Comparison of algorithms for monitoring wheat powdery mildew using multi-angular remote sensing data

Li Song ,Luyun Wng ,Zheqing Yng ,Li He,c ,Ziheng Feng,b ,Jinzho Dun,c ,Wei Feng,c,*,Tinci Guo,*

a State Key Laboratory of Wheat and Maize Crop Science,Agronomy College,Henan Agriculture University,Zhengzhou 450046,Henan,China

b Information and Management Science College,Henan Agricultural University,Zhengzhou 450046,Henan,China

c CIMMYT-China Wheat and Maize Joint Research Center,State Key Laboratory of Wheat and Maize Crop Science,Henan Agricultural University,Zhengzhou 450046,Henan,China

Keywords:Characteristic wavelength selection Estimation model Machine learning Multi-angular remote sensing Wheat powdery mildew

ABSTRACT Powdery mildew is a disease that threatens wheat production and causes severe economic losses worldwide.Its timely diagnosis is imperative for preventing and controlling its spread.In this study,the multiangle canopy spectra and disease severity of wheat were investigated at several developmental stages and degrees of disease severity.Four wavelength variable-selected algorithms: successive projection(SPA),competitive adaptive reweighted sampling (CARS),feature selection learning (Relief-F),and genetic algorithm (GA),were used to identify bands sensitive to powdery mildew.The wavelength variables selected were used as input variables for partial least squares (PLS),extreme learning machine(ELM),random forest(RF),and support vector machine(SVM)algorithms,to construct a suitable prediction model for powdery mildew.Spectral reflectance and conventional vegetation indices(VIs)displayed angle effects under several disease severity indices(DIs).The CARS method selected relatively few wavelength variables and showed a relatively homogeneous distribution across the 13 viewing zenith angles.Overall accuracies of the four modeling algorithms were ranked as follows:ELM(0.70-0.82)>PLS(0.63-0.79) >SVM (0.49-0.69) >RF (0.43-0.69).Combinations of features and algorithms generated varied accuracies,with coefficients of determination(R2)single-peaked at different observation angles.The constructed CARS-ELM model extracted a predictable bivariate relationship between the multi-angle canopy spectrum and disease severity,yielding an R2 >0.8 at each measured angle.Especially for larger angles,monitoring accuracies were increased relative to the optimal VI model (40% at -60°,33% at +60°),indicating that the CARS-ELM model is suitable for extreme angles of-60°and+60°.The results are proposed to provide a technical basis for rapid and large-scale monitoring of wheat powdery mildew.

1.Introduction

In recent years,global climate change has led to higher rates of crop disease.Among them,wheat powdery mildew is one of the most destructive diseases worldwide,severely affecting both the yield and quality of wheat.When the disease occurs,a lack of timely diagnosis will delay effective intervention to control the disease.For its control,the spraying dose is determined by the diagnosis and directly influences the disease control effect and environmental safety.Timely and accurate detection of disease stress is needed for improving crop yield and quality,lowering disease control costs,and reducing environmental pollution of the environment.

Spectral remote-sensing data contain a wealth of information.Sensitive vegetation indices,sensitive angles,and optimization methods are all fundamental issues in research on spectral remote sensing.Remote sensing technology has shown high potential for use in agriculture in three ways.(I) It can overcome drawbacks of conventional monitoring methods of diseases,providing a new way to monitor and predict diseases of crops cultivated over large areas [1].When a plant is attacked by disease-causing pathogens,its cell activity,biochemical components,leaf morphology,leaf inclination angle distribution,canopy structure,and density change,in turn altering the plant’s leaf canopy reflectance spectrum.(II) It can show the characteristics of crop reflectance changes caused by changes in leaf structure [2].Using sensitive wavelengths and vegetation indices constructed from them,researchers have investigated the spectral diagnosis of crop diseases and determined that the sensitive wavelengths for wheat powdery mildew were 580-710 nm [3],512-634 and 692-702 nm[4],and 680-760 nm[5].(III)It can increase the accuracy of disease monitoring.Using the area of the red edge peak(∑dr680-760nm)to increase the accuracy of wheat powdery mildew detection[6],and a double green vegetation index was constructed using the 584 and 550 nm green bands to estimate the severity of wheat powdery mildew in a disease outbreak [3].Previous research has focused on deriving vegetation indices via selection of characteristic wavelengths.But owing to the limited information in vegetation indices containing few bands,the low precision of the associated models limits their practical applicability for precise disease control.

