Comparison and Transferability of Nitrogen Content Prediction Model Based in Winter Wheat from UAV Multispectral Image Data

Abstract

Information about the nitrogen dynamic in wheat is important for improving in-season crop precision nutrient management and cultivated land sustainability. To develop unmanned aerial vehicle (UAV)-based spectral models for an accurate and effective assessment of the plant nitrogen content in the key stages (jointing, booting, and filling) of wheat growth, winter wheat experiment plots in Henan Province, China, were used in this study. Based on the K6 multichannel imager, 5-band (Red, Green, Blue, Red edge, and Near-infrared (Nir)) multispectral images were obtained from a UAV system and used to calculate 20 vegetation indices and 40 texture features from different band combinations. Combining the sensitive spectral features and texture features of the nitrogen content of winter wheat plants, BP neural network (BP), random forest (RF), Adaboost, and support vector machine (SVR) machine learning methods were used to construct plant nitrogen content models, and compared for the model performance and transferability. The results showed that the characteristics of different spectral features were different, but most of them had a partial normal distribution. Compared with spectral features, the distribution of texture features was more discrete. Based on Pearson’s correlation analysis, 51 spectral and texture features were selected to build four machine learning models. The estimates of plant nitrogen by the RF and Adaboost methods were relatively concentrated, mostly close to the 1:1 line; while the estimates of plant nitrogen from the BP and SVR methods were relatively scattered. The RF method was the best, with coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) of 0.811, 4.163, and 2.947 g/m², respectively; the SVR method was the worst, with R², RMSE, and MAE of 0.663, 5.348, and 3.956 g/m², respectively. All models showed strong transferability, especially the RF and Adaboost methods, in predicting winter wheat nitrogen content under rainfed and irrigation water management.

1. Introduction

Wheat is the largest sown area and the most widely distributed food crop in the world [1]. As a major wheat producer, China contributed to more than 17% of the world’s total wheat production in 2021 [1, 2]. Stable wheat production is essential to China and the global food supply. Nitrogen is a key and essential nutrient for wheat yield and quality, and accurate information about the nitrogen content of the wheat plant is important for in-season crop growth monitoring, precision fertilizer application, and environmental quality [3, 4, 5].

Given the labor-intensive nature of field sampling and laboratory analysis, research has been conducted over the past decades to develop alternative methods for rapid monitoring, and accurate prediction of crop nitrogen content based on spectral imaging collected from near-ground, unmanned aerial vehicle (UAV), and satellite remote sensing systems [6, 7, 8]. Particularly, progress has been made in sensitive band screening, vegetation index construction, optimization of prediction and inversion methods, and accuracy improvement [7, 8, 9, 10]. However, the efficiency of data collection and performance of various nitrogen prediction models vary with the sensing platforms.

In terms of near-ground sensors, Yang et al. [11] constructed a wheat leaf nitrogen content prediction model based on wheat canopy hyperspectral data collected from a handheld analytical spectral devices (ASD) spectrometer at different growth stages; using the 39 sensitive characteristic bands, the R² of the prediction model was as high as 0.998. Similarly, Zhang et al. [12] used ASD wheat canopy hyperspectral data and constructed 14 different vegetation indices such as the Soil Adjusted Vegetation Index (SAVI) to retrieve the nitrogen content of wheat leaves based on the characteristic bands; they found that the combined use of multiple indices could significantly improve the model accuracy compared with a single vegetation index (R² = 0.92 and 0.83, respectively). With the progress of remote sensing technology, airborne imagers have also been widely used. For example, Li et al. [13] combined handheld canopy spectral data and airborne canopy hyperspectral images at different growth stages of wheat using the N-PROSAIL model to estimate the nitrogen content of wheat canopy and obtained a model R² of 0.83; although there was a spectral difference between the handheld ASD and airborne spectrometers, the data fusion did not affect the N-PROSAIL model’s performance. The promising prediction results achieved with near-ground and airborne hyperspectral data attract researchers to use the lower-cost UAV remote sensing data for rapid monitoring and inversion of crop nitrogen, but the low spectral resolution of UAV multispectral data affects the prediction performance of nitrogen models. In addition to spectral models based on spectral reflectance, the rich texture information of the UAV multispectral images has not been widely used in constructing plant nitrogen models. Previous studies have shown that the texture feature can improve the identification of useful spatial features from the original images and enhance the inversion accuracy when retrieving crop parameters [14, 15, 16]. For example, Jia and Chen [14] established a model using principal component regression analysis for predicting the nitrogen content of winter wheat using UAV image features at a spectral resolution of 0.06 m; the accuracy of the model established by fusing the spectral and texture features (R² = 0.68) of UAV multispectral images was improved by more than 10% compared with that established by a single vegetation index (R² = 0.66) or texture feature (R² = 0.65). Zheng et al. [17] evaluated the potential of integrating texture and spectral information from UAV-based multispectral imagery for improving the quantification of nitrogen status in rice crops, indicating model vegetation indices (Vis) with a R² of 0.14, texture features with a R² of 0.41, and the combination of VIs and texture features with a R² of 0.68. The results revealed the potential of image textures derived from UAV images for estimation of winter wheat nitrogen status. Therefore, a comprehensive analysis of the sensitivity of spectral and texture features to crop nitrogen content, and the predictive model establishment using appropriate methods are of great significance to improve the accuracy of nitrogen content prediction, enhance the applicability of models, and reduce costs.

