Spectral indices measured with proximal sensing using canopy reflectance sensor,chlorophyll meter and leaf color chart for in-season grain yield prediction of basmati rice

VARINDERPAL-SINGH,KUNAL,MEHTAB-SINGH and BIJAY-SINGH

1Department of Soil Science,Punjab Agricultural University,Ludhiana 141004(India)

2Shree Guru Gobind Singh Tricentenary(SGT)University,Chandu-Budhera Road,Gurugram,Haryana 122505(India)

3College of Agriculture,Punjab Agricultural University,Ludhiana 141004(India)

ABSTRACT A number of optical sensing tools are now available and can potentially be used for refining need-based fertilizer nitrogen(N)topdressing decisions.Algorithms for estimating field-specific fertilizer N needs are based on predictions of yield made while the crops are still growing in the field.The present study was conducted to establish and validate yield prediction models using spectral indices measured with proximal sensing using GreenSeeker canopy reflectance sensor,soil and plant analyzer development(SPAD)chlorophyll meter,and two different types of leaf color charts(LCCs)for five basmati rice genotypes across different growth stages.Regression analysis was performed using normalized difference vegetation index(NDVI)recorded with GreenSeeker sensor and leaf greenness indices measured with SPAD meter and LCCs developed by Punjab Agricultural University,Ludhiana(India)(PAU-LCC)and the International Rice Research Institute,Philippines(IRRI-LCC).The exponential relationship between NDVI and grain yield exhibited the highest coefficient of determination(R2)and minimum normalized root mean square error(NRMSE)at panicle initiation stage and explained 38.3%–76.4%variation in yield using genotype-specific models.Spectral indices pooled for different genotypes were closely related to grain yield at all growth stages and explained 53.4%–57.2%variation in grain yield.Normalizing different spectral indices with cumulative growing degree days(CGDD)decreased the accuracy of yield prediction.Normalization with days after transplanting(DAT),however,did not reduce or improve the predictability of yield.The ability of each model to predict grain yield was validated with an independent dataset collected from two experiments conducted at different sites with a range of fertilizer N doses.The NDVI-based genotype-specific models exhibited a robust linear correlation(R2=0.77,NRMSE=7.37%,n=180)between observed and predicted grain yields only at 35 DAT(i.e.,panicle initiation stage),while the SPAD,PAU-LCC,and IRRI-LCC consistently provided reliable predictions(with respective R2 of 0.63,0.60,and 0.53 and NRMSE of 10%,10%,and 13.6%)even with genotype invariant models with 900 data points obtained at different growth stages.The study revealed that unnormalized values of spectral indices,namely NDVI,SPAD,PAU-LCC,and IRRI-LCC,can be satisfactorily used for in-season estimation of grain yield for basmati rice.As LCCs are very economical compared with chlorophyll meters and canopy reflectance sensors,they can be preferably used by small farmers in developing countries.

KeyWords:GreenSeeker canopy reflectance sensor,IRRI-LCC,need-based nutrient management,nitrogen,optical sensing tools,PAU-LCC,SPAD meter

INTRODUCTION

Basmati rice,also known as‘scented pearls’,is a premium rice with excellent aroma and extraordinary cooking and organoleptic qualities,which makes it a delicacy among world foods.The best basmati rice is produced in the northwestern foothills of the Himalayas in the Indian subcontinent.India exported 4.4 million tons of basmati rice to the world in 2019–2020(APEDA,2020).Nitrogen(N)is the principal plant nutrient required for obtaining desired yields of basmati rice.Farmers apply fertilizer N topdressings based on visual estimation of leaf greenness.Using lush green color as the threshold,they often end up applying fertilizer N in excess of the crop requirements.In recent years,field-specific and need-based fertilizer N management has been introduced for crops such as rice,wheat,and maize using optical sensing tools such as GreenSeeker crop reflectance sensor,soil and plant analyzer development(SPAD)chlorophyll meter,and leaf color chart(LCC).Based on in-season measurements of leaves and crop canopies,these tools estimate the fertilizer N requirement of the crop to be top-dressed(Varinderpal-Singhet al.,2010;Swarbrecket al.,2019;Bijay-Singh and Ali,2020;Bijay-Singhet al.,2020).