In aiming to improve the identification accuracy of powdery mildew,it is critical to select robust and effective information from numerous remote sensing data sources.Three types of characteristic selection algorithms,namely Filter,Wrapper,and Embedded[7],are typically relied upon for selection of sensitive factors for the remote-sensing detection of crop diseases.Besides the selected characteristic factors,the modeling algorithm used affects the accuracy of crop disease detection remote sensing.Currently,remote sensing models for crop disease monitoring include statistical models and artificial intelligence models.Because the first category are less generalizable in spatial and temporal dimensions,statistical models tend to have limited generalization ability,especially when they are fitted to data that is fully independent from the training set,investigators have proposed machine learning algorithms,such as support vector machines (SVM),neural networks(NN),and extreme learning machine(ELM),which can evaluate training error and generalization ability,and have applied them to remote sensing monitoring of wheat powdery mildew severity [8].Machine learning algorithms can tolerate outliers and noise while handling much data in a variety of forms,but machine learning-based remote sensing monitoring of crop diseases requires massive data samples and suffers from overlearning,local extrema,and curse of dimensionality [9].Crops and disease types have their own applicable machine learning algorithms,and the performance of machine learning methods often varies widely,even under the same disease conditions[10].It is desirable to investigate multiple algorithms to increase the accuracy of disease monitoring and the applicability of various models.

Bidirectional reflection is the basic feature of surface reflection in nature [11].Early in the development of remote sensing,owing to limitations of instruments and of our understanding,the remote sensing data acquired during vertical observation were readily affected by the soil background and vegetation canopy structure,resulting in the phenomena ‘‘same spectrum and different object”and‘‘same object and different spectrum”[12].These arise because the spectral information of a target obtained from a single angle is one-sided and cannot convey the spatial information of the monitored object.Using a multi-angle observation method permits the collection of three-dimensional canopy information of crops,providing more detailed and accurate information about crop growth[13].It has been proposed[14,15]that the inversion of the vertical component of chlorophyll in winter wheat can be better achieved by using multi-angle data to identify different plant types.Wheat powdery mildew has a ‘‘bottom-up” pathogenesis: in the early stage of disease onset,its symptoms are concentrated in lower parts of the plants.Consequently,changes in vegetation canopy structure due to disease stress manifest only when the biological system has already suffered severe injury.Multi-angle spectral technology can make full use of data from multiple observation angles and can convey more information about the lower part of the canopy.

In this study,multi-angle hyperspectral reflectance data were processed by four band-selection methods.The objectives were to:(1)identify the relationships between DI and common spectral parameters using multi-angle spectra;(2) compare the performance of machine learning algorithms in monitoring wheat powdery mildew using multi-angle spectral data;(3) evaluate the ability of the band-selection methods combined with machine learning algorithms to accurately monitor wheat powdery mildew.It was hypothesized that: (1) spectral reflectance and its derived vegetation index have large observation angle effects,leading to the low predictability of the linear model at extremely large angles;(2) nonlinear modeling methods improve the monitoring of wheat powdery mildew for precise crop management;(3)machine learning combined with variable selection methods mitigate the influence of observation angle and increase model accuracy at extreme angles.

2.Materials and methods

2.1.Experimental design

Five experiments were conducted in Zhengzhou,China,during two growing seasons.Various degrees of infection,inoculation methods,and cultivars of hexaploid wheat (Triticum aestivum L.)were studied (Table 1).All five experiments were conducted at the experimental station of Henan Agricultural University,located in Zhengzhou city(34°86′N,113°59′E).Experiments 1-2 were conducted in 2016-2017 and experiments 3-5 in 2019-2020.The susceptible cultivar Yanzhan 4110 and the moderately susceptible cultivar Guomai 301 were used.In a disease nursery field,artificially inoculated field,and naturally infected open field,the canopy spectra of wheat at multiple developmental stages and under multiple degrees of disease severity were acquired.The field sampling stages were booting,anthesis and filling.From the jointing stage,plants infected with powdery mildew in the nursery were taken out and inoculated after 16:00 in the field.Inoculation with the powdery mildew fungus was divided into four main categories of disease severity;(1) healthy control,not inoculated with the fungus;(2)mild disease degree,inoculated every 5 days;(3)moderate disease degree,inoculated every 3 days,and (3) severe disease degree,inoculated every day.The inoculated plots were covered with a transparent shed,opened during the daytime and closed at night,to keep them moist and the night temperature not lower than 10 °C.The transparent shed was removed when there were clear signs of powdery mildew infection in the inoculated plots.Experiments 1 and 3 were planted in 4 m × 2.1 m plots,experiments 2 and 4 were planted in 7 m×2.9 m plots,and experiment 5 was planted in 2.9 m × 15 m plots,all in along the north-south direction,spaced 18 cm apart(Fig.S1).All experiments followed a randomized complete block design,and each treatment was repeated three times.Management procedures followed local practice for winter wheat production.

