Estimation of spectral responses and chlorophyll based on growth stage effects explored by machine learning methods

Dehu Go ,Lng Qio ,Lulu An ,Ruomei Zho ,Hong Sun,c,* ,Minzn Li,c ,Weijie Tng,Nn Wng

a Key Laboratory of Modern Precision Agriculture System Integration Research,Ministry of Education,China Agricultural University,Beijing 100083,China

b Key Laboratory of Agricultural Information Acquisition Technology,Ministry of Agriculture and Rural Affairs,China Agricultural University,Beijing 100083,China

c Yantai Institute of China Agricultural University,Yantai 264670,Shandong,China

Keywords:Wheat chlorophyll content Growth stage effects Sensitive wavelengths Spectral response Random forest

ABSTRACT Estimation of leaf chlorophyll content(LCC)by proximal sensing is an important tool for photosynthesis evaluation in high-throughput phenotyping.The temporal variability of crop biochemical properties and canopy structure across different growth stages has great impacts on wheat LCC estimation,known as growth stage effects.It will result in the heterogeneity of crop canopy at different growth stages,which would mask subtle spectral response of biochemistry variations.This study aims to explore spectral responses on the growth stage effects and establish LCC models suited for different growth stages.A total number of 864 pairwise samples of wheat canopy spectra and LCC values with 216 observations of each stage were sampled at the tillering,jointing,booting and heading stages in 2021.Firstly,statistical analysis of LCC and spectral response presented different distribution traits and typical spectral variations peak at 470,520 and 680 nm.Correlation analysis between LCC and reflectance showed typical red edge shifts.Secondly,the testing model of partial least square (PLS) established by the entire datasets to validate the predictive performance at each stage yielded poor LCC estimation accuracy.The spectral wavelengths of red edge (RE) and blue edge (BE) shifts and the poor estimation capability motivated us to further explore the growth stage effects by establishing LCC models at respective growth periods.Finally,competitive adaptive reweighted sampling PLS(CARS-PLS),decision tree(DT)and random forest(RF) were used to select sensitive bands and establish LCC models at specific stages.Bayes optimisation was used to tune the hyperparameters of DT and RF regression.The modelling results indicated that CARS-PLS and DT did not extract specific wavelengths that could decrease the influences of growth stage effects.From the RF out-of-bag(OOB)evaluation,the sensitive wavelengths displayed consistent spectral shifts from BE to GP and from RE to RV from tillering to heading stages.Compared with CARS-PLS and DT,results of RF modelling yielded an estimation accuracy with deviation to performance(RPD)of 2.11,2.02,3.21 and 3.02,which can accommodate the growth stage effects.Thus,this study explores spectral response on growth stage effects and provides models for chlorophyll content estimation to satisfy the requirement of high-throughput phenotyping.

1.Introduction

Estimation of leaf chlorophyll content (LCC) by proximal sensing is a critical technique for photosynthesis evaluation in high-throughput phenotyping [1,2].Traditionally,crop biochemical pigments are measured in laboratory following destructive sampling and time-consuming tasks.Timely and non-destructive monitoring of LCC by spectroscopy is an efficient approach for connecting genomics to phenomics.However,spectral reflectance is always affected by the temporal variability of crop biochemical and biophysical properties,especially referring to the growth periods,so it interferes with wheat (Triticum aestivum L.) LCC estimation at different growth stages.Thus,studies should focus on the exploration of spectral responses on growth stage effects and improvement of crop LCC estimation for in-field wheat.

According to the spectroscopy theory of Kubelka-Munk (K-M),spectral reflection has a close relationship to the biochemical concentration of LCC under a specific measurement of optical distance,in which the ideal condition of K-M theory is based on assumptions of surface homogeneity.The heterogeneity of crop canopy always exists and changes greatly across different growth stages,referred to as growth stage effects,which makes crop estimation challenging.Growth stage effects are defined as changing biophysical traits across different growth stages that would mask spectral subtle differences in biochemical variations[3].Prey et al.[4]suggested that spectral response variations,such as amplitude variations and spectral position shifts,were presented following different growth variations.Feng et al.[5]also indicated that the predictive capacity of different vegetation indices was highly correlated with the spectral development traits of specific growth stages.Recent evidence suggested that crop variables,such as plant nitrogen (PNC),could be predicted accurately at later growth stages,while the estimation results are quite poor at early stages before canopy closure[6].Ravier et al.[7] indicated that establishing wheat LCC models considering each growth stage is better rather than a single overall model.Sembiring et al.[8] reported great challenges to overcome the influences of growth stage effects and suggested the need to further calibrate accurate LCC models at specific stages.In addition,many other researches pointed out that growth stage effects should be explored and solved due to significant impacts on crop LCC estimation,such as applications in paddy rice (Oryza sativa L.) [9,10],potato (Solanum tuberosum L.) [11],maize (Zea mays L.)[6,12] and winter wheat [13,14].