Increased accuracy in grain yield prediction can further refine need-based N fertilizer topdressing decisions such that grain yield can be sustained and better fertilizer N use efficiencies can be achieved.Several researchers have attempted to predict crop yield at maturity using in-seasonmeasurements in different crops(Tealet al.,2006;Aliet al.,2014,2020;Tagarakis and Ketterings,2017;Tagarakiset al.,2017;Bijay-Singh and Ali,2020).Canopy reflectance data recorded with GreenSeeker explained 83% of grain yield prediction of winter wheat(Raunet al.,2001),77%of corn(Tealet al.,2006),63%of dry direct-seeded rice(DSR)(Aliet al.,2014),and 33%of sorghum(Tagarakiset al.,2017).Leaf greenness intensity recorded with a chlorophyll meter explained 82%of the variation in grain yield prediction of wet DSR(Rameshet al.,2002)and 53%of the variation of winter wheat(Aliet al.,2020).

Spectral indices such as the normalized difference vegetation index(NDVI)values,SPAD readings,and LCC scores,when normalized with cumulative growing degree days(CGDD)or number of days from transplanting or time of sowing of the crop(when growing degree days(GDD)>0),may further improve the prediction of yield at maturity.For example,improved grain yield predictions were observed when NDVI values were normalized with CGDD of corn(Tagarakis and Ketterings,2017),cabbage(Jiet al.,2017),and sugarcane(Loftonet al.,2012).However,grain yield prediction did not improve for dry DSR(Aliet al.,2014),sorghum(Tagarakiset al.,2017),or corn(Tealet al.,2006).Normalizing LCC scores with CGDD improved grain yield prediction for dry DSR(Aliet al.,2014).Data normalized with days after transplantation(DAT)improved theR2from 0.65 to 0.75 for corn(Tagarakis and Ketterings,2017)and from 0.44 to 0.67 for sorghum(Tagarakiset al.,2017),but did not improve the accuracy of yield prediction for cabbage(Jiet al.,2017).

Studies on basmati rice grain yield predictions from inseason measurements of spectral indices are not yet available.We conducted field experiments for three years with the following objectives:i)to develop regression models accurately representing the relationships between in-season measurements of different spectral indices using GreenSeeker canopy reflectance sensor,SPAD 502 chlorophyll meter,and two types of LCCs(one developed by Punjab Agricultural University(PAU),Ludhiana,India(PAU-LCC)and the other by the International Rice Research Institute,Philippines(IRRI-LCC))and the grain yield of different basmati rice cultivars at maturity,ii)to determine the optimal crop growth stage for in-season spectral indexmeasurements for accurate prediction of basmati rice grain yield,and iii)to validate the grain yield prediction equations using datasets from independent field experiments.

MATERIALS AND METHODS

Experimental sites and soil characteristics

Two field experiments were conducted during the summer season(Kharif crops)of 2016,2017,and 2018 at the research farm of PAU,Ludhiana,India.The research farm is located in a semi-arid subtropical region that received,on average,459 mm of rainfall per crop growth season in the three years.Average minimum and maximum temperatures of 26.5 and 34.9◦C,respectively,were recorded during the experimental period.Experiment 1 was conducted at site 1 during 2016 and 2017,while experiment 2 was conducted at sites 1 and 2 during 2018.

Before transplanting the basmati rice seedlings,surface(0–15 cm)soil samples were collected and air-dried for analysis of their physical and chemical properties.The soil at experimental site 1 was sandy loam with a pH of 7.2(1:2 soil/water suspension),4.1 g kg−1organic carbon(Walkley and Black,1934),0.19 dS m−1electrical conductivity(EC;1:2 soil/water suspension),107.9 kg ha−1ammonium acetate-extractable potassium(Merwin and Peech,1951),and 16.8 kg ha−1sodium bicarbonate-extractable phosphorus(Olsenet al.,1954).The diethylenetriaminepentaacetic acid(DTPA)-extractable micronutrients(Lindsay and Norvell,1978)were:0.72 mg kg−1zinc,4.86 mg kg−1iron,0.52 mg kg−1copper,and 3.90 mg kg−1manganese.The soil at experimental site 2 was sandy loam in texture with a pH of 7.4,3.9 g kg−1organic carbon,0.20 dS m−1EC,104.5 kg ha−1ammonium acetate-extractable potassium,17.1 kg ha−1sodium bicarbonate-extractable phosphorus,and DTPAextractable nutrients of 0.63 mg kg−1zinc,4.64 mg kg−1iron,0.47 mg kg−1copper,and 4.11 mg kg−1manganese.