The methods for prediction and inversion of plant nitrogen content mainly include statistical and physical methods. Statistical methods used univariate and multiple regression to establish linear, logarithmic, and power function models [14, 18]. Physical models are mainly radiative transfer models and geometric optical models, and the crop nitrogen content predicted by screening feature bands through sensitive parameter analysis, using lookup table method and artificial neural network (ANN) method [8, 13, 19]. In recent years, with the advancement of data mining techniques, various programs, such as artificial neural networks (ANN), genetic algorithms (GA), random forest (RF), and other hybrid methods, have been increasingly applied to the prediction of crop nitrogen content, which outperform traditional models in terms of accuracy [9, 20, 21]. For example, Chlingaryan et al. [20] compared the analysis of traditional statistical analysis methods and machine learning methods in crop nitrogen content and yield prediction. They found that machine learning regression methods, such as least squares support vector machines (LS-SVR) and back-propagation neural networks (BPNN), have promised a higher accuracy. In addition, different machine learning methods differ in prediction accuracy [22, 23]. For example, Yang et al. [11] used backward transmission neural network (BP), SVR, and radial basis neural network (RBF) methods for the prediction of canopy nitrogen in winter wheat with model R² of 0.82 (SVR) to 0.98 (RBF); Qiu et al. [22] used Adaboost, ANN, K-neighborhood (KNN), partial least squares (PLSR), RF, and SVR machine learning regression methods for the prediction of rice nitrogen nutrient index, and found that the RF achieved the highest model accuracy with a R² of 0.98 during the filling stage. Although machine learning methods are proven to be superior to traditional statistical analysis methods, few studies investigated the transferability of the models when applied to UAV images. To explore the model prediction effects and transferability of general machine learning methods (BP and SVR) and ensemble learning methods (RF and Adaboost) for constructing plant nitrogen content models, especially in agricultural applications, this study aims to compare the performance of different machine learning models in predicting the wheat nitrogen content with a combination of spectral and texture features and provide insights for future model deployment to support the rapid in-season assessment of nitrogen nutrition and precision fertilization across large extents. Specifically, our study will focus on the following aspects:

1. Extracting spectral and texture features of UAV images based on the acquired multispectral images at key growth stages of winter wheat.

2. Feature optimization by Pearson’s correlation analysis using the extracted spectral and texture features.

3. Comparing the model accuracy for estimating wheat canopy nitrogen content using different machine learning methods (BP, RF, Adaboost, and SVR).

4. Evaluation of the computation efficiency and transferability of different machine learning methods under rainfed and irrigation water management.

2. Materials and methods

2.1 Study area and experimental design

The study area is located in Shangshui County, Henan, China, as shown in Figure 1. The study area is flat and belongs to the temperate continental monsoon climate with cold and dry winters and high temperatures and rainy summers. Winter wheat, maize, cotton, and other crops are mainly cultivated. The growth cycle of winter wheat is 8 months. Generally, it is sown in October and harvested at the end of May and early June of the following year. The experiment was conducted in the jointing, booting, and filling stages of winter wheat, and the field cover did not change much during these three phenological periods. Field activities affecting the ground surface, such as plowing and sowing, were avoided in the experimental plots. The soil type in the study area is sandy ginger black soil (Chinese Soil Taxonomy Classification).

Figure 1.

Location of the study area and the spatial distribution of the experimental design.

