Estimation of nitrate nitrogen content in cotton petioles under drip irrigation based on spectral indices and wavelet neural network

Background: Estimating nitrate nitrogen (NO 3- -N) content in petioles is one of the key approaches for monitoring nitrogen (N) nutrition in crops. Rapid, non-destructive, and accurate monitoring of great NO 3- -N content in cotton petioles under drip irrigation is of great signicance. Methods: NO 3- -N content in cotton petioles under drip irrigation and the corresponding canopy spectral reectance of cotton plants grown in experimental plots under various N application levels were analyzed. The correlations among ‘trilateral parameters’ and six vegetation indices, and NO 3- -N content in petioles were determined. A traditional regression model of NO 3- -N content in cotton petioles under drip irrigation was established, and a wavelet neural network (WNN) model with different index numbers was developed. The WNN model was veried using independent data, and compared with the random forest algorithm , radial basis function neural network and back propagation neural network. Results: Based on the analyses of ‘trilateral parameters’ and petiole NO 3- -N content, blue edge amplitude (Db) and blue edge area (SDb) of the blue edge parameters exhibited a strong positive correlation with petiole NO 3- -N content, and the correlation coecients was 0.90. Among the blue edge parameters, the coecient of determination (R 2 ) of the Db polynomial regression equation and petiole NO 3- -N content was the highest (R 2 = 0.89), while the root mean square error (RMSE) of the linear regression model was the lowest (RMSE = 1.04). R 2 value of the traditional regression model developed using blue edge parameters and petiole NO 3- -N content signicantly increased, while RMSE value decreased when compared with those of the red edge and yellow edge parameters. Analyses results of the vegetation index developed using original spectral reectance data and the vegetation index developed using the rst set of derivative spectral reectance data and petiole NO 3- -N content, revealed that the rst derivative vegetation index, normalized difference spectral index (ND705) exhibited a strong negative correlation, with a correlation coecient of -0.90. The rst derivative vegetation index, ND705 and petiole

g/L, the R 2 value was 8.6% higher than the R 2 the rst derivative vegetation index model, in which RMSE and MAE reduced by 18.7% and 20.5%, respectively. The model was tested using independent veri cation data, and which revealed that the R 2 value of the model was 0.88, RMSE was 0.65g/L, and MAE was 0.47g/L based on the blue edge parameters, predicted value of WNN, and true value of the veri cation model, which was superior other models. The performance of the WNN model based on the blue edge parameters improved by 7.3%, and RMSE and MAE reduced by 25.2% and 30.9%, respectively when compared with those of the vegetation index model.
Conclusion: The present study demonstrated that an inexpensive approach consisting of WNN algorithm and spectrum can be used to enhance the accuracy of NO 3 --N content estimation in cotton petioles under drip irrigation, which re ects their practical application potential.

Background
Optimal management of nitrogen (N) fertilizer is of great signi cance to the improvement of yield and quality of cotton [1], and the reduction of waste and environmental problems caused by excessive N fertilizer input [2]. A reasonable amount of N fertilizer is conducive for the maintenance of a balance between vegetative growth in cotton, and promotion of N absorption and utilization e ciency [3,4]. N fertilizer is generally stored and assimilated by cotton plants in the form of nitrate nitrogen (NO 3 − -N). NO 3 − -N content varies in different parts of the cotton plant, and follows the order of petioles > stems > leaves [5]. The relationships among cotton leaves, petioles, buds, and bolls re ect the coordinated relationship between vegetative and reproductive growth in cotton plants [6], which subsequently in uence the yield and quality of cotton [7]. Therefore, petiole NO 3 − -N content is an effective parameter that re ects the overall N nutrition status of cotton, and petioles can be used as primary plant parts in the diagnose N nutrition [8,9]. Petioles also facilitate rapid determination of N nutrition status of plants to guide rational application of N fertilizer [10,11].
The traditional methods of evaluating cotton N nutrition predominantly include soil mineral N determination, laboratory analysis of plant, determination of petiole NO 3 -N, etc.; [12,13] however, the methods are associated with certain demerits such as the procedures are cumbersome, time consuming, poor timing of analyses results, and they involve destructive sampling of numerous plants [14,15]. Hyperspectral remote sensing technology has been widely used in the estimation of physiological parameters during crop growth and development due to its non-destructive, cheap, and e ciency characteristics [16]. The diagnosis of N nutrition in crops based on spectral data has made considerable progress [17]. The technique has been applied in several crops to rapidly obtain crop N nutrition status spectral indices derived from spectral [18][19][20]. Several researchers have developed crop N nutrition monitoring models based on spectral indices, and achieved high accuracy. Wang et al. [21] demonstrated that the ratio spectral index (RSI, 822,738) could be used as an indicator for monitoring N accumulation in rice and wheat leaves. Liu et al. [22] proposed a ratio vegetation index (RVI, 764,657) based on re ectance at wavelengths of 764 and 657 nm as an effective indicator of N status of oilseed rape in winter.
In addition, as the parameters associated with spectral location characteristics, 'trilateral parameters' cannot only re ect the spectral characteristics of vegetation, but are also sensitive to variations in N content [23]. The red edge parameter, which is one of the 'trilateral parameters' has been successfully used to the estimate N nutrition in a variety of crops with satisfactory results [24,25]. The red edge 'blue shift' phenomenon exists in the re ectance spectra of numerous crops. Railyan [26] and Gilbert [27] established that the position and red edge slope in triticale and maize varied constantly during the entire growing season, and were closely associated with the phonological period of crops. The red edge shifted to the long wave direction in the vegetative growth stage, and shifted to the short wave direction in the reproductive growth stage.
The spectral indices of crops can be obtained by developing linear or nonlinear relationship or learning method of arti cial neural networks, and the application of spectral indices combined with arti cial neural network algorithms to estimate N content has presented numerous research results. Feng et al.
[28] established the quantitative estimation models of N content rice canopy leaves based on adaptive differential optimization extreme learning machine, radial basis function (RBF) and particle swarm optimization BP. To estimate the N content of maize in natural environments rapidly and accurately, Xiu et al. [29] proposed a method for measuring maize N content based on wavelet energy coe cient and back propagation neural network (BP). The method improved the accuracy of corn N content estimation when compared with the regression analysis model. Wavelet neural network (WNN) [30] is a type of an arti cial neural network, which is generated by applying wavelet analysis theory to neural network theory, and exhibits strong nonlinear mapping and learning capabilities [30]. The current research on WNN covers several elds such as medicine [31], industry [32], and nance [33] and has achieved satisfactory results.

