sharpness in it,but defersY.Wang et al./Field Crops Research149(2013)33–3935Fig.1.The image of a rice canopy captured by a digital camera and the same image processed using Matlab:(a)original image,(b)scaled green channel minus red channel (GMR)value and displays as image and(c and d)segmented images using GMR threshold15and30,respectively.Black portion of the images is regarded as non-canopy(soil and plant residues).the processing.Therefore,all the changes made on a raw imagefileare non-destructive.The images in CR2raw format were adjusted to be identical incolor saturation,contrast and brightness with Photoshop(AdobeSystems Inc.)and converted to JPGfiles.After that JPG imageswere processed with Matlab for calculation of feature parametersdirectly or after segmentation.The computation of parameters usedthe mean value of all the pixels in an image.Canopy cover(CC)is thepercentage of the number of pixels reflecting the canopy against thetotal of the whole image.Normalized redness intensity(NRI)andnormalized greenness intensity(NGI)are the proportion of red orgreen color intensity in all the three colors.GMR is the differencebetween green channel and red channel.G/R is greenness intensitydivided by redness intensity.GMR,NRI,NGI and hue(H)were cal-culated using the following equations(Jia et al.,2004;Kawashimaand Nakatani,1998;Rorie et al.,2011).R,G and B denote meanvalues of the red,green and blue channels,respectively.Data anal-ysis was done using the SPSS13.0(SPSS Inc.)and the correlationcoefficients were obtained with the Spearman correlation.GMR=G−R(1)NRI=RR+G+B(2)NGI=GR+G+B(3)HueIf max(RGB)=R,H=60×G−Bmax(RGB)−min(RGB)If max(RGB)=G,H=60×2+B−Rmax(RGB)−min(RGB)If max(RGB)=B,H=60×4+R−Gmax(RGB)−min(RGB)(4)3.Results3.1.Above-ground biomass,N content and LAI of different ricecultivarsAnalysis of variance of the N content indicated no significantdifference between the three cultivars(Table2).But the LAI ofLiangyoupeijiu(LYP9)was about43%and76%higher than that ofNanjing46(NJ46)and Nanjing45(NJ45),respectively.The biomassof LYP9was similar to that of the other two,but the differencebetween NJ45and NJ46was rather significant.3.2.Relationships between image feature parameters and rice Nstatus before image segmentationThe image feature parameters,i.e.,GMR,G/R,NGI,NRI and H,extracted from intact images were significantly correlated with riceplant indices(above-ground biomass,N content and LAI,Table3).Among them,GMR,G/R,NGI and H were positively correlated withthe three plant indices,while NRI was significantly but negativelycorrelated with plant indices.GMR had the highest correlation coef-ficient with biomass,N content and LAI,reaching0.93**,0.93**and0.94**,respectively,and followed by G/R and NGI(both about0.9),Table2Mean values of above-ground biomass(Biomass),nitrogen content(N content)andleaf area index(LAI)of the three cultivars.Cultivars Biomass(g m−2)N content(g m−2)LAI(m2m−2)SamplenumbersNJ46395.6b a8.44a 2.12a96NJ45285.9a7.36a 1.72a72LYP9363.9ab9.09a 3.03b72a Means within the same column followed by different letters were significantlydifferent according to the Tukey test(P<0.05).36Y.Wang et al./Field Crops Research149(2013)33–39Table3Spearman correlation coefficients between rice plant indices and the image feature parameters extracted from intact images.GMR G/R NGI NRI H Number ofsamples Biomass0.93**0.88**0.89**−0.42**0.19**240N content0.93**0.93**0.89**−0.53**0.14240LAI0.94**0.91**0.91**−0.45**0.22**240**P<0.01.while NRI and H had much lower correlation coefficient compared to the others.However,H had no significant relationship with N content.The image feature parameters,GMR and G/R,were well corre-lated with the three plant indices.This paper focuses on the analysis of the relationships between the two feature parameters and rice N status.Regression analysis showed that GMR and G/R had exponen-tial relations with the three plant indices(Fig.2),and the following exponential function was identified to bestfit the nonlinear rela-tionships:y=ae bx(5) where y is a dependent variable,representing above-ground biomass,N content or LAI,x is an independent variable,represent-ing GMR or G/R.Both a and b were parameters obtained by the least square method.Regression analysis was performed using the concatenation of the three cultivars,the R2between GMR and the three plant indices reached0.84,0.81and0.82,and that between G/R and the three were0.71,0.81and0.71(Fig.2).The feature parameters were in good exponential relations with above-ground biomass,N content and LAI(Fig.2).With the rice growth,GMR and G/R values of the whole image were getting higher.Fig.2clearly shows that the high-est value of GMR was around30and of G/R around1.4.When GMR and G/R reached the maximums,the plant canopy coverage of the image approached100%.Separate regression analyses between NJ45,NJ46and LYP9and plant indices are shown in Fig.2.The relationships between GMR and the above-ground biomass,N content and LAI of NJ45were highly significant with R2reaching0.93,0.93and0.95,respectively. The R2between GMR and the three plant indices of the three culti-vars ranged from0.79to0.95,and between G/R and the three plant indices of the three cultivars in the range of0.66–0.93.Different sampling dates,cultivars and N application rates did not influence the trends of the relationships.3.3.Relationships between image feature parameters and rice N status after image segmentationAccording to the distribution of GMR values in the canopy image,afixed threshold value was set at20for image segmen-tation.Significant relationships were found between the feature parameters extracted from the segmented images and the plant indices(Table4).The correlation coefficients of CC with above-ground biomass,N content and LAI reached0.90**,0.91**and0.95**, Table4Spearman correlation coefficients between rice plant indices and the image feature parameters extracted from segmented images.CC GMR G/R NGI NRI H Number ofsamples Biomass0.90**0.47**0.76**0.47**−0.58**0.44**240N content0.91**0.41**0.83**0.48**−0.69**0.55**240LAI0.95**0.35**0.76**0.42**−0.62**0.49**240**P<0.01.Table5Regression analysis between the rice plant indices and canopy cover(CC)using Eq.