Five nitrogen application levels were set using a randomized block experiment design, which were N0 (0), N6 (60 kg/hm²), N12 (120 kg/hm²), N18 (180 kg/hm²), and N24 (240 kg/hm²), and of which 50% was applied as base fertilizer and the remaining 50% was applied at the jointing stage. All treatments had 150 and 90 kg/hm²of phosphorus and potassium fertilizer, all of which were applied as base fertilizer. Two levels of water set were: rainfed only (W0) and normal irrigation (W1). Each nitrogen level treatment was repeated three times, so a total of 30 plots. The spatial distribution of the experimental design is shown in Figure 1.

2.2 Data acquisition and processing

2.2.1 UAV-based multispectral image acquisition and preprocessing

This study employed a UAV system consisting of a six-rotor DJI M600 UAV and a K6 multichannel multispectral imager (Anzhou Technology, Beijing, Co., Ltd.) to acquire multispectral images. The K6 multispectral imager mounted onboard the UAV had an incident light sensor and five spectral bands with center wavelengths at Blue (450 nm), Green (550 nm), Red (685 nm), Red edge (725 nm), and Near-infrared (780 nm, Nir). During the growth period of winter wheat from 2020 to 2022, the multispectral images of the winter wheat canopy were acquired at the jointing, booting, and filling stages with a flight height of 50 m. When the aircraft flew, the lens was vertically downward with a field angle of 30° and a heading overlap of 70%, and a lateral overlap of 75%. The canopy reflectance data of winter wheat were extracted by format conversion, image mosaicing, geographic information correction, and radiometric calibration. The digital number (DN) values of the images were transformed into reflectance values per band by applying the empirical line model derived from the measured reflectance values and DN values of the five calibration canvases.

2.2.2 Ground data acquisition and processing

The acquisition of ground-truth data is synchronized with the acquisition of the UAV multispectral images. The area with uniform growth is selected in every plot, and 20 of the single stem samples were taken in sealed bags by fixing the total number of stems in a 1-meter double row. The plant organs were separated into leaves, stems, and ears in the laboratory and placed in paper bags respectively. They were killed at 105°C and dried to a constant weight at 80°C. After the organs were crushed, the nitrogen content was determined by the Kjeldahl method. Finally, 360 measured nitrogen values of winter wheat plant nitrogen content were obtained for the three growth stages. The samples were divided into training and test datasets according to the 1:1 ratio, and the models were trained using cross-validation.

2.3 Methods

2.3.1 Feature extraction

1. Spectral vegetation index calculation

Since the launch of the Earth Resources Satellite, scientists have begun to study the relationship between spectral response and vegetation [24, 25, 26, 27]. Crops with nitrogen deficiency will show obvious apparent characteristics such as reduced coverage and yellowing of leaves [28, 29]. This study selected 20 common vegetation indices used for nitrogen content prediction, and the formulas are shown in Table 1.

1. Texture feature extraction

Vegetation indices	Abbreviation	Calculation formulas	References
Green Normalized Vegetation Index	GNDVI	(Rnir-Rgreen)/(Rnir + Rgreen)	[30]
Green Optimized Soil Adjusted Vegetation Index	GOSAVI	1.16 × ((Rnir-Rgreen)/(Rnir + Rgreen + 0.16))	[31]
Normalized Difference Vegetation Index	NDVI	(Rnir-Rred)/(Rnir + Rred)	[32]
Modified Simple Ratio Index	MSR	((Rnir/Rred)-1)/(((Rnir/Rred) + 1)^0.5)	[33]
Rededge Optimized Soil Adjusted Vegetation Index	REOSAVI	1.16 × ((Rnir-Rred)/(Rnir + Rred + 0.16))	[34]
Rededge Renormalized Difference Vegetation Index	RERDVI	(Rnir-Rred edge)/(Rnir + Rrededge)^0.5	[30]
Chlorophyll Absorption Ratio Index	CARI	(Rred edge-Rred)-0.2 × (Rrededge + Rred)	[35]
Optimized Soil Adjusted Vegetation Index	OSAVI	(Rnir-Rred)/(Rnir + Rred + 0.16)	[35]
Normalized Green-Blue Difference Index	NGBDI	(Rgreen-Rblue)/(Rgreen + Rblue)	[36]
Enhanced Vegetation Index	EVI	2.5 × ((Rnir-Rred)/(Rnir + 6 × Rred-7.5 × Rblue+1))	[37]
Triangular Vegetation Index	TVI	0.5 × (120 × (Rnir-Rgreen)-200 × (Rred-Rgreen))	[38]
Atmospherically Resistant Vegetation Index	ARVI	(Rgreen-Rred)/(Rgreen + Rred-Rblue)	[39]
Excess Green Index	EXG	2 × Rgreen-Rred-Rblue	[40]
Ratio Vegetation Index	RVI	Rnir/Rred	[41]
Modified Triangular Vegetation Index	MTVI	(1.5 × (1.2 × (Rnir-Rgreen)-2.5 × (Rred-Rgreen)))/(((2 × Rnir + 1)^2–6 × Rnir-5 × (Rred)^0.5–0.5)^0.5)	[42]
Soil Adjusted Vegetation Index	SAVI	1.5 × (Rnir-Rred)/(Rnir + Rred + 0.5)	[43]
Normalized Blue-Green Difference Vegetation Index	GBNDVI	Rnir-(Rgreen + Rblue)/(Rnir + (Rgreen + Rblue))	[30]
Renormalized Difference Vegetation Index	RDVI	(Rnir-Rred)(Rnir + Rred)^0.5	[42]
Difference Vegetation Index	DVI	Rnir-Rred	[44]
Optimized Vegetation Index	VIplot	1.45 × (R^2nir + 1)(Rred + 0.45)	[45]