Experimental design
The eld experiment was carried out in 2019 at the teaching test site of Shihezi University, Shihezi City, Xinjiang Uygur Autonomous Region(86°02′E, 44°18′N) ( Fig. 1 a, b). Soil fertility in the 0-20 cm soil layer in the experimental plots was determined; total N was 1.13 g/kg, alkali-hydro N was 44.26 mg/kg, available phosphorus content was 19 mg/kg, available potassium content was 486 mg/kg, organic matter content was15.50 g/kg, pH value was 8.17. Lumianyan 24 cotton variety, which is middle-late maturing variety, with a growth period of approximately130 days was used as the experiment material.
Five N gradients were designed as follows: 0 kg/ha (N0), 195.5 kg/ha (N1), 299 kg/ha (N2), 402.5 kg/ha (N3) and 506 kg/ha (N4). The total amount of phosphate (P 2 O 5 ) and potassium (K 2 O) fertilizers were 109.8 kg/ha and 91.8 kg/ha, respectively. One lm, three rows, and three belts were used in the experiment. The row spacing was 76 cm and the plant spacing was 10 cm. Each treatment was repeated three times and arranged in completely randomized blocks covering a plot area of 2.25m × 15m. Cotton is was rst crop to be grown in the experimental eld, and protective rows were set around the cotton plants. Other eld management measures were in accordance with the requirements of high-yield cultivation. Fertilizers were applied with irrigation water during the cotton growth period under drip irrigation with lm.
Validation test data were obtained from a high-yield cotton eld in Shihezi University Teaching Experimental Field (Fig. 1c). The independent test cotton eld was divided into 15 plots. The total amount of N fertilizer applied was 300 kg/ha, while the total amount of P 2 O 5 and K 2 O fertilizers applied were 109.8 kg/ha and 91.8 kg/ha, respectively.

Spectral data acquisition
The key growth period of cotton was divided as follows: full bud period (65 days after sowing), initial owering period (77 days after sowing). full owering period (88 days after sowing), and initial boll stage (107 days after sowing). Analytical Spectral Devices ASD FieldSpec 3 portable spectrometer (Analytical Spectral Devices Inc., Boulder, CO, Colorado, USA) was used to obtain spectral data of the cotton canopy. The band range was 350-1075 nm, and the eld of view was 25°. Three rows of cotton plants with uniform growth in different treatment plots were randomly selected. The spectrometer probe was placed vertically downward at 25 cm above the canopy. The trigger was pulled during scanning and the spectral data obtained were automatically saved. The spectral data acquisition time was three hours. The average values of the three curves were calculated using 'ViewspecPro' software (Analytical Spectral Devices, Inc., Boulder, CO, Colorado, USA) as the re ectance values of the cell spectra.

Determination of NO 3 --Ncontent in cotton petioles
After the collection of canopy spectral data, 20 cotton plants with petioles (10 days after topping) and two leaves (10 days after topping) were randomly selected from the experimental. Cotton petioles and leaves were separated, the petioles were cut and pressed, and the sap was immediately measured using the 'LAQUA twin NO 3 meter' (a brief description of NO 3 meter is presented in Table 1).