Specific bands used to estimate wheat LCC are correlated to growth stage effects.Differences and commonalities of sensitive wavelengths co-exist at different growth stages.Spectral responses within specific wavelengths have close relationships to pigment photosynthesis,molecule structure and optical plant interaction[15,16].For commonalities across different growth stages,spectral wavelengths within 400-700 nm are strong absorption regions of plant pigments which absorb most of incident radiations [17].The spectral profile of red edge (RE) at 690-780 nm is consistent under different ground coverages and growth stages.The critical point of 730 nm is sensitive to growth stage effects to detect subtle spectral signal from crop pigments [18].For differences in growth stages,green band near 550 nm is sensitive to leaf chlorophyll differences.Peng et al.[19] pointed out that the blue edge (BE),yellow edge (YE) and red edge (RE) can reveal the green vegetation traits at different growth stages.During the early growth stages of open canopy closure,Hatfield and Prueger [20] indicated that blue band near 460 nm combined with red band at 660 nm have great predictive potentials for LCC due to the sensitivity of saturation effects in spectral diagnosis.Thus,the differences and commonalities of sensitive wavelengths at different growth stages should be determined.

To sum up,sensitive wavelength selection and LCC modelling performance are critical to explore the growth stage effects and both processes are always interacted with each other.Two methods can be used to explore variations in wheat LCC estimation.The first method seeks to explicitly establish parametric models by importing more observational variables to explain the contributions of variables on wheat LCC estimation [21-23].The other method adopts advanced non-parametric regression to implicitly express estimation models aiming at timely and accurate detection[24-26].Feature selection methods listed [27] indicated that ensemble wavelengths could improve the estimation accuracy of crop properties.Due to large numbers of dimensions,methods such as multiple linear regression (MLR) or stepwise MLR can induce overfitting for suffering from multicollinearity [28].Partial least square (PLS) regression transforms high-dimensional wavelengths into few latent variables for prevent modelling from overfitting [29].Competitive adaptive reweighted sampling (CARS) is an efficient variable selection method that is widely applied in in-field spectroscopy analysis [30].However,many researchers pointed out the presence of complex non-linear coupling between biochemical pigments and biophysical traits under different growth stages [31].

Machine learning algorithms are used to address the non-linear problems and construct LCC estimation models.Decision tree(DT)operates in a binary tree structure to establish the regression process [32].Random forest is a machine learning method combined with single DT using bagging (bootstrap aggregating) with randomisation to construct ensemble decision learning [33,34].To characterise growth stage effects by specific spectral wavelengths,regardless of biochemical indictors of crop pigments or crop canopy properties,random forest (RF) has great potential to gain better estimation performance and has been used to estimate crop variables of chlorophyll content [35,36],nitrogen [37],and leaf area index [33,38].

This study aims to explore the spectral responses on growth stage effects and the improvement of wheat LCC estimation.A method is proposed to evaluate and reveal growth stage effects,determine sensitive band and improve wheat LCC estimation at different growth stages.This paper is organised in four parts: (1)analysis of wheat LCC and reflectance distributions and mutual relationships at different growth stages;(2)evaluation of the overall modelling performance and validation of the predictive performance at each growth stage;(3) exploration and comparison of sensitive wavelengths and modelling performance of CARS-PLS,DT and RF at different growth stages;and (4) discussion of variations in sensitive wavelengths,modelling performance comparison and potential applications.