Table 1.

Vegetation indices and the calculation formulas.

Texture feature extraction methods mainly include statistical methods, such as gray level cooccurrence matrix (GLCM), texture spectrum and geometric methods; model methods of random field model and fractal model methods; signal processing methods and structural analysis methods [46, 47]. Among these methods, the GLCM method is an image recognition technology currently recognized by the academic community as an image recognition technique with strong robustness and adaptation characteristics, which can effectively achieve the classification and retrieval of images and maximize the accuracy of remote sensing image classification processing [16, 26]. In this study, the texture features from five bands in multispectral images are extracted through the GLCM method, and the extracted texture feature information mainly includes eight indicators of con, cor, dis, ent, hom, mean, sm, and var. The specific calculation method is described in Zhou et al. [26].

2.3.2 Machine learning regression method

1. Back-propagation neural network

Back-propagation neural network (BP) is a multilayer feedforward network using error back-propagation for model training, which is one of the most widely used neural network models [48, 49]. This study uses identity as the plant nitrogen content training activation function of the model. Also, to prevent overfitting, parameters such as learning rate and regularization are often introduced to optimize the model [50]. In this study, a 3-layer network structure was used, and a quasi-Newtonian method family optimizer (lbfgs) was used to improve the running speed. In this study, the detailed parameters for BP are shown in Table 2.

1. Random forest

BP		RF		Adaboost		SVR
Parameters	Parameter value	Parameters	Parameter value	Parameters	Parameter value	Parameters	Parameter value
Data cut	0.5	Data cut	0.5	Data cut	0.5	Data cut	0.5
Data shuffling	Yes	Data shuffling	Yes	Data shuffling	Yes	Data shuffling	Yes
Cross-validation	30% off validation	Cross-validation	30% off validation	Cross-validation	30% off validation	Cross-validation	30% off validation
Activation function	identity	Identity node split evaluation criterion	mse	Number of base classifiers	100	Penalty factor	1
Solver	lbfgs	Minimum number of samples for internal node splitting	2	Loss function	linear	Kernel function	linear
Learning rate	0.1	Minimum number of samples of leaf nodes	1	Base classifier	Decision tree classifier scale	Kernel function coefficient	scale
L2 regular term	1	Maximum depth of tree	10	Learning rate	1	Maximum number of terms in kernel function	3
Number of iterations	1000	Maximum number of leaf nodes	50	/	/	Error convergence condition	0.001
Number of hidden layers 1st neurons	100	Number of decision trees	100	/	/	Maximum number of iterations	1000

Table 2.

Parameters for BP, random forest (RF), Adaboost, and support vector machine (SVR) methods.