Experimental design,treatments,and crop management

The study involved two sets of experiments.In experiment 1,seven fertilizer N levels(0,10,20,30,40,50,and 60 kg N ha−1)were applied to establish plots with variable soil N supply.The treatments were replicated three times and,in each case,fertilizer N was applied in the form of urea at 14 DAT.The experiment was laid out in a split-plot design with fertilizer N treatments as the subplots and three basmati rice genotypes,CSR30,Punjab Basmati 4(PB4),and Punjab Basmati 5(PB5),in 2016 and five genotypes,CSR30,PB4,PB5,PUSA1637,and PUSA1718,in 2017 as the main plots.

Experiment 2,conducted during 2018 at site 1,was laid out in a split-plot design.The subplots consisted of four N application levels(0,20,40,and 60 kg N ha−1)at predetermined times and two PAU-LCC-guided,needbased fertilizer N applications.The main plots were five basmati rice genotypes(CSR30,PB4,PB5,PUSA1637,and PUSA1718).Experiment 2 was also conducted at site 2,differing only in the four predetermined N application levels,which were 0,25,50,and 75 kg N ha−1.The basmati rice cultivar CSR30 is a tall,photoperiod-sensitive genotype possessing extra-long slender grains that yields about 3.4 t ha−1,whereas PB4,PB5,PUSA1637,and PUSA1718 aresemi-dwarf genotypes having resistance to many pathotypes of bacterial leaf blight,and yield between 3.8 and 4.4 t ha−1.While CSR30 shows a grain yield response up to 20 kg fertilizer N ha−1,PB4,PB5,PUSA1637,and PUSA1718 respond up to 40 kg fertilizer N ha−1(PAU,1979).Therefore,the PB4,PB5,PUSA1637,and PUSA1718 genotypes were grouped and are referred to as the PB-PUSA group.

Seedlings of all the genotypes were transplanted into well-puddled plots at a spacing of 20 cm×15 cm in all the experiments.In 2016,all basmati rice seedlings were transplanted on July 15.In 2017,seedlings of genotype CSR30 were transplanted on July 25 and those of the PBPUSA group on July 6.In experiment 2,conducted in 2018,all genotypes were transplanted on July 15 at site 1 and on July 25 at site 2.In all the experiments,the water was kept standing for 14 d after transplanting,followed by irrigation applied 2 d after the infiltration of the ponded water into the soil.Subsequently,irrigation was applied to avoid water stress,depending on the rainfall experienced.Weeds were controlled by applying the pre-emergence herbicide Machete 50 EC(butachlor).To prevent zinc deficiency disorders,100 kg zinc sulphate heptahydrate(21%zinc)per hectare was applied at the time of puddling(PAU,1979).

At maturity,the basmati rice was manually harvested.Grain and straw yield data for all the experiments were collected from a net area of 12.8 m2in the center of each plot.Samples of both grain and straw were collected and oven-dried at 65◦C to estimate the moisture content.Straw yields were expressed on an oven-dry basis,and grain yields were adjusted to 14%moisture content.

Measurement of spectral indices

The GreenSeeker model 505 canopy reflectance sensor(NTech Industries Inc.,USA),Minolta SPAD-502 chlorophyll meter(SPAD 502 plus,Konika Minolta®Inc.,Japan),PAU-LCC,and IRRI-LCC were used to obtain and evaluate different in-season spectral indices.Fully opened leaves from the top of ten randomly selected normal plants were used to record LCC and SPAD readings.The color of the middle part of a leaf was matched with LCC panels by placing the adaxial surface of the leaf on LCC(Varinderpal-Singhet al.,2007).Readings of LCC were taken in the shade to prevent variation due to subtle differences in color caused by direct sunlight.Chlorophyll meter values were recorded by placing the central portion of the indexleaf in the slit of the SPAD meter(Bijay-Singhet al.,2016).While taking SPAD meter readings,wet leaves and tall and widely spaced plants were avoided.The NDVI data obtained using the GreenSeeker canopy reflectance device was recorded at a height of 0.75 m by keeping the sensor perpendicular to the crop row at 0.5 m s−1walking speed(Bijay-Singhet al.,2015).