2.Materials and methods

2.1.Experimental design and in-field spectral sampling

Experiments were conducted in the field station of Dry Farming and Agriculture Institute,Hebei Academy of Agricultural and Forestry Sciences(HAAFS)located at Hengshui,Hebei province,China.This farming site operates planted maize-wheat intercropping systems for testing growth differences (Fig.1).Wheat cultivar Heng 4399 was planted with a tillering rate of 300 kg ha-1.Fertilisation treatments were combined under six compound fertilizer levels of nitrogen proportions (0,90,180,360,540 and 720 kg ha-1) and phosphorus proportions (0,60,120,240,360,480 kg ha-1) with a ratio of 3:2 and three replicates,abbreviated as L1,L2,L3,L4,L5 and L6.In each plot,buffer strip was set to eliminate the mutual influence of each plot.

In the field experiment,canopy spectral sampling was conducted on March 18,March 30,April 20 and May 10,2021 representing wheat growth stage of tillering,jointing,booting and heading,respectively.Wheat canopy reflectance was measured by an ASD FieldSpec HandHeld 2 portable spectroradiometer(Analytica Spectra Devices.,Inc.)with a spectral range of 325-1075 nm,1 nm sampling interval and a spectral resolution of 3 nm.The reflectance spectra were acquired at 10:00-14:00 in sunny days.A white panel with 98% reflectance was used for radiation correction.

2.2.Wheat LCC extraction

UV1800 was used to extract wheat LCC according to the Lambert-Beer law on Dry Farming and Agriculture Institute,HAAFS,Hengshui,Hebei province,China.Before absorbance determination,we weighed 0.4 g of wheat leaves,cut into 2-4 cm patches,and soaked in 40 mL of 95% ethanol in the dark with shaking in a 12 h interval until fading of the green leaves.Absorbance values at 649 and 665 nm were determined by Lichtenthaler method[39].

Fig.1.Experiment description of georeferenced sampling sites and flowchart of data processing.

where Cais the content of chlorophyll a,mg L-1;Cbis the content of chlorophyll b,mg L-1;Ctis the sum of chlorophyll content,mg L-1;and D649and D665are the relative absorbance(%)at the wavelength of 649 and 665 nm,respectively.

2.3.Flowchart of data processing

The matrix of spectral reflectance(X)means predictor variables,and the response vector of wheat LCC (Y) means response variables.The matrix X was constructed with rows as observation samples and columns as wavelength variables.Firstly,a total of 864 datasets (marked as Stotal) were split into four datasets across the growth stages of tillering,jointing,booting and heading (marked as STillering,SJointing,SBootingand SHeading).Each stage has 216 pairwise observations of reflectance spectra and wheat LCC.Secondly,methods of CARS-PLS,DT and RF were used to select characteristic wavelengths and establish estimation models of wheat LCC at specific growth stage,respectively.Finally,further discussion was conducted to evaluate the corresponding sensitive wavelength variations and the modelling potential across different growth stages.These methods were all implemented in Matlab (2019a),and the flowchart of data processing is illustrated in Fig.1.

2.4.Spectral preprocessing and dataset partition

Multiplicative scattering correction (MSC) was used to reduce spectral variations induced by ambient alterations [40].To ensure that the model calibration set maximises the characterisation of the sample uniform distribution and improves model stability,the dataset partition method of SPXY [41] was used to separate the total dataset into 2/3 as calibration sets and 1/3 as validation sets.The sets of calibration and validation (SC and SV) were described by the number of samples,mean,standard deviation(SD),minimum,maximum,range and subscripts and were marked with ‘total’ and growth stage names.The calibration set of SCtotaland the validation set of SVtotalcorresponded the aggregation calibration of SCtillering,SCjointing,SCbootingand SCheadingand the validation set of SVtillering,SVjointing,SVbootingand SVheading.To validate the existence of growth stage effects,the ensemble model of SCtotalwas calibrated to validate the prediction accuracy of the validation set of SVtillering,SVjointing,SVbootingand SVheading.The LCC statistics of different datasets are illustrated in Table 1.LCC range of the calibration set involves the corresponding validation set,which contributes to the modelling generalisation and the range of SCtotaland SVtotalcontains the range of the four growth stage sets.

Table 1 LCC statistics of calibration and validation datasets at different growth stages (mg L-1).