Table 1 Seasons,wheat cultivars,soil status,treatments,and sampling dates for the five field experiments.

2.2.Data acquisition

2.2.1.Measurement of canopy multi-angular hyperspectral reflectance

Spectral reflectance data at the canopy level was collected in a 0.44 m2area of each plot under clear sky conditions,between 10:00 and 14:00 local time,using an ASD Field Spec handheld data logger(Analytical Spectral Devices,Boulder,CO,USA)fitted with a 25°-field-of-view fiber optic adaptor.This sensor has a sampling interval of 1.5 nm and a resolution of 3.5 nm,with a spectral detection range spanning 325-1075 nm.The multi-angular reflectance data of the wheat canopy were obtained in the principal plane of the sun.The term‘‘multi-angular”refers to(only)the zenith angle of observation,defined as the forward direction (measured positively away from the sun) and the backward direction (measured negatively toward the sun);the view zenith angle(VZA)is defined as zero at the nadir position (Fig.S1).The observation direction went from backward to forward,comprising 13 observation angles(-60°,+60°,-50°,+50°,-40°,+40°,-30°,+30°,-20°,+20°,-10°,+10°,and 0°),as designed according to the field goniometer system(FIGOS) [16].The distance between the probe of the spectrometer and the plant target was 80-100 cm,with each measurement repeated five times in the canopy per sampling point,with the average measurement taken as the spectral value of that point.To calculate the baseline reflectance,a 0.4 m × 0.4 m BaSO4calibration panel was used.Calibration panel reflectance measurements were taken both before and after the wheat vegetation measurements.

2.2.2.Measurement of disease severity

The DI,defined as the proportion of leaf area covered with disease spots,was quantified by visual interpretation,following the rules for the investigation,and forecasting of wheat powdery mildew(Chinese Standard:NY/T 613-2002)[17].The survey was conducted using the five-point survey method was used: at each spectral sampling point in each survey plot,20 plants were selected and the disease index of that sample was obtained by averaging five points.To minimize human error and bias,the same plant protection specialists oversaw all disease investigations in the field.The disease severity degree of all fully expanded leaves was used to express a plant’s condition for a given sampling point,and the grid method (lesion area as the percentage of total fully expanded leaf area)was used to calculate the severity of powdery mildew divided into nine scoring grades: 0,1%,10%,20%,30%,45%,60%,80%,and 100%[18].The number of wheat leaves per grade was recorded separately,and disease index(DI)per sampling point calculated as.

where x is the level of each gradient,n is the highest disease gradient (n=9),and f is the total leaf number for each gradient.

2.3.Methods of feature selection

2.3.1.Competitive adaptive reweighted sampling (CARS)

CARS is a variable selection method that tries to mimic Darwin’s evolutionary theory of ‘‘survival of the fittest” [19].Each set of spectral bands is treated as an individual,and the adaptive reweighted sampling(ARS)technique identifies wavelength points with large absolute regression coefficients in the PLS model,removes any low-weighted wavelength points,and finally selects the most accurate subset of the regression model after crossvalidation and successive optimization.In this study,removed those bands whose lost the key information of the spectrum.

2.3.2.Successive projections algorithm (SPA)

The SPA selects characteristic variables by calculating the size of the projection vector of remaining variables and selected variables,minimizing the latter’s covariance,thereby reducing the complexity of the model-building process and increasing its overall stability and accuracy,to achieve the purpose of selecting informative feature variables [20].In this study,the bands in the 400-900 nm spectral were calculated and the eigenvariables with the smallest covariance were selected.

2.3.3.Genetic algorithm (GA)

The GA is designed to seek an optimal solution,by simulating evolutionary processes in nature,but with a strong adaptive capacity and global optimization [21].Based on an obtained fitness value,genetic processes such as selection,crossover,mutation,and replication are performed on ‘‘individuals”.In these operations,individuals with low fitness are removed,so that populations with higher fitness than the previous generation are generated.In this study,fitting values were obtained from the 400-900 nm range.Spectral bands with low compatibility with DI were also excluded.A band population more suitable for multi-angular DI prediction was generated.