Statistical analysis

Statistical analyses were performed using the Excel software(Microsoft Corporation,USA)and the Statistical Package for Social Sciences(IBM Corp,USA).Duncan’s multiple range test was performed on grain yield data to analyze the effect of fertilizer N level using SPSS.The grain yield of basmati rice at maturity was regressed,using Excel,against the different spectral indices measured at different crop growth stages with GreenSeeker,SPAD meter,and LCCs.Exponential,polynomial,and power functions were used to develop equations to predict grain yield from the spectral indices measured at different crop growth stages.Prediction equations were also generated using spectral indices normalized with CGDD.The CGDD was calculated as the sum of GDD from the date of transplanting to the sensing time,while GDD(◦C)was calculated as:

whereTmaxandTminare the daily maximum and minimum temperatures(◦C),respectively,and 10 is the base temperature(◦C)required for rice cultivation(Barger,1969).

The NDVICGDD,SPADCGDD,PAU-LCCCGDD,and IRRILCCCGDDwere the normalized values obtained by dividing the respective spectral index values by the CGDD from transplanting to sensing date,while the NDVIDAT,SPADDAT,PAU-LCCDAT,and IRRI-LCCDATwere the normalized values obtained by dividing the respective spectral indexvalues by DAT,i.e.,the number of days on which GDD>0 between transplanting and sensing time(Bijay-Singhet al.,2015).

Prediction equations were validated using a linear regression model between observed and predicted grain yields obtained from the independent field experiment conducted during 2018.Agreement between predicted and observed grain yields was assessed by normalized root mean square error(NRMSE,%),which was calculated as follows(Wallach and Goffinet,1989):

whereYi,,andnrepresent theith observed yield,ith predicted yield,mean of the observed yield values,and total number of samples,respectively.The agreement between predicted and observed grain yields was deemed to be excellent,good,fair,and poor when NRMSE was<10%,10%–20%,20%–30%,and>30%,respectively(Loague and Green,1991).

RESULTS

Basmati rice yield as affected byfertilizer N levels

Basmati rice grain yields from CSR30 and PB-PUSAgroup plants at various fertilizer N levels are shown in Fig.1.Fertilizer N use significantly improved the grain yield of all genotypes.Compared with control(0 kg N ha−1),the average grain yield of CSR30 was 21.2%and 42.8%higher at 10 and 20 kg N ha−1,respectively,whereas the average yields of the PB-PUSA group at N levels of 10,20,30,and 40 kg N ha−1were 14.6%,26.2%,35.9%,and 49.8%higher,respectively.The grain yield of the CSR30 genotype responded to fertilizer N use at only 20 kg N ha−1,whereas the PB-PUSA group responded to an N dose of 40 kg N ha−1.Further increases in fertilizer N levels resulted in significantly lower grain yield in all genotypes.Therefore,spectral indices obtained from the 50 and 60 kg N ha−1treatments were excluded from the regression analysis for grain yield prediction.

Fig.1 Responses of the grain yields of basmati rice genotypes(CSR30 and PB-PUSA group)to fertilizer N levels in the two field experiments conducted at the research farm of Punjab Agricultural University,Ludhiana,India during the summer seasons of 2016,2017,and 2018.The PB-PUSA group included four genotypes:Punjab Basmati 4(PB4),Punjab Basmati 5(PB5),PUSA1637,and PUSA1718.

Identification of the most useful grain yield prediction function

Regression analysis was performed with exponential,polynomial,and power functions to develop in-season grain yield prediction models using spectral indices(NDVI,SPAD,PAU-LCC,and IRRI-LCC)at various growth stages.The regression models were validated using the grain yield data from two independent field experiments.The coefficients of determination(R2)and NRMSE values obtained during the validation of the different regression functions at different growth stages are provided in Table I.Different regression functions provided comparableR2values at different growth stages.However,NRMSE values as high as 75.3%(in polynomial)and 59.1%(in power)led to poor validation with these models at certain growth stages.The exponential function consistently gave comparable values ofR2with lower values of NRMSE at different growth stages and across all genotypes.Thus,the exponential regression model was used to determine the sensing time for grain yield predictions.