2.5.Variable selection and modelling based on PLS and CARS-PLS

For handling high-dimensional data,PLS is used to replace the original wavelengths with orthogonal latent variables (nLV) [42].The nLV considers both response variables of wheat LCC and predictor wavelengths to estimate the wheat LCC.Based on the PLS modelling,CARS proposed by Li et al.[43] regards PLS regression coefficients as weights to iteratively select sensitive wavelengths of wheat LCC.Three critical steps are conducted for variables selection and the modelling calibrations.Monte Carlo sampling(MCS)is the first step to carry out subset sampling of sub-models.Then,exponentially deceasing functions (EDF) is used to remove redundant variables with outputs of PLS regression coefficients according to the order.Thus,adapted weighted sampling (ARS) and root mean square errors of cross validation are used to iteratively select optimal sub-models and sensitive wavelengths.

For the RF of each decision tree,the so-called out-of-bag data refer to the data that each time a decision tree is built by repeated sampling for training,while 1/3 of the data is not utilised for the decision tree building.This part of the data (out-of-bag) can be used to evaluate the performance of the decision tree and calculate the prediction error rate of the model,called out-of-bag data error(errOOB1).To maintain modelling comparability,the top 10 bands were selected as sensible variables of wheat LCC.Thus,with consistent number of sensible variables,the sensible wavelength position and modelling performance will be further analysed to explore growth stage effects.

2.6.Decision tree and random forest

2.6.1.Decision tree and random forest

Decision tree operates in a binary tree structure to establish the regression models [32].The DT regression is influenced by hyperparameters of root node,internal branch nodes,terminal node(leaves nodes).The information gain ratios quantified by entropy are used to select the sensitive wavelengths to split the dataset into different values.

Random forest is a machine learning method combined with single DT using bagging strategy (booststrap aggregating) to construct ensemble decision learning.For excellent nonlinear generalisation ability,RF is employed to established LCC estimation due to the simplicity of parameters tuning[44].Random forest regression can simplify models to conduct a tentative operation by permuting OOB errors.Higher importance demonstrates the selected variable have more spectral response information of wheat LCC.In this study,the predictor bands of X have 751 wavelengths,while the spectral responses of chlorophyll absorption position vary across different growth stages.

2.6.2.Parameters tuning of DT and RF

Radom forest tends to over-fitting,but shallow tree tends to under-fitting.Thus,the numbers of observations per leaf node(minLS) and the predictor variables (numPTS) to be sampled should be tuned [45].There are three model-sensitive parameters including numPTS,minLS and splitting depth.Bayes optimisation is used to minimise a scalar objective function in a bounded domain.Mean squared error (MSE) is the main criterion to optimise the number of leaves and trees.The pairwise parameters of numPTS and minLS are 200 and 5.

The parameters of numPTS,minLS and splitting depth from the DT optimisation can be shared with the RF modelling process.Thus,different parameters were used to construct the LCC estimation with hyperparameters of the number of trees (nTree) and the number of variables (nTry).To stabilise the estimation errors,nTree and mTry values were tested to yield the lowest RMSE at different growth stages.The nTree values were set from 100 to 1000,and one-third of input variables were set to minimise the estimation errors[46,47].The sensitive variables and estimation accuracy of wheat LCC were obtained by iterating the RF models 100 times.

2.7.Spectral regions descriptions

The spectral range is divided into different intervals to facilitate the evaluation of spectral sensitive changes at different growth stages.According to the photochemical traits,chlorophyll content leads to stronger reflection in green bands (centred at 550 nm),known as green peak(GP),than red bands(650-690 nm)and blue bands(430-490 nm),and the inflection points before and after the formation of GP curves were designated blue edge(BE)and yellow edge (YE) [48].The strong absorption of red regions is called red valley (RV).The reflectance spikes of red edge (RE) are induced by strong absorption of red bands and strong reflection of near infrared regions.BE,YE and RE constitute ‘three edge’ parameter[19,49,50].The high reflectance of near infrared plateau (NIRP) is caused by complex interactions between incident radiation and leaf internal structures (cell wall,mesophyll and air cavities).The spectral range of 325-1075 nm was separated as Violet blue (VB:325-470 nm),BE (470-530 nm),GP (530-580 nm),YE (580-640 nm),RV (640-680 nm),RE (680-780 nm) and NIRP (780-1075 nm) to describe the sensitive bands at different growth stages.