Random forest is a typical representative of ensemble learning with bagging idea, a supervised machine learning method constructed by integration with decision tree as the base learner, and it introduces randomness in the training process of decision trees to make it have excellent resistance to overfitting as well as noise resistance; moreover, RF can be trained in parallel during model training to improve the efficiency of training, while feature importance can be obtained [51, 52]. RF reflects its randomness from two aspects: sample selection and feature selection. Combined with the study of Liepe et al. [53], the parameters of RF node splitting evaluation criterion and maximum depth of the tree in this study are set in Table 2 after several runs and the algorithm steps are as follows:

① Draw the training dataset from the original sample dataset. In each round, n training samples (with put-back sampling) are drawn from the original sample dataset using the Bootstrap method. A total of k rounds are performed and k training sets are obtained.

② Each time a training dataset is used to get a model, k training datasets get a total of k models.

③ For the classification problem: the k models obtained in the previous step are voted to obtain the classification results, and the mean value of the above models is calculated as the final result.

1. Adaboost

Adaboost is the abbreviation of “adaptive boosting,” which was proposed by Yoav Freund and Robert Schapire in 1995. It is a typical representative of the idea of ensemble learning with boosting idea, and the operation is performed by continuous iterations, adding a new weak learner in each round until some predefined sufficiently small error rate is reached. Adaboost method is very sensitive to noisy and abnormal data and is less prone to overfitting than most other learning algorithms. The Adaboost method operates by iterating continuously, adding a new weak classifier in each round, until a predetermined small enough error rate is reached. Each training sample is given a weight indicating the probability of being selected by a classifier into the training set, and the weights are continuously adjusted so that the Adaboost method can “focus” on those samples that are difficult to distinguish [54, 55, 56]. The detailed parameters for Adaboost are shown in Table 2.

1. Support vector machine regression

Support vector machine regression (SVR) characterizes data into a high-dimensional data feature space using a nonlinear mapping, so that the independent and dependent variables have good linear regression characteristics in the high-dimensional data feature space, which is fitted in that feature space and then returned to the original space [57, 58]. The details can be shown in the Figure 2. Given the training sample D = {(x₁,y₁),(x₂,y₂),…,(x_n,y_n)}, it is desired to learn an f(x) such that it is as close as possible to y. w and b are the parameters to be determined. In this model, the loss is zero only when f(x) is identical to y. The SVR method assumes that we can tolerate a deviation of at most ε between f(x) and y. The loss is calculated when and only when the absolute value of the difference between f(x) and y is greater than ε. This is equivalent to constructing an interval band of width 2ε centered on f(x), and if the training sample falls into this interval band, it is considered to be predicted correctly.

Figure 2.

Schematic diagram of support vector machine (SVM) regression.

In regard to the kernel functions, linear, polynomial, radial basis, sigmoid, etc., are commonly used [50, 59]. Among them, the linear kernel function has the advantages of high efficiency and wide range of applications, combined with the study of Yi et al. [60], the linear kernel function was selected in this study to meet the demand and also improve efficiency, and other parameter settings are shown in Table 2.

2.3.3 Model evaluation metrics

Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used to measure the prediction effect of nitrogen content in winter wheat plants. The smaller the values of RMSE and MAE, the more accurate the model is. R² compares the predicted values with the case where only the mean value is used. The closer the result is to 1, the more accurate the model is. The RMSE, MAE, and R² are calculated according to Bellis et al. [61].

3. Results

3.1 Statistical analysis of UAV multispectral image data features

3.1.1 Spectral features

The physiological and biochemical parameters of winter wheat plants differed across the spectral features as shown in Figure 3. The skewness and kurtosis of the Nir band reflectance were smaller than zero, showing a left-skewed and flat broad peaks distribution; while the skewness coefficient and kurtosis coefficient of the Red edge, Red and Blue bands’ reflectance were larger than zero, showing a right-skewed and sharp peaks distribution. The skewness of the Green band reflectance was less than zero and the kurtosis coefficient was greater than zero, showing a left-skewed and flat broad peaks distribution. Combined with the magnitude of the kurtosis coefficient and skewness coefficient values, there were differences in the distribution of reflectance data in different bands, indicating the different nitrogen content or other growth parameters of winter wheat plants.

Figure 3.

Statistical analysis of spectral features.