TABLE ICoefficient of determination(R2)and normalized root mean square error(NRMSE)obtained during validation of different regression functions(polynomial,power,and exponential)for developing grain yield prediction models using various spectral indicesa)at different growth stages of basmati rice grown in the two field experiments conducted at the research farm of Punjab Agricultural University,Ludhiana,India during the summer seasons of 2016,2017,and 2018

Time of sensing for making grain yield predictions

The NRMSE andR2values for the exponential regression function for predicting grain yield of basmati rice at maturity from different spectral indices(NDVI,SPAD,PAULCC,and IRRI-LCC)measured at different growth stages are shown in Tables II–V.The NDVI measurements with the GreenSeeker sensor at 35 DAT explained 76.4%and 38.3%of yield variation in genotype CSR30 and the PB-PUSA group of genotypes,respectively,with the highestR2values being 0.764 and 0.383,respectively(Table II,Fig.2)and the lowest NRMSE values being 7.08%and 14.5%,respectively(Table II).TheR2values at the other growth stages were 0.069(21 DAT)and 0.478(49 DAT)for CSR30,and 0.281(21 DAT)and 0.089(49 DAT)for the PB-PUSA group,with relatively higher NRMSE values than at 35 DAT.Pooling the data across different growth stages and genotypes resulted in a very poor correlation(R2=0.042,Table II).

Fig.2 Coefficient of determination(R2)for the exponential relationships between basmati rice grain yield and spectral indices at different days after transplanting(DAT)for the CSR30 genotype and the PB-PUSA group of genotypes.NDVI=normalized difference vegetation index;SPAD=leaf greenness indexmeasured with an SPAD chlorophyll meter;PAU-LCC=reading of the leaf color chart developed by Punjab Agricultural University,Ludhiana,India;IRRI-LCC=reading of the leaf color chart developed by the International Rice Research Institute,Philippines.The PB-PUSA group of genotypes included Punjab Basmati 4(PB4),Punjab Basmati 5(PB5),PUSA1637,and PUSA1718.

TABLE IIRegression coefficients(a and b),coefficient of determination(R2),and normalized root mean square error(NRMSE)for the exponential relationship(y=aebx)between grain yield(y)and the spectral indexnormalized difference vegetation index(NDVI)and its normalized values(x)recorded at different growth stages of basmati rice grown in the two field experiments conducted at the research farm of Punjab Agricultural University,Ludhiana,India during the summer seasons of 2016,2017,and 2018

TABLE IIIRegression coefficients(a and b),coefficient of determination(R2),and normalized root mean square error(NRMSE)for the exponential relationship(y=aebx)between grain yield(y)and the spectral indexleaf greenness indexmeasured with an SPAD chlorophyll meter(SPAD)and its normalized values(x)recorded at different growth stages of basmati rice grown in the two field experiments conducted at the research farm of Punjab Agricultural University,Ludhiana,India during the summer seasons of 2016,2017,and 2018

TABLE IVRegression coefficients(a and b),coefficient of determination(R2),and normalized root mean square error(NRMSE)for the exponential relationship(y=aebx)between grain yield(y)and the spectral indexreading of the leaf color chart(LCC)developed by Punjab Agricultural University(PAU),Ludhiana,India(PAU-LCC)and its normalized values(x)recorded at different growth stages of basmati rice grown in the two field experiments conducted at the research farm of Punjab Agricultural University,Ludhiana,India during the summer seasons of 2016,2017,and 2018

TABLE VRegression coefficients(a and b),coefficient of determination(R2),and normalized root mean square error(NRMSE)for the exponential relationship(y=aebx)between grain yield(y)and the spectral indexreading of the leaf color chart(LCC)developed by the International Rice Research Institute(IRRI),Philippines(IRRI-LCC)and its normalized values(x)recorded at different growth stages of basmati rice grown in the two field experiments conducted at the research farm of Punjab Agricultural University,Ludhiana,India during the summer season of 2016,2017,and 2018

The SPAD readings exhibited consistently high correlations with grain yield for CSR30(R2=0.661–0.715)and the PB-PUSA(R2=0.402–0.529)genotypes across different growth stages(Table III).TheR2for the relationship of PAU-LCC(Table IV)with grain yield varied from 0.628 to 0.665 for CSR30 and from 0.410 to 0.533 for the PB-PUSA genotypes,while theR2for the relationship of IRRI-LCC(Table V)with grain yield varied from 0.625 to 0.680 for CSR30 and from 0.373 to 0.424 for the PB-PUSA genotypes,thereby exhibiting consistent correlations at different growth stages.Data pooled across genotypes and different growth stages explained 57.1%,57.2%,and 53.4%of the variation in grain yield with NRMSE values of 12.3%,12.3%,and 12.8%for SPAD,PAU-LCC,and IRRI-LCC,respectively(Tables III–V).