Spectral parameter selection
Spectral indices were associated with cotton photosynthesis, soil fertility level, nutrient management, etc.
Six spectral indices and 'trilateral parameters' that are sensitive to N nutrition in cotton under drip irrigation were selected based on the spectral response characteristics of cotton canopy under drip irrigation and previous studies, as shown in Table 2. The determination of the number of hidden layer nodes is a key factor in uencing the accuracy of the WNN prediction model. Therefore, the number of hidden layer nodes is determined under the condition of meeting the model accuracy, and the compactness of the model structure is ensured to avoid redundancy. In the present study, the number of hidden layer nodes was set to ve, and the model was trained with ve, eight, 10, 12, 16, and 20 hidden layer nodes. The training error, test error, and model training time are presented in Table 3. Prediction MRE is considered minimum when the number of hidden nodes is 10. Therefore, the number of hidden nodes was set to 10, the learning rate was 0.01, the number of iterations was 1000, and the maximum allowable error was 0.001. WNN was created in MATLAB R2019b software (MathWorks, Inc. Natick, Massachusetts, USA).
After classifying, processing, and ltering the data based on the relevant theory of WNN, the parameters were initialized and the training data were input. In addition, the error value was calculated and the parameters were corrected. A WNN prediction model for petiole NO 3 --N content in cotton under drip irrigation was developed based on spectral indices through repeated training and iterations. The prediction accuracy of the model was continuously enhanced and errors were reduced. The speci c ow chart of developing WNN is illustrated in Fig 2. RF [42] is an algorithm that integrates multiple trees through ensemble learning and its basic unit is a decision tree. RF is widely used in high-dimensional data classi cation and regression. The RF algorithm was developed using in MATLAB R2019b software (MathWorks, Inc. Natick, Massachusetts, USA). The number of classi cation trees in RF algorithm was 1070.
RBF [42] can t continuous nonlinear functions, and its hidden layer adopts RBF, which will responds to input signals locally. In the present study, RBF neural network was developed using MATLAB R2019b software (MathWorks, Inc. Natick, Massachusetts, USA), and the variance parameter of RBF kernel function was set to 0.3.
BP neural network [42] is a learning algorithm of feedback networks, which re ects the input-output relationship of samples, and has strong nonlinear fuzzy approximation ability. In the present study, a BP neural network was developed in MATLAB R2019b software (MathWorks, Inc. Natick, Massachusetts, USA). The BP neural network adopts a three layer structure, with 10 hidden layer nodes, 1000 iterations, and 0.01 learning rate.

Results
Relationship between petiole NO 3 --N content and 'trilateral parameters' The correlation analyses results of NO 3 --N content and 'trilateral parameters' of cotton petioles under drip irrigation are presented in Table 4.

Discussion
Feasibility of remote sensing monitoring NO 3 --N content in cotton petioles under drip irrigation Timely and accurate monitoring of N nutrition in crops is the key to accurate application of N fertilizer [43]. The rapid development of remote sensing technology presents a potential novel method for monitoring crop nutrition [44]. Local and international researchers have used the technology to monitor plant N content and N accumulation, although research NO 3 --N in cotton petioles under drip irrigation is scarce [45]. Monitoring of petiole NO 3 --N content is a widely used approach in crop nutrition diagnosis and topdressing recommendation [44]. In the present study, the correlations among six 'trilateral parameters' and six vegetation indices, and NO 3 --N content in cotton petioles under drip irrigation was analyzed and the results revealed that large proportion of the spectral index was strongly correlated with  [49] determined the N nutrition in winter wheat by conducting hyperspectral analyses, and established that blue-violet light was sensitive to N. Stroppiana et al. [50] demonstrated that blue light was the ideal wave segment for N estimation in rice. The results obtained in the present study could be due to the variations in crop canopy structure and biomass, or the unique climatic conditions in Xinjiang, drip irrigation fertilization methods, and other factors.
Most of the researchers are more interested in the red edge parameters and pay less attention to the blue edge parameters, which cover wavelengths between 490 and 530 nm, when selecting spectral characteristic parameters. Therefore, blue edge parameters should be taken into consideration when determining N nutrition based on spectral data. The present study further demonstrated the potential of blue edge parameters in the estimation of N content in crops.

Application of neural networks in remote sensing monitoring
The R 2 value of the WNN estimation and veri cation models were relatively high, while RMSE and MAE values were relatively low, suggesting that the stability of the model is high, and the estimation capacity of the model is superior. Combined WNN maintains the advantages of arti cial neural networks and wavelet analysis, which accelerates network convergence in turn, preventing the algorithm from falling into local optimum and occasionally makes local analysis more frequent [51,52]. The RBF neural network algorithm confers the advantages of rapid training and convergence speed, strong input-output mapping ability, and strong generalization ability when compared with