2.3.4.Feature selection Relief-F algorithm (Relief-F)

The Relief-F algorithm assigns unique weights to features based on the relevance of each and their categorization [22].Features whose weights are less than a certain threshold are removed,and the larger the weight of the feature,the greater its contribution to the classification.Choosing features contributing substantially to the classification scheme can be optimized by selecting a subset of feature parameters.In this study,only those bands with high importance were selected;their reflectance and corresponding DI data were then re-inputted into the Relief-F,to update all feature weights based on sample neighbors.

2.4.Modeling approaches

2.4.1.Extreme learning machine algorithm (ELM)

Extreme learning machine is a learning algorithm that uses a single hidden-layer feedforward neural network (SLFN).To work,this algorithm needs only to set the number of hidden layer nodes in the network.because it does not require adjusting the input weights and the bias of the network’s hidden elements during its execution,and generates a unique optimal solution,with fast learning speed and good generalization performance [23,24].For all samples in the experiment randomly divided into two groups for training and testing data sets.First,the spectral bands are input into ELM for training,and then,the trained ELM is used to predict the DI,ELM model can get the prediction result according to the memory.

2.4.2.Partial least squares (PLS)

PLS is a stochastic method that combines multiple linear regression and principal component analysis.PLS transforms the original variables with high data redundancy into fewer variables by selecting the optimal latent variables to describe the relationship between the predicted and true value [25].The PLS analysis used the reflectance data corresponding to 400-900 nm wavelengths.

2.4.3.Random forest (RF)

RF is a well-known classification algorithm proposed by Breiman [26].It implements a bootstrap resampling technique,where n samples are repeatedly and randomly selected for replacement from the original training sample set N,to generate a new training sample set for training decision trees;m decision trees are then generated according to the above steps,to form a random forest.This entire process was repeated 1000 times to weaken the effect caused by the random of sample.

2.4.4.Support vector machine (SVM)

The SVM algorithm,first proposed by Cortes and Vapnik in 1995[27],is based on statistical learning theory.It seeks the best compromise between a model’s complexity(specific learning accuracy of the training samples) and its learning ability (unbiased random sample recognition ability),based on vapnik chervonenkis (VC)dimensionality theory and the principle of structural risk minimization,to thereby obtain the best generalization capability.In this study,reflectance selected from 13 observation angles was used as input parameter and DI was used as dependent variable for training and prediction with SVM algorithm.

2.5.Model accuracy evaluation

A random data allocation method with a 7:3 ratio was used.In each training session,7/10 of the total samples were randomly selected as the training data set and the remaining 3/10 served as the validation data set.This data allocation method was applied to test each model,by randomly selecting 79 of the 114 data samples as training samples for model construction and using the remaining 35 samples as validation samples for model evaluation.The coefficient of determination R2(a measure of goodness of fit)of the regression of the measured DI on the estimated DI value of a model and the root-mean-square error RMSE were used to evaluate a model’s accuracy.The larger the R2and the smaller the RMSE,the higher is the estimation accuracy of a model.

where n is the number of test samples yiis the measured value of the disease index of wheat powdery mildew;y-is the average of the quantified disease index value,andis the value predicted by a given model.

2.6.Data analysis

To further eliminate the influence of non-target factors such as baseline drift and scattering on the spectra,only spectral data in the 400-900 nm band were retained for screening the feature variables and these used as variable inputs to a model.The bestperforming algorithm was deemed the one with the maximum R2and minimum RMSE.To investigate the differences in wheat canopy’s reflectance among observation angles in section 3.1,the spectral reflectance of a survey sample in a lightly diseased state was taken as an instance (DI=7.68).All spectral data,agronomic parameters,and algorithm operations were run with MATLAB 2019a,Origin 2019b,and MS Excel software programs,and compared with 10 conventional vegetation indices chosen based on prior studies (Table 2).

Table 2 Summary of selected vegetation indices (VIs) reported in the literature.