Normalized spectral indices and grain yield predictions

The values of NDVI,SPAD,PAU-LCC,and IRRI-LCC were normalized with CGDD and DAT to test the hypothesis that results would be improved by removing environmental influences that may occur due to variations in weather conditions during different years.Normalizing the NDVI values with CGDD(NDVICGDD)reduced theR2values of the exponential equations for predicting basmati rice grain yield to 0.029,0.748,and 0.439 for CSR30 and to 0.089,0.194,and 0.014 for PB-PUSA for measurements made at 21,35,and 49 DAT,respectively(Table II).TheR2in the case of NDVICGDDfor data pooled across the growth stages and genotypes was as low as 0.005(Table II).Similarly,the decreasedR2and increased NRMSE values for the relationships between SPAD,PAU-LCC,and IRRI-LCC measurements and basmati rice grain yield after adjusting for CGDD(Tables III–V)indicate reduced accuracy of inseason grain yield prediction.However,when NDVI,SPAD,PAU-LCC,and IRRI-LCC were normalized with DAT,theR2values remained the same and the accuracy of grain yield prediction was not thereby improved(Tables II–V).

Goodness of fit of grain yield prediction models

The reliability of the prediction models was validated with an independent dataset recorded from two different experiments with different fertilizer N management treatments.The observed grain yield in each experimental treatment was plotted against the grain yield predicted from NDVI,SPAD,PAU-LCC,and IRRI-LCC data recorded at different growth stages using genotype-and growth stage-specific exponential prediction equations(Fig.3).The degree of agreement between the predicted and observed values was determined usingR2and NRMSE of the linear regression between the two.Validation of grain yield predicted with pooled NDVI data recorded at different growth stages led to unfavorable NRMSE(27.0%)and poorR2(0.069)for the linear correlation between observed and predicted grain yields(Fig.3a).However,pooled data points pertaining to observed grain yield and those predicted with SPAD-,PAU-LCC-,and IRRI-LCC-based prediction equations were evenly scattered around the 1:1 line.Linear relationships between the observed and predicted yields with equations based on SPAD(R2=0.625,NRMSE=10.0%),PAU-LCC(R2=0.602,NRMSE=10.0%),and IRRI-LCC(R2=0.528,NRMSE=13.6%)registered acceptable predictions(Fig.3b–d).Separate validation of the predicted yields based on NDVI data recorded at 35 DAT improved the linear correlation(R2=0.773,NRMSE=7.37%),which provided the best predictions(Fig.3e).

Fig.3 Relationships between observed grain yields and grain yields predicted using the NDVI(a),SPAD(b),PAU-LCC(c),and IRRI-LCC(d)regression models with data pooled across growth stages and using the NDVI regression model with data recorded at day 35 after transplantation(e).NDVI=normalized difference vegetation index;SPAD=leaf greenness indexmeasured with an SPAD chlorophyll meter;PAU-LCC=reading of the leaf color chart developed by Punjab Agricultural University,Ludhiana,India;IRRI-LCC=reading of the leaf color chart developed by the International Rice Research Institute,Philippines.R2=coefficient of determination;NRMSE=normalized root mean square error.The solid lines denote the linear regressions and the dashed lines are the 1:1 lines.

DISCUSSION

In-season estimation of basmati rice grain yield

The maximum grain yield of CSR30 was obtained at a fertilizer N application rate of 20 kg N ha−1,whereas the PB-PUSA genotypes responded to a double N dose(40 kg N ha−1)(Fig.1).Further increases in N level decreased the grain yield by 21.3%and 12.9%for CSR30 and PB-PUSA genotypes,respectively.Hence,excessive N application leads to yield loss,in addition to the escape of reactive N fromcrop lands to water bodies and the atmosphere(Bhatiaet al.,2012;Varinderpal-Singhet al.,2021).