The 20 vegetation indices constructed based on spectral reflectance, the mean, standard deviation (SD), skewness coefficient, and kurtosis coefficient of TVI, RVI and MSR were greater than the others, and the coefficient of variation (the ratio of SD to mean) values were mostly around 0.20. From the data distribution characteristics, the skewness coefficient and kurtosis coefficient of VIplot, DVI, GOSAVI, and GNDVI were all greater than zero, showing a right-skewed and sharp peaks distribution; the skewness coefficient and kurtosis coefficients of RDVI, SAVI, MTVI, RVI, EXG, CARI, RERDVI, and MSR were less than zero, showing a left-skewed and flat broad peak; the skewness coefficients and kurtosis coefficients of GBNDVI, VARI, TVI, EVI, NGBDI, OSAVI, REOSAVI, and NDVI were less than zero and the kurtosis coefficients were greater than zero, showing the distribution of left-skewed and sharp peaks. In summary, different spectral features had different numerical magnitudes and distributions, but most of them show close to skew-normal distributions, indicating the similarities and dissimilarities in the response of these spectral features to wheat nitrogen content, which can provide a basis for the construction of nitrogen content model using vegetation indices.

3.1.2 Texture features

Texture features are inherent attributes of remote sensing images, especially for the high spatial resolution UAV data. The basic characteristics of these 40 texture features were analyzed and are shown in Figure 4.

Figure 4.

Statistical analysis of texture features.

The texture features of con, dis, ent, mean, and var. increased with increasing wavelength, while hom and sm decreased with increasing wavelength. Compared with the spectral feature values, the difference between the maximum, minimum, and median values of texture features was larger, which had a differential amplification effect on the construction of the nitrogen content model for winter wheat. From the perspective of data distribution characteristics, the skewness coefficient and kurtosis coefficient of 12 texture feature values, including var_780, var_685, var_550, var_450, sm_780, sm_725, hom_780, dis_685, cor_685, con_685, con_550, and con_450, were all greater than zero, showing a right-skewed and sharp peaks distribution; the skewness coefficient and kurtosis coefficient of sm_685, ent_725, ent_550, ent_450, hom_550, dis_450, cor_780, and mean_780 were less than zero, showing a left-skewed and flat broad peaks distribution; hom_725, mean_725, dis_725, sm_450, cor_450, mean_685, mean_450, hom_450, dis_780, sm_550, var_725, con_725, dis_550, ent_685, cor_725, cor_550, and con_780 texture eigenvalues had skewness coefficients larger than zero and kurtosis coefficients smaller than zero, showing a right-skewed and flat broad peaks distribution; 685hom and 780ent had skewness coefficients smaller than zero and kurtosis coefficients larger than zero, showing a left-skewed and sharp peaks distribution. Compared with the spectral features, the texture eigenvalues were more discrete and the coefficient of variation (CV) was also larger. This dispersion provides a basis for the construction of an accurate nitrogen content model for winter wheat.

3.2 Sensitive analysis of the feature responses to the nitrogen content

To screen out the sensitive characteristics of plant nitrogen content, Pearson’s correlation analysis was conducted between 25 spectral features and 40 texture features and the measured nitrogen content, and the results are shown in Figure 5. For spectral features, 25 features passed the 0.01 extremely significant level test, except for three features of 550-nm band reflectance, GNDVI, and GOSAVI; for texture features, except for 11 features, mean_450, var_725, sm_725, mean_725, hom_725, ent_725, dis_725, con_725, cor_685, and dis_780, whereas the other 29 texture features passed the 0.01 extremely significant level test. Specifically, some correlation coefficients (r) had absolute values greater than 0.5, such as 17 species including spectral reflectance of 450-nm, 685-nm, 780-nm bands and vegetation indices of TVI, VARI, REOSAVI, OSAVI, NDRGI, RVI, MSR, NDVI, VARI, CARI, RDVI, SAVI, MTVI, RERDVI, and 4 texture features, including cor_550, cor_725, cor_725, and mean_780. Moreover, there was a positive correlation between spectral features and wheat nitrogen content, and a negative correlation between texture features and plant nitrogen content. Therefore, to retain the sensitive characteristics of plant nitrogen content as much as possible, the 51 spectral features and texture features that passed the 0.01 extremely significant level test were taken as input variables for the construction of the plant nitrogen content prediction model in the next step.

Figure 5.

Correlation coefficients between spectral and texture features and nitrogen content in winter wheat plants.

3.3 Prediction of nitrogen content in winter wheat plant based on machine learning methods

Based on the 51 spectral features and texture features, BP, RF, Adaboost, and SVR regression methods were used to predict the nitrogen content in winter wheat plants. The model was first trained based on the training dataset and then evaluated using the test dataset. The measured and predicted nitrogen for the test dataset are shown in Figure 6. The machine learning models had different effects on the prediction of winter wheat plant content. From the 95% confidence interval (CI), the confidence interval span between the measured and predicted nitrogen values of BP and SVR methods was larger than that of RF and Adaboost methods, and the data distribution was more scattered, while the measured and predicted nitrogen values of RF and Adaboost methods were more close to the 1:1 line.