3.Results

3.1.Canopy spectral reflectance changes in wheat under multiple viewing zenith angles

Five representative wavelengths: 450 nm (blue),560 nm(green),650 nm (red),730 nm (red-edge),and 860 nm (nearinfrared) were chosen to analyzed changes in reflectance across zenith angles(Fig.1).As mentioned above,the spectral reflectance values of wheat plant canopies showed similar tendencies in the visible and near-infrared regions,with higher values in the backward than in the forward scattering direction.In the backward direction,as the viewing zenith angle (VZA) increased,reflectance initially increased and then decreased in the red and blue regions,but generally increased in the green band (Fig.1a) and red-edge and near-infrared bands(Fig.1b).In the forward direction,canopy reflectance declined with a greater angle in the visible bands(Fig.1a),whereas in the red-edge and near-infrared regions it generally increased (Fig.1b).

3.2.Relationships between mildew disease severity and vegetation indices under multiple viewing zenith angles

Fig.2a shows that the coefficient of determination between the selected VIs and DI showed similar angular effects,with R2values varying closely with zenith angle.Higher R2values (better fits)were obtained in the forward than backward observation direction,and the R2values gradually increased as angles changed from-60°to+30°.The R2values peaked at 10°to 30°in the forward direction.Among the 10 VIs,the PMI was the best predictor parameter at the+20°viewing angle,while another three VIs(SAVI,RIDA,and RPMI)gave the best forecasting results at the +10° viewing angle.Theremaining six VIs (SIPI,CI2,NDRE,REP,RES,and RVSI) yielded the highest R2values at the +30° viewing angle.

Fig.1.Observed reflectance changes of wheat canopy spectra in visible light band(a),and the red edge and near infrared bands(b).Reflectance changes in red,blue,green,red-edge,and NIR bands at multiple zenith angles.X-axis shows the viewing angle(negative angles are in the backscattering direction and positive angles are in the forward scattering direction),and the Y-axis shows spectral reflectance values.

Fig.2.Relationships between disease index and vegetation indices at multiple viewing zenith angles.X-axis shows viewing angle(negative angles are in the backscattering direction,and positive angles are in the forward scattering direction),the Y-axis shows vegetation indices,and the Z-axis shows the fitted relationships(R2)of VIs and DI at 13 viewing zenith angles.Model calibrations and validations of RPMI,RES,and SAVI (b-g).The black dashed lines indicate a 1:1 relationship.

The plotted quantitative relationships between the three common VIs (RPMI,RES,and SAVI) and DI under three extreme representative viewing angles are shown in Fig.2b-g.

Fig.2 shows that the performance of RPMII,RES and SAVI models varied greatly across three viewing angles of-60°,0°and+60°.The model fitting accuracy of RES surpassed that of SAVI,but was lower than that of RPMI.Among the 3 VIs,all the 0° was the best predictor at -60°,0°,and 60° viewing angle.At -60°,0°,and 60°viewing angle,the RPMII models with corresponding R2values of 0.41,0.67,and 0.50 in the calibration set (Fig.2b),and of 0.41,0.67and 0.44 in the validation set (Fig.2e).Compared with the RES and SAVI models,the R2of the RPMI models was increased by >17% in the nadir direction for both calibration and validation sets,and increased by respectively 12% and 9% at the -60° angle and 13%and 2%at the+60° angle in the calibration and validation(Fig.2).The R2of RES and SAVI models declined to greater extent as the observation angle increased in the forward direction,and reduced the powdery mildew prediction effect at the extreme angles.

3.3.Screening of sensitive wavelength variables

Fig.3 shows the sensitive wavelengths selected by the four methods.Also shown is the best band position and the number of wavelength variables selected per viewing angle.As depicted in Fig.3a,because only variables with minimum collinearity are selected by SPA,the wavelengths and adjacent wavelengths occur in different regions.Fewer wavelengths were evidently obtained by the SPA than by the CARS method (Fig.3b),while the wavelengths obtained by the GA or Relief-F method included not only all or most of the wavelengths obtained by the method but also the adjacent wavelengths and the wavelengths that were not yet selected (Fig.3c,d).

For all four methods,the selected wavelength variables were distributed in the visible region and the red-edge and nearinfrared regions.However,those selected by CARS and GA were evenly distributed there,whereas SPA and Relief-F selected more wavelength variables selected in the visible region of the forward versus backward scattering direction.Except for Relief-F,the other three wavelength variable-selection algorithms generally showed the same trend,with the number of wavelength variables selected increasing with viewing angle in both the forward and backward scattering directions.

3.4.Model building based on machine learning

Model building was clearly improved using the CARS method,as it had the highest accuracy.The second preferred method was the GA,whereas the lowest accuracy was obtained with Relief-F(Fig.4).