Canopy reflectance estimates of biomass(NDVI)account for temporal and spatial variability and can predict grain yield accurately(Harrellet al.,2011),which would provide substantial information for making N management decisions.The NDVI,NDVICGDD,and NDVIDATvalues were first used to predict grain yield from data pooled for years,genotypes,and growth stages,and then separately for individual growth stage and different sets of genotypes(Table II).The very lowR2values for prediction equations based on NDVI measurements(pooled across the genotypes,growth stages,and years)and grain yields revealed that reflectance data cannot be pooled across growth stages and genotypes,which vary in their response to N fertilizer.The NDVI values are specific to genotypes and growing environments(Varinderpal-Singhet al.,2021).

The best exponential relationship between grain yield and NDVI values was obtained at 35 DAT(the panicle initiation stage in basmati rice),better than those at early(21 DAT,active tillering stage)and late(49 DAT,heading stage)growth stages.Aliet al.(2014)and Harrellet al.(2011)also noticed that at early and late growth stages,sensing is not reliable for grain yield prediction for rice.The weak relationships at the early growth stage can be attributed to slow growth,lower N uptake,and the strong influences of water and soil background on canopy reflectance(Gnypet al.,2014),while the saturation of NDVI due to canopy closure,which influences the field of view of the sensor,leads to poor correlations at later growth stages(Tealet al.,2006;Harrellet al.,2011).The NDVI considers both the greenness and the biomass of the crop simultaneously.However,a potential factor for its weaker relationship was its failure to show differentiation between the treatments with variable fertilizer N,as biomass measurements act as a limiting factor at both early and later growth stages of the crop.Hence,NDVI failed to sustain a strong correlation with grain yield throughout the growth stages.Therefore,growth stage is a major and decisive factor in determining the potential of NDVI data in predicting grain yield.The determination coefficient of CSR30(R2=0.764)at 35 DAT was higher than that of the PB-PUSA group(R2=0.383).This might be due to the pooling of NDVI data recorded for different genotypes and the cumulative effect of the individualR2values of each genotype in the PB-PUSA group on the finalR2,as the canopy geometry and NDVI values are genotype-specific.

Spectral measurements with SPAD meters and LCCs have shown great potential for sustaining high yields while defining low N optimum doses in rice(Varinderpal-Singhet al.,2007;Bijay-Singhet al.,2016;Swarbrecket al.,2019).Aliet al.(2014)attempted to predict the grain yield of dry DSR using SPAD and LCC.Unlike canopy reflectance data,spectral measurements with SPAD,PAU-LCC,and IRRILCC consistently provided better correlations with results pooled across years,genotypes,and growth stages,as well as at individual growth stages in CSR30 and PB-PUSA(Tables III–V,Fig.3b–d).Leaf greenness measured with SPAD,PAU-LCC,or IRRI-LCC,as an indexof leaf chlorophyll content,exhibited consistently better correlations with basmati rice grain yield than canopy reflectance data,indicating its ability to differentiate fertilizer N treatments and thus plant N requirements,irrespective of the genotype and growth stage.The PAU-LCC provided results comparable with those from the expensive SPAD meters and performed better than IRRI-LCC in both the CSR30 and PB-PUSA genotypes.In addition,PAU-LCC measured leaf greenness more precisely than IRRI-LCC,as it can differentiate between leaves with SPAD variations of 5 units,whereas the leaf greenness discrimination level of IRRI-LCC is at least 10 SPAD units.Nevertheless,IRRI-LCC provided acceptable correlations with grain yield.The SPAD,PAU-LCC,and IRRI-LCC data recorded at different growth stages can thus be used for accurate in-season yield prediction using exponential equations in basmati rice.Therefore,achieving valid results with SPAD meters and LCCs does not require definition of the growth stage,and these tools can be used throughout the growing period to support need-based fertilizer N topdressing decisions.