2.8.Modelling evaluation

The calibrated models were evaluated by comparing the determination coefficients(R2)and root mean square errors(RMSE).The ratio of standard deviation(SD)to RMSE,termed as RPD,was used to evaluate the model predictive accuracy,which was categorized in four levels: poor prediction of category 1 (RPD <1.5),common prediction of category 2 (1.5 ≤RPD<2.0),approximate prediction of category 3(2.0 ≤RPD<3.0)and excellent prediction of category 4 (RPD ≥3.0) [45,51].The univariate linear fitting relationships between measured LCC and predicted LCC values for validation sets were used to assess the estimation accuracy with linear parameters of slopes and biases.The values of R2,RMSE,SD and RPD were calculated with the following equations:

3.Results and discussion

3.1.Statistics and correlation analysis of LCC and spectra

The LCC ranged from 0 to 50 mg L-1and divided into five groups with the interval of 10 mg L-1.The distribution of LCC showed obvious difference in the four growth stages (Fig.2A-D).Correspondingly,reflectance spectra and coefficients of reflectance of the five LCC groups showed different profiles (Fig.2E-H) with the LCC range of 6.45-28.16,9.99-40.11,11.95-40.52 and 8.45-45.44 mg L-1for tillering,jointing,booting,and heading stage,respectively.

From different growth stages,there was an overall tendency for the variation coefficients of reflectance.In light of spectral response in corresponded LCC intervals,the variation coefficient displayed a top-down order of I2 >I1 >I3 (tillering stage),I3>I2>I4(jointing stage),I4>I3>I2(booting stage),and no significant difference(heading stages).The variation coefficients presented three consistent peaks near 385,520 and 682 nm,which may explain the spectral response of wheat LCC.The statistical analysis of both LCC and spectral reflectance reveals that there are significant variations due to growth stage effects.

The correlation coefficients at wavelengths of 325-1075 nm appear relatively large differences at different growth stages.Variations near 970 nm presented water absorption valley[19].The RE ranges represent apparent position shifts at different growth stages.Greater than 740 nm and<740 nm show positive and negative correlations,respectively.The correlation coefficients at the tillering stage presented the second highest absolute values within visual bands but the lowest values in NIR bands.The jointing and booting stages displayed similar tendency for the fluctuations of correlation coefficients.For the heading stage,the correlation coefficients presented the biggest absolute values for both visual and NIR bands.Thus,results of correlation analysis indicate that the spectral response regions sensitive to wheat LCC are located at visual bands especially at RE bands of 682-762 nm and the NIR bands exhibit commonality traits under different growth stages.

3.2.Differences in spectral features and predictive performance at different growth stages

The positions and values of BE,GP,YE,RV,RE and NIRP were statistically analysed (Fig.2).There are no significant spectral shifts of GP,RV and NIRP.The positions of BE and RE indicated that the growth stage variations could lead to spectral position shifts.The statistical analysis revealed that the differences and commonality features due to growth stage effects are surely coexisted and these phenomena may motivate us to further evaluate LCC modelling performance with total dataset of SCtotaland single stage datasets of SCtillering,SCjointing,SCbootingand SCheading.

The PLS model was established based on SCtotalcalibration set.The model parameters for calibration wereRMSEV=4.80 mg L-1and RPD=1.49,and those for validation wereand RMSEV=5.05 mg L-1(Fig.3A).However,the utilization of individual growth stage dataset resulted in quite different model paraments,which indicated different evaluations of model accuracy(Fig.3B-D,F).This suggests that the LCC model might be specific to wheat growth stage,and model based on the entire datasets has a limit to predict LCC in individual growth stage.Validating the wheat LCC modelling performance at specific stage is essential rather than single entire model.

3.3.Variable selections across different growth stages

The selected bands by CARS-PLS,DT and RF are illustrated together in Fig.4.The regions of VB,BE,GP,YE,RV,RE and NIRP were used to clearly visualise the selected bands.The upmost axis presents the photochemical traits of chlorophyll a,chlorophyll b and typical functional groups in VNIR (325-1075) regions.Subsequently,the growth stage effects can be elucidated by wavelength positions and spectral absorption traits.