(5).Biomass N content LAI NJ46(n=96)Eq.y=37.3e4.5x y=0.84e4.36x y=0.14e5.06x RMSE a107 2.730.71R20.870.810.84NJ45(n=72)Eq.y=39.8e4.3x y=1.03e4.29x y=0.17e4.98x RMSE51 1.340.28R20.940.930.96LYP9(n=72)Eq.y=16.9e5.1x y=0.30e5.54x y=0.09e5.68x RMSE178 4.18 1.13R20.670.700.80a Root mean squared error.respectively.All the feature parameters showed positive relation-ships,except for NRI,which displayed a negative one.An additional feature parameter CC was extracted from the seg-mented images,which is closely related to above-ground biomass, N content and LAI(Table5).The correlation coefficients of the other feature parameters extracted from a segmented image with rice plant indices varied to a various degree from those extracted from its intact one.Those of GMR,G/R and NGI extracted from a seg-mented image with above-ground biomass,N content and LAI were significantly lower than those from its intact ones,about49%,56% and63%lower for GMR,and14%,11%and16%lower for G/R,respec-tively.The correlation coefficients of NRI and H with the three plant indices increased somewhat after image segmentation.Those of H in particular increased sharply by232%,393%and223%,separately.Regression analysis between CC and rice plant indices of the three cultivars was done separately(Table5).CC showed good exponential relations with above-ground biomass,N content and LAI(Fig.3).Thefitting curve of LYP9deviated quite far from those of NJ46and NJ45,but the curves of NJ45and NJ46were quite close to each other,which is attributed to the difference in plant type between the two rice cultivars,hybrid indica rice LYP9vs.japonica rice NJ45and NJ46(Huang et al.,2008;Zong et al.,2000).4.Discussion4.1.Above-ground biomass,N content and LAI of differentcultivarsCultivar is a considerable factor in the difference of LAI.LYP9 was much bigger than NJ45and NJ46in LAI,mainly because LYP9 is a hybrid indica rice,with a loose shape,wide leaves and large mean tilt angle(Huang et al.,2008;Zong et al.,2000).These char-acteristics of LYP9were presented in the picture in a much bigger coverage than those of the other two cultivars.4.2.Relationships between image feature parameters and crop N statusAccuracy of the image captured with a digital camera is subject to the influence of changing climate factors,like illumination inten-sity(Graeff and Claupein,2003;Pagola et al.,2009).It has always been a challenge to measure colors in the natural environment.A color normally depends on the spectrum of the incident illumina-tion and the reflectance properties of the object surface,as well as potentially on the angles of illumination and viewing,and many other factors.Restoration of the color through a digital camera is also subject to the influences of the sensors,photometric system and processing system of the camera.Rorie et al.(2011)calibrated the pictures with the standard color card in photographing,but failed to compare the images before and after the calibration.It is,Y.Wang et al./Field Crops Research149(2013)33–3937Fig.2.Relationships of GMR(the difference between green channel and red channel)and G/R(greenness intensity divided by redness intensity)against(a and d)above-ground biomass,(b and e)nitrogen content and(c and f)leaf area index,respectively,for the rice cultivars Nanjing46(NJ46),Nanjing45(NJ45)and Liangyoupeijiu(LYP9),fitted with Eq.(5).therefore,hard to evaluate the accuracy of the image restored with this calibration method.Segmentation of images changed the correlation coefficients of feature parameters with above-ground biomass,N content and LAI to a varying degree.In the segmented images,correlation coefficients of GMR,G/R and NGI with above-ground biomass,N content and LAI were significantly reduced(Tables3and4),which was because in a segmented image,only the part of rice canopy stood out,making the color of the image simple,and in turn narrow-ing the variation ranges of parameters.In this case,the influences of climate factors,like illuminance,would be amplified.But on the contrary,in an intact image,the rice canopy was only a part of the image,the variation ranges of parameters were wider,and the calculation of parameters used the average of all pixels,thus per-mitting relatively larger errors and having a certain buffer capacity to influencing factors.Before the image segmentation,thefitting curves of the three cultivars for GMR and rice plant indices were quite close to each other(Fig.2).There was no significant difference between varieties,especially in the japonica and indica rice.Therefore,GMR would be a quite universal indicator applying to N diagnosis.But when it comes to