Adjusted yield prediction equations using CGDD and DAT

The NDVICGDD,SPADCGDD,PAU-LCCCGDD,and IRRILCCCGDDvalues were calculated by adjusting the values of all four reflectance indices with CGDD,which measures the degree of warmth the plants received during the growing season.An additional index,DAT,was used as the divisor of NDVI,SPAD,PAU-LCC,and IRRI-LCC values.Some studies have reported improvements in grain yield prediction when NDVI and LCC values were normalized with CGDD(Aliet al.,2014;Tagarakis and Ketterings,2017)and DAT(Tagarakiset al.,2017).However,in this study,the accuracy of in-season yield prediction was inadequate when spectral indices measured with the GreenSeeker sensor,SPAD,PAU-LCC,and IRRI-LCC were normalized with CGDD;prediction validity was not influenced when normalization was performed with DAT(Tables II–V).Normalizing the NDVI values with CGDD decreased the accuracy of inseason yield estimation in sorghum(Tagarakiset al.,2017)and transplanted rice(Harrellet al.,2011),while adjusting the NDVI values with DFP(days from transplanting when GDD>0)led to similarR2values as with normal NDVI values in cabbage(Jiet al.,2017).Therefore,CGDD and DAT did not provide consistently reliable normalization of spectral indices and proximal sensing data for making grain yield predictions.

Comparison of various grain yield prediction functions

Regression analysis was performed with exponential,polynomial,and power functions to find the best function for making in-season grain yield predictions using spectral indices measured with GreenSeeker,SPAD meter,PAULCC,and IRRI-LCC at various growth stages.Prediction equations based on all spectral indices generally explained similar variation pattern inR2at different growth stages,but polynomial and power functions showed unexpected surges in NRMSE values for data recorded at certain growth stages.HigherR2and lower NRMSE were consistently observed for exponential prediction equations at all growth stages.Therefore,exponential equation was selected as a more appropriate equation for grain yield prediction than power and polynomial equations.Thenkabailet al.(2000)also observed that exponential regression models were better models for describing variation in crop biophysical parameters and spectral vegetative indices in many crops.Furthermore,the exponential regression model has been the best option in studies in which the crop growth rate at a particular time is proportional to the remaining crop growth(Kutneret al.,2004).Therefore,exponential equations generated using NDVI data recorded at 35 DAT,and SPAD,PAU-LCC,and IRRI-LCC values recorded at different growth stages,were used to make grain yield predictions.

Validation of grain yield prediction models

Validation of the prediction models was performed with an independent dataset obtained from two different experiments.As regards NDVI,poor agreement(NRMSE of 27.0%)was obtained between observed and predicted yields when the data were pooled across genotypes,growth stages,and years(Fig.3a).However,yield predictions based on NDVI measurements made at 35 DAT showed low NRMSE values(7.37%),indicating excellent agreement between predicted and observed values and validating the developed regression function for predicting grain yield with 77.3%variability(Fig.3e).However,good agreement(10%

CONCLUSIONS

Genotype and time of sensing greatly impact the accu-racy of yield predictions using functions based on NDVI.Optical sensing data recorded at 35 DAT(approximating the panicle initiation stage)provided better predictions(higherR2and lower NRMSE)using genotype-specific regression models for the CSR30 and PB-PUSA group genotypes.Measurements of NDVI correlated poorly with grain yield at other growth stages due to its failure to differentiate between treatments with variable fertilizer N application rates.The grain yield predictions based on NDVI data pooled for genotypes and/or growth stages correlated poorly with the observed values.However,SPAD,PAU-LCC,and IRRI-LCC could accurately predict in-season grain yield from the equations obtained by pooling data across different genotypes,years,and growth stages because,unlike NDVI,they could differentiate between treatments with variable fertilizer N rates at all growth stages.The use of CGDD and DAT to normalize the yield prediction equations of NDVI,SPAD,PAU-LCC,and IRRI-LCC did not improve the accuracy of yield prediction for basmati rice.Unaltered NDVI,SPAD,PAU-LCC,and IRRI-LCC values thus offer a promising solution for predicting end-of-season grain yield of basmati rice.We therefore advocate the use of spectral indices for developing need-based fertilizer N use strategies to ensure optimum yield of basmati rice and to achieve better agronomic and recovery efficiencies of applied fertilizer N.The variety-specific algorithms generated may be helpful for the plant breeders to predict in-season grain yield of basmati rice,provided that data are recorded under different agroclimatic conditions with diverse groups of genotypes that can be further processed to achieve robust algorithms.

ACKNOWLEDGEMENT

The research was funded by the Department of Biotechnology(DBT),Government of India(No.BT/IN/UKVNC/42/RG/2014-15)and the Biotechnology and Biological Sciences Research Council(BBSRC)under the international multi-institutional collaborative research project entitled Cambridge-India Network for Translational Research in Nitrogen(CINTRIN)(No.BB/N013441/1).