For CARS-PLS,the commonalities of sensitive wavelengths were mainly situated at RE regions near 755 nm and NIRP near regions 970 nm,which meant RE shift variations and leaf water absorption features[19].With respect to OOB permutations of DT,the selected wavelengths of tillering stages by DT was located at bands of VB:328;BE: 509;516,RE: 766;NIRP: 927 and 930 nm.The jointing stage only selected wavelengths of NIRP:927,934 nm,which could be assigned to the third overtone of C-H stretch.Sensitive wavelengths of booting stages were situated at bands of BE: 521;YE:587;RE: 722,742 and 753;NIRP: 881,972,977,1060 and 1073 nm and the heading stage,VB: 391 and 438;GP: 556;RE:691 and 733 nm.There is inconsistent distribution tendency for the selected bands of DT regression.Thus,the selected wavelengths by DT could not seek out the critical wavelengths that were resistant to growth stage effects.

Fig.2.Statistical analysis of LCC and spectral responses of growth stages.(A)LCC distribution of tillering.(B)Spectral response of tillering.(C)LCC distribution of jointing.(D)Spectral response of jointing.(E) LCC distribution of booting.(F) Spectral response of booting.(G) LCC distribution of heading.(H) Spectral response of heading.

The selected wavelengths of tillering stages by RF were located at bands of BE:504,506,510,511,513,514,515 and 516;RE:770;NIRP: 977 nm.The selected wavelengths of jointing stage by RF were located at bands of BE: 519;RE: 727,736,739,740,741,742,743,744 and 746 nm.The selected wavelengths of booting stage by RF were located at bands of BE: 522;RE: 727,736,739,740,741,742,743,744 and 746 nm.The selected wavelengths of heading stage by RF were located at bands of GP:545 and 552;RE:725,729,730,732,733,734,735 and 741 nm.Compared with DT,RF models frequently selected narrow sets of neighbouring wavelengths with OOB predictions permutation of variables importance.These properties could facilitate the exploration of the growth stage effects on the wheat LCC estimation.The RF selected bands near 520 nm at different stages,which is consistent with the variation coefficients of reflectance and the other research of[35].This indicates that this specific wavelength near 740 nm is resistant to growth stage effects,which echoes the correlation analysis.These results demonstrate that the spectral response of 504-552 nm and RE shoulder provide a potential for measuring wheat LCC[52].It was also shown that wavelengths of BE and RE could provide great estimating potentials of wheat LCC [53].

Compared to CARS and DT,RF selected robust wavelength spaces at different growth stages.The sensitive wavelengths presented spectral shifts from BE to GP and RE to RV from tillering to heading stages.Growth stage effects have a great influence on the sensitive band selection.Changing biophysical traits at different growth stages may mask spectral subtle differences of biochemistry variations [3].

3.4.LCC estimation at different growth stages

For exploring LCC estimation variations within specific growth stage,four pairwise sets of SCtillering,SCjointing,SCbootingand SCheadingand SVtillering,SVjointing,SVbootingand SVheadingwere used to calibrate the LCC models and compare the predictive performance,respectively.In this part,PLS,CARS-PLS,DT and RF were used to establish LCC estimation models (Fig.5).The blue and red circles represent the calibration and validation datasets.The colour transparency is used to illustrate LCC values derived from six fertilisation treatments.

Fig.5.PLS,CARS,DT,RF Modeling at different growth stages.(A) PLS modeling of tillering.(B) PLS modeling of jointing.(C) PLS modeling of booting.(D) PLS modeling of heading.(E) CARS modeling of tillering.(F) CARS modeling of jointing.(G) CARS modeling of booting.(H) CARS modeling of heading.(I) DT modeling of tillering.(J) DT modeling of jointing.(K) DT Modeling of booting.(L) DT Modeling of heading.(M) RF modeling of tillering.(N) RF modeling of jointing.(O) RF modeling of booting.(P) RF modeling of heading.

3.4.1.LCC modelling based on PLS and CARS-PLS

For the validation results,the models of the four growth stages outperformed the whole model.In comparison with the PLS models,results of CARS-PLS at the four growth stages showed some improvements.However,the modelling results of both PLS and CARS-PLS at the jointing stage presented poor estimation results.As such,other effective methods that are resistant to growth stage effects should be developed.

3.4.2.LCC modelling based on DT and RF

The modelling results revealed that the estimation performances appeared a merit rank of RF >DT >CARS with highand RPD and low RMSEV.From the modelling results of the three estimation methods,the growth stage effect has a great influence on wheat LCC estimation.In contrast to the CARS-PLS modelling results,the DT modellingof the tillering stage decreased to 0.61 and 0.51.This result indicates that single DT regression could not represent substantial relevance between spectral reflectance and LCC values at different growth stages.In light of RF regression,overall improvements of LCC estimation accuracy were detected over the four growth stages compared with CARS-PLS and DT modelling,which shows an excellent modelling potentials as RPD evaluation indicators [45].