Fig.4 illustrates how the R2is strongly dependent on the modeling algorithm applied,and how this value varies with observation angle and selection method.The predictive ability of monitoring models developed by the ELM method outperformed those based on PLS and SVM methods,with the RF model giving the lowest monitoring accuracy.In comparing their accuracy,the four model algorithms ranked as follows: ELM (0.70-0.82) >PLS(0.63-0.79) >SVM (0.49-0.69) >RF (0.43-0.69).

When the wavelength variables selected by the different methods were input into the ELM model,the accuracy values for predicting powdery mildew severity were relatively stable at each angle.After they were input into the PLS model,its R2was relatively stable between -40° and +40° viewing angles,but sharply decreased at the extreme (larger) angles.When these wavelength variables were input into the SVM model,the R2decreased with a larger view zenith angle in the backward scattering direction,whereas in the forward direction,the R2initially increased but decreased as the angle increased.Under the RF model,the R2decreased as the view zenith angle increased,in both the forward and backward scattering directions,and generally showed a singlepeak trend.

Fig.3.Location and number of feature bands selected by the four feature selection methods at 13 viewing zenith angles.(a)Location and number of feature bands selected by the SPA (400-900 nm).(b) Location and number of feature bands selected by the CARS (400-900 nm).(c) Location and number of feature bands selected by the GA (400-900 nm).(d) Location and number of feature bands selected by the Relief-F (400-900 nm).

Fig.4.Predictive models constructed by four feature selection methods.(a) SPA method.(b) CAR method.(c) GA method.(d) Relief-F method.

Among the combined prediction models incorporating diverse wavelength selection algorithms and modeling methods,the CARS-ELM model gave the best prediction performance,having the highest R2at each angle,and being the most stable across the 13 viewing angles (R2=0.81-0.83).Not be overlooked,however,is the GA-ELM model which is another candidate accurately monitoring disease under multi-angle observation conditions(R2=0.75-0.81).

3.5.Model evaluation based on machine learning

Fig.5 (a,d) shows that the R2and RMSE of the CARS-ELM and GA-ELM models varied across the viewing zenith angles.Compared with the GA-ELM model,the fitting accuracy of the CARS-ELM model was greater and showed superior angular stability by maintaining a larger R2even at the extreme angles(Fig.5a).For the GAELM model,its monitoring accuracy was like that of CARS-ELM at the near nadir,but its R2declined to greater extent as the observation angle increased,especially at extremely large angles,where R2was reduced by >15% (Fig.5d).

To emphasize the CARS-ELM model’s advantages at improving the angular viewing,a scatterplot shows this combined model’s excellent ability to monitor disease severity (Fig.5b-c,e-f).Compared with the GA-ELM model,the R2of the CARS-ELM model was increased by about 3% in the nadir direction for both calibration and validation sets,and by respectively 10% and 8% at the-60° angle and 7% and 6% at the +60° angle in the calibration and validation.The CARS-ELM model provided the highest estimation accuracy in the nadir direction(Cal-R2=0.81,Cal-RMSE=1.95;Val-R2=0.82,Val-RMSE=1.76),and greatly increased the prediction effect at the extreme (larger) angles.

4.Discussion

4.1.Combining variable-selection methods and modeling algorithms for plant disease monitoring

Previous study has confirmed that spectral remote-sensing data in crop disease monitoring importance and application prospects.The sensitive bands of wheat powdery mildew are located at covering the visible and the near-infrared spectra(380-1300 nm)[38].Previous researchers[39-42]have linked sensitive wavelengths to specific diseases.The disease VI can be constructed according to the inter-band combination and transformation based on a screening of sensitive spectral bands [43-46].As the disease progresses and becomes more severe,it changes in the host plant’s foliage,morphology,and physiological status [47].The performance of selecting a single band for extracting disease information during modeling is limited,and the lack of useful band information leads to a decrease in the accuracy of the linear model.The highest R2of the optimal vegetation index RPMI in our study was 0.67.In the early stage of powdery mildew disease,the collected canopy spectral information is different from the actual incidence rate,which makes powdery mildew monitoring more challenging.Accordingly,more effort is still needed to improve the inversion process and its results.

Fig.5.Comparison between the models CARS-ELM and GA-ELM at different viewing zenith angles.The CARS-ELM model’s calibration and validation (a-c),the GA-ELM model’s calibration and validation (e-f).The black dashed lines indicate a 1:1 relationship.