Figure 6.

Relationship between the predicted and measured nitrogen content in a winter wheat plant with the four different machine learning methods.

4. Discussion

4.1 Training efficiency of different machine learning methods

Under the same environmental conditions such as data segmentation, shuffling methods, cross validation, etc., as shown in Table 2, the training time of the model varies greatly. Using a computer with Intel Core i7-9700K CPU and 64 GB RAM, the SVR method takes the shortest time of 0.022 s, the RF and Adaboost methods take time of 0.778 s and 0.831 s, respectively, and the longest time is the BP method (3.122 s), which is 142 times longer than that of the SVR method. This is consistent with the results obtained in other studies, such as in Jeung et al. [51], Du et al. [48], Fernández Habas et al. [52], and Lin and Liu [62], on the prediction of flow scouring efficiency, thermal efficiency, pasture quality, and soil total nitrogen using these machine learning methods.

4.2 Effects of different machine learning methods on the nitrogen content prediction models

The model evaluation metrics for the training and test datasets from the four methods are shown in Table 3. In terms of the training dataset, the R² of the different models follow the order of Adaboost (0.988), RF (0.964), BP (0.837), and SVR (0.703); the values of RMSE and MAE were similar to that of the R². However, this is not the case for the test dataset. For R², the best model is RF (0.811), followed by Adaboost (0.791), BP (0.712), and SVR (0.663). This indicates that the Adaboost model has been overfitted compared to RF, BP, and SVR.

Methods	Datasets	RMSE (g/m^2)	MAE (g/m^2)	R^2
BP	Training dataset	3.955	3.040	0.837
	Test dataset	4.808	3.684	0.712
RF	Training dataset	1.749	1.299	0.964
	Test dataset	4.163	2.947	0.811
Adaboost	Training dataset	0.202	0.067	0.988
	Test dataset	4.437	3.207	0.791
SVR	Training dataset	5.210	4.047	0.703
	Test dataset	5.348	3.956	0.663

Table 3.

Evaluation of the models constructed by random forest (RF), Adaboost, and support vector machine (SVR) methods.

Furthermore, the model evaluation was conducted using the test datasets for the constructed model, and the relationship between the measured and predicted nitrogen values of test dataset was shown as per sample given in Figure 7. The local fitting effects of the four methods have good performance, which were mainly related to the distribution of plant nitrogen content data characteristics. In this study, the nitrogen content values were mainly concentrated in the range of 12–28 g/m², and the trained model had better prediction ability for the local values in this range, so the agreement between the measured and predicted nitrogen is high for the test datasets; overall, there was a trend of underestimation in the high nitrogen content area, while the estimation was relatively good in the low nitrogen content area. Along with the over−/−underprediction of different models demonstrated earlier, the results indicate that the efficiency and accuracy of the prediction models constructed by the RF and Adaboost methods are outstanding, which is inextricably linked to the principles of the algorithms.

Figure 7.

Curve fitting effects of the test datasets.

4.3 Transferability of the prediction model of nitrogen content in winter wheat plants

In this study, BP, RF, Adaboost, and SVR machine learning methods were used to construct the nitrogen prediction model for winter wheat, and good prediction results were achieved when both water management (rainfed and irrigation) treatments were combined. How about the transferability of the established model across different water treatments? This is an important issue regarding the generalizability of the model [63], which has practical implications for researchers who want to apply the models developed from one water management regime to another. To answer this question, the models constructed by different machine learning methods using datasets under W0 or W1 water treatments were evaluated. For example, models were first fitted using datasets from W1 treatment only and predicted onto the W0 treatment plots and vice versa (Figures 8 and 9). The prediction effects of the four methods on the nitrogen content of W0 and W1 treatments trained using W1 and W0 treatments were the same as those trained using both W0 and W1 datasets, both of which were closer to the 1:1 line for the RF and Adaboost methods. The R² of transfer prediction results for the models constructed by BP, RF, Adaboost, and SVR methods were 0.751, 0.723, 0.720, and 0.660 for the prediction of nitrogen content in W0 treatment and 0.512, 0.693, 0.612 (trained using data from W1 treatment) and 0.452 for the prediction of nitrogen content in W1 treatment (trained using data from W0 treatment), respectively. This is also the case for the RMSE and MAE. As a result, the transfer prediction ability of the plant nitrogen content prediction model constructed by RF and Adaboost methods was better than that of BP and SVR methods. Although the nitrogen content prediction model constructed in this study has good local transferability, future research is needed to test whether such models could be transferred to other wheat production areas or even to other crops.