The wheat LCC estimation by RF are based on ensemble traits of sub-models of input subset and variables partition,which contributes to prevent the changing canopy structure traits at different growth stages from masking subtle spectral responses of wheat LCC.In addition,establishing estimation models based on dataset from specific growth stages keep the canopy properties consistent within a certain growth period,thereby reducing the influences of the heterogeneity of crop canopy.This phenomenon may contribute to the resistance of structure traits compared with DT and CARS-PLS.The RF modelling results demonstrated that LCC modelling performance and sensitive wavelengths could decease the influences of growth stage effects.

3.4.3.RF estimated performances for combination datasets of multiple growth stages

Considering agronomic applications,the combination datasets of multiple growth stages were used to build and evaluate LCC estimating model across two or more growth stages,such as the adjacent two stages of tillering-jointing,jointing-booting,booting-heading and four stage combinations.

The models for the tillering-jointing stages presented the validation results (=0.77,RMSEV=2.45 mg L-1and RPD=1.93)(Table 2).The models for the jointing-booting stages presented thevalidation results (=0.81,RMSEV=2.37 mg L-1and RPD=2.02).The models for the booting-heading stages presented the validation results(=0.86,RMSEV=1.95 mg L-1and RPD=2.78).For the full growth stages,the modelling results represented the validation results(=0.81,RMSEV=2.28 mg L-1and RPD=2.49).The estimation results of multiple growth stages indicated that the RF models could overcome the growth stage effects.Thus,it was used to construct georeferenced LCC maps at the four growth stages (Fig.6).These findings provide insights in wheat LCC estimation of multiple growth stages to satisfy the requirement of agronomy management recommendation and high-throughput phenotyping.

Table 2 RF estimated performances for combinations datasets of multiple growth stages.

Fig.6.Geo-referencedLCCmapsofthefour growth stages.

4.Conclusions

The calibration model of partial least square (PLS) established by entire datasets to validate the predictive performance of specific growth stage displayed a poor estimation accuracy with the ratios of deviation to performance(RPD)of 1.11,0.73,1.61 and 1.78.The spectral wavelengths of red edge(RE)and blue edge(BE)shifts and poor estimation capability motivate us to further explore the growth stage effects by establishing LCC models at different growth periods.Wheat LCC modelling was conducted at the tillering,jointing,booting and heading stages by CARS-PLS,DT and RF.The modelling results revealed that the estimation performance showed a merit rank of RF >DT >CARS with high R2vand RPD and low RMSEV.Compared with CARS and DT,the selected wavelengths by RF indicated that specific wavelengths near 520 nm and 740 nm are resistant to growth stage variations following spectral shifts from BE to GP and RE to RV.The RF results yielded an estimation accuracy with RPD of 2.11,2.02,3.21 and 3.02 to accommodate the growth stage effects.Thus,the estimation models established and sensitive wavelength selected at specific growth stages can reduce the influence of the heterogeneity of crop canopy and decrease the influences of growth stage effects.This study explores spectral response on growth stage effects and provides models for chlorophyll content estimation to satisfy the requirement of highthroughput phenotyping.

CRediT authorship contribution statement

Dehua Gao:Visualization,Writing -original draft.Lang Qiao:Visualization.Lulu An:Conceptualization.Ruomei Zhao:Data curation.Hong Sun:Writing -review &editing,Resources.Minzan Li:Writing-review&editing,Resources.Weijie Tang:Visualization.Nan Wang:Data curation.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work was supported by the National Key Research and Development Program (2019YFE0125500),University-Locality Integrative Development Project of Yantai (2020XDRHXMPT35),the National Natural Science Foundation of China (31971785 and 41801245),and the Graduate Training Project of China Agricultural University (JG2019004,JG202026,YW2020007,QYJC202101,and JG202102).We would like to acknowledge the cooperation with University of Nottingham and UbiPOS UK Ltd.(107461:iAgriWatch-Intelligent Remote Sensing for Smart Farm).We thank Di Song and Jinbo Qiao for their help with field data collection.We appreciate the field assistance of Institute of Dry Farming,Hebei Academy of Agriculture and Forestry Sciences.