Some scholars have attempted to accurately distinguish and retain only the most sensitive bands or spectral regions in modeling when using hyperspectral data for disease monitoring,and achieved good monitoring results [48,49].Understandably,because of large inherent differences between monitoring targets,different methods are likely appropriate for screening the most sensitive wavelength regions.In our study,more wavelength variables were obtained by the GA and Relief-F methods,a reason for this could be that when one wavelength is selected as relevant,the adjacent wavelength has a high probability of being selected as well,and covariance among the variables may have affected model accuracy.Because the fewer wavelength variables identified by CARS were evenly distributed in the visible and red-edge and near-infrared regions,information in the entire band could be fully utilized.These sensitive band screening methods could provide a useful reference for monitoring research on other related plant diseases.Applying differentially effective wavelength-screening methods to different monitoring objects for differing purposes would result in disparities in their dynamic monitoring ability.These methods could not successfully handle extreme conditions in real field environments,such as large viewing angles,highnitrogen application conditions,and high planting densities.

The choice of a modeling method will directly influence the effectiveness of plant disease monitoring[50,51].In the ELM model only the number of hidden layer nodes of the network must be set;the input weights of the network and the bias of the hidden elements during the execution of its algorithm need not be considered.Overall,then,the ELM model generated a unique optimal solution and showed high learning speed and generalizability.In this study,in terms of monitoring accuracy,models were ranked as ELM >PLS >SVM >RF.A suitable modeling method is crucial for acquiring special characteristics in diverse target objects.Among the algorithm combinations tested in this study,the CARS-ELM model increased predictor accuracy,especially at extreme angles,suggesting that the CARS-ELM model is a robust algorithm for powdery mildew surveillance.In this way,an appropriate machine learning algorithm could be employed to better coordinate multivariate factors,thereby enhancing the covariant relationship between canopy spectrum information and wheat powdery mildew status.Previous researchers [52,53] also made similar attempts,by combining certain variable selection methods with machine learning algorithms,to solve large-dimensional problems in spectral data.These findings together indicate that the effectiveness of combinations of algorithms is strongly influenced by the selected sensitive features,by differences in the monitored objects,and by coordination between algorithms.The CARS-ELM model evaluated in this study provides a method for band selection and a modeling framework for monitoring crop disease.Whereas in our study,the optimized model was built presuming conditions of a canopy structure that is relatively stable(in the booting to mid-filling stage),wheat powdery mildew usually appears at the late jointing stage during which canopy structure is changing.Whether the optimized model in this study can be applied as effectively to the late jointing stage awaits further validation and refinement.

4.2.Application of multi-angular hyperspectral remote sensing to plant disease monitoring

In recent years,with the diversification of remote-sensing earth observation methods,stronger requirements have been put forward for disease monitoring systems that rely on remotely sensed data.Multi-angle remote sensing that incorporates abundant spectral information has become a research focus in the development of remote sensing technology [54,55].Spectra harboring different angle information are usually obtained from the main plane of the sun,by using a self-made multi-angle observation frame based on that described by Sandmeier et al.[16].In this study,the correlations of 10 conventional VIs with disease severity at 13 viewing angles were compared to evaluate the impact of observation angle on monitoring models.As evidenced by Fig.2a,the correlation strength varied considerably among the angles,showing better performance in the forward than backward scattering direction.Multi-angle remote sensing technology improves the applicability of existing vegetation indices to crop production management[57],and increases the accuracy of monitored targets by extracting suitable observation angles for pertinent crop growth characteristics.In actual field production settings,the application of a single optimal observation angle is limited.In this study,the results suggested that the +10° to +30° viewing angles were more sensitive than the nadir direction for the DI detection of wheat powdery mildew.Some researchers have tried to attenuate the angular effects and extend the range of viewing angles[56,57].He et al.[35]constructed a novel spectral parameter (RPMI) that could be more appropriate for monitoring powdery mildew at 0°to+30°in wheat,implying an expansion of suitable angles,and thereby facilitated the monitoring of diseases in precision agriculture.In previous studies,the choice of sensitive bands at several viewing angles was often inconsistent with that using multi-angle remote sensing data [58],and differing methods to select the feature variables increased variation in the selection of sensitive bands (Fig.3).In our study,four modeling approaches were used to analyze the effectiveness of varying band selection sensitivity for powdery mildew’s monitoring.Angle effect was a formidable barrier in both vegetation index and reflectance modeling,resulting in lower R2values at larger angles.The problem of angular effects emerging at large angles must be resolved to permit the flexible application of remote sensing technology in precise agriculture.The CARS-ELM model clearly performed better at mitigating such large angle effects,having significantly improved an R2for angles of -60° to+60 (R2values >0.8),whereas other models had significant single-peak effects when monitoring powdery mildew.Multiple algorithms should be considered and the development of an angle-independent inversion model will bolster the usefulness of monitoring crop production and reduce the bias of viewing angle to disease prediction.