Figure 8.

Transferability of models constructed by the four machine learning methods under rainfed (W0) and irrigation (W1) treatments.

Figure 9.

Comparison of the transferability of models constructed by the four machine learning methods for rainfed (W0) and irrigation (W1) treatments.

4.4 Mechanism analysis of the four machine learning methods for nitrogen content prediction

Machine learning is a data-driven method, which can achieve accurate prediction by fully mining the information in the dataset. Machine learning has become a research hotspot for prediction in many disciplines [48, 51, 52, 62, 63, 64, 65, 66]. However, there are differences in the design of the machine learning methods, and this study focuses on the influence of BP, RF, Adaboost, and SVR models on the prediction of nitrogen content in winter wheat. The BP method has a relatively strong learning ability, but it requires more parameters to be fitted; the model training also takes a long time, and different solvers and activation functions may affect the efficiency of the model. The training of RF and Adaboost methods is adjustable with relatively simple parameters and fast fitting speed. The SVR method can solve high-dimensional problems with strong generalization ability and relatively low dependence on the overall data, but it is difficult to determine the appropriate kernel function. Therefore, there is a tradeoff between the efficiency and accuracy of the machine learning models.

In the study, the RF and Adaboost methods are more prominent with the R² above 0.8, mainly because both methods belong to ensemble learning based on the idea of bagging and boosting, respectively. During processing of the models’ construction, a number of learners are combined to get a new learner, so as to achieve a better learning effect, which fully reflects the “group wisdom” of machine learning. In addition, both methods are extracted from the original dataset using the Bootstrap strategy and reorganized to form a subset as large as the original dataset. This means that samples inside the same subset can be recurring, and samples in different subsets can also be recurring. Moreover, unlike a single decision tree that selects an optimal feature to segment nodes after considering all features in the segmentation process, the RF method selects the optimal feature variable among these features by randomly examining certain feature variables in the base learner, similar to “democratic voting,” and this randomness makes the generalization ability and learning ability of the RF model superior to those of the individual learner. This performance has also been verified in the literature [49, 52, 53, 64]. The Adaboost method takes into account the weights of each classifier in the sampling process, similar to “elite selection,” but if the data are not balanced, the model accuracy decreases [57]. Comprehensively, the RF and Adaboost methods are more effective for plant nitrogen content prediction models when considered together.

5. Conclusions

Based on the 5-band multispectral reflectance acquired by the K6 multichannel multispectral imager, 20 vegetation indices and 40 texture features were obtained by computational analysis. Different spectral features had different numerical magnitudes and distribution characteristics with an approximate skew-normal distribution, while the texture features were more discrete compared with the spectral features. In total, 51 spectral features and texture features that passed the 0.01 significant level test were selected to construct models using the BP, RF, Adaboost, and SVR methods with validation R² of 0.712, 0.811, 0.791, and 0.663, respectively. The RF and SVR methods tend to underestimate the wheat nitrogen content, while BP and Adaboost slightly overestimated the wheat nitrogen content. When predicting the nitrogen content in winter wheat under different water treatments, the model shows a strong transferability, especially the RF and Adaboost methods. Combining R², RMSE, and MAE, the RF and Adaboost methods have better computation time, accuracy, and transferability for nitrogen content prediction in winter wheat.

However, it is undeniable that this study has limitations in exploring the transferability of the model based only on data from different irrigation treatments at the same sampling sites. Regarding the transferability of the model, the applicability between different regions, species, and years will be the next direction of our research.

Acknowledgments

This research was funded by the Henan Province Key R&D and Promotion Projects (No. 232102111030), and the Science and Technology Innovation Leading Talent Cultivation Program of the Institute of Agricultural Economics and Information, Henan Academy of Agricultural Sciences (No. 2022KJCX01). The research was also funded by the Independent innovation projects of Henan Academy of Agricultural Sciences (No. 2023ZC062).

Conflicts of interest

The authors declare no conflict of interest.