the CC after image segmentation,thefitting curves of NJ45and NJ46were quite far from the LYP9(Fig.3),which is attributed to the different plant types between the two rice sub-species(Huang et al.,2008;Zong et al.,2000).For this reason,the use of CC to evaluate N status of japonica and indica rice needs different parameters.Canopy cover has an exponential relationship with above-ground biomass,N content and LAI(Li et al.,2010),which tallies with the actual growing process of rice.GMR,G/R and NGI extracted from the intact images are also in exponential relationship with the three plant indices(Fig.2),which is mainly because intact images contain two parts,rice canopy and soil,and the calcula-tion of parameters uses the average of all pixels in an image.As the canopy coverage increases,the proportion of the canopy in the whole image increases,and reflected in the feature parame-ters,thus making the feature parameters of the whole image closer to the mean value of the canopy area.As a result,GMR,G/R and NGI38Y.Wang et al./Field Crops Research149(2013)33–39Fig.3.Relationships of canopy cover(CC)against(a)above-ground biomass,(b)N content and(c)leaf area index for the rice cultivars Nanjing46(NJ46),Nanjing45 (NJ45)and Liangyoupeijiu(LYP9),fitted with Eq.(5).show a variaion trend the same as the CC does,and also similar to the senescence of plants,Adamsen et al.(1999)demonstrated with the G/R in the intact images.Although GMR,G/R and CC are well reflected the N status of rice, their application is subject to limitations.The CC ranges from0to 1theoretically.When the canopy of rice approaches full coverage, CC in the image would no longer increase,but the above-ground biomass,N content and LAI keep on increasing(Fig.3)and at this time CC will not truly reflect N status of the crop.GMR and G/R extracted from the intact images performed similarly to CC.When the parameters approach saturation(about30for GMR and1.4for G/R in Fig.2),the three plant indices of rice have not reached their maximum yet.Therefore,assessment of N status using these feature parameters can only be done before the canopy reaches saturation in coverage.The canopy,however,is far from saturation at thefinal top dressing,so it is still applicable in most cases.4.3.Image segmentation methodPrevious researchers did not pay attention to the research on how to segment images,and most of them adopted different meth-ods in image segmentation.Adamsen et al.(1999)did not perform any segmentation of images in their study on parameter G/R. Researches by Jia et al.(2009)and Kawashima and Nakatani(1998) used Photoshop manually to extract canopy images,which,though rather illustrative,is very time-consuming and may introduce man-made errors.Li et al.(2010)counted pixels with SAVI Green>0as canopy in the image,the image segmentation method is quick,but parameters in the equation need adjusting in light of images.On the basis of previous researches,this study puts forth a new method for segmenting images of rice canopy by setting threshold value based on the magnitude and distribution of GMR.This method is simple and easy to apply.It only needs a unified threshold value set for a group of images and is also very quick in data processing(less than1s to segment a15mega pixels image on a desktop com-puter,3.4GHz CPU).Moreover,this method is applicable to extract canopies of most other green plants.When the canopy coverage of a crop is quite high,the CC value extracted with this method is often slightly lower than the actual one(Fig.3),which is mainly because the lower part of the canopy is growing in the shade of the upper part,thus making the lower part darker in the image and lower in GMR value.When its GMR is low-ered below the threshold,this part will be regarded as background for deletion.Therefore,the extracted CC deviates somewhat from its actual value,but this deviation will not significantly affect the relationship of CC with crop N status.5.ConclusionsThe image segmentation method used in this study is simple in operation,rapid in data processing and also applicable to seg-ment canopies of other green plants.The feature parameters,either before or after image segmentation,are closely related to rice N sta-tus.GMR and G/R extracted from intact images and CC extracted from segmented images are all in exponential relationship with above-ground biomass,N content and LAI of rice(mean R2=0.83). However,the use of CC to evaluate N status of japonica and indica rice needs different parameters.The use of GMR,G/R and CC to diagnose N status of rice has to be done before the canopy reaches its saturation in coverage.Once the canopy gets saturated,these parameters no longer reflect the plant growth conditions accurately.However,the last fertilization of rice cultivation is typically done before earing stage,and the canopy is far from reaching its saturation in coverage at this time. AcknowledgmentsThis research was supported by the Knowledge Innovation Pro-gram of the Chinese Academy of Sciences(KSCX-YW-440)and the Agricultural Science and Technology Innovation Foundation of Jiangsu Province(CX(12)1002).We wish to thank the editor and two anonymous reviewers for their very helpful comments on ear-lier drafts.ReferencesAdamsen,F.J.,Pinter Jr.,P.J.,Barnes,E.M.,LaMorte,R.L.,Wall,G.W.,Leavitt,S.W., Kimball,B.A.,1999.Measuring wheat senescence with a digital camera.Crop Sci.39,719–724.。