In the present study,an angle-independent model was developed to increase the generalizability of monitoring crop production for diseases.Owing to the bottom-up onset characteristics of powdery mildew,minor changes to the canopy at the early stage of infection can easily influence the extraction of key spectral information,thereby reducing the overall accuracy of disease monitoring.In this study,a variety of feature screening methods were used to strengthen the extraction of key spectral information of powdery mildew,thus making the spectral variables of the input model more sensitive.Using the feature screening method reduces the number of variables,rendering the model operation simpler and faster and increasing modeling efficiency.Because the generalizability of the ELM modeling method is higher than that of other algorithms,the constructed model is usually more stable and more accurate.The powdery mildew prediction model finally constructed,CARS-ELM,increased the extraction of large-angle information while reducing the effect of extreme angles on the powdery mildew monitoring model (Fig.4b).In terms of a more flexible application and fewer angular limits to crop disease diagnosis,we found that,unlike conventional VIs,CARS-ELM was stable across the 13 VZAs.The reflectance data sources have also been extended,and reflectance measured from any angle can be used to build models for accurately monitoring disease.This enabled us to construct a common algorithm to predict the severity of wheat powdery mildew over a wide-angle range using portable monitors and multi-angular satellite data.However,the validity of this model for other disease monitoring now awaits further verification.Different scales of remote sensing technology have their own advantages.Satellite-and UVA-based remote sensing mainly used to guide large-scale disease monitoring plans,providing technical support and reference methods for crop protection.Whether remote sensing data from various observation angles and from unmanned aerial vehicles and satellite platforms can be flexibly used in our model and the problem of large angle effects can be overcome remains to be verified.For ground-based remote sensing,real spectral information extracted is more suitable for powdery mildew monitoring of small plots.A future challenge is to implement the optimal model in portable disease monitoring equipment,such as smart phones and small handheld instruments,or use remote sensing platforms such as drones and satellites,to estimate and compare the degree of disease infection and realize realtime disease monitoring in situ.Integrated with plant protection knowledge,automatically formulated but nonetheless effective prevention and control plans will permit the early detection and control of plant diseases.However,the reliability of such investigations should be extensively tested over longer time spans(multiple years) and under multiple disease conditions across sites.

5.Conclusions

This comparative study of four sensitive-band selection methods showed that their selected spectral regions and bands were similar,suggesting that wheat powdery mildew is characterized by specific spectrum-sensitive regions.This study integrated the ELM model approach with CARS-selected optimum wavelengths,and the CARS-ELM model was distinguished by increased accuracy,especially under extreme angle conditions.The problematic angle effect in monitoring powdery mildew severity was smaller for this CARS-ELM model than for the other tested prediction models.Accordingly,the angle range of powdery mildew monitoring can now be expanded to include both positive and negative 60°,thereby allowing disease surveillance to be pursued at any viewing angle,whether on the ground,on flying drone platforms,or at the satellite scale.These findings add to the body of knowledge about angular remote sensing,as well as providing potential applications in the direction of plant disease monitoring conducted by unmanned aerial vehicles or by multi-angular remote sensing satellites.These applications may permit the early prevention of crop diseases,improving protective interventions and reduce yield losses.To further increase the robustness of the proposed model,it should be tested and perfected for diverse crop types and environmental conditions.

CRediT authorship contribution statement

Li Song:Visualization,Investigation,Validation,Writing-Original.Luyuan Wang:Investigation.Ziheng Feng:Software.Zheqing Yang:Visualization.Li He:Supervision.Jianzhao Duan:Data Curation.Wei Feng:Conceptualization,Resources,Writing -Review &Editing.Tiancai Guo:Conceptualization,Project administration,Writing -Review &Editing.

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 Natural Science Foundation of China (31971791) and the National Key Research and Development Program of China (2017YFD0300204).

Appendix A.Supplementary data

Supplementary data for this article can be found online at https://doi.org/10.1016/j.cj.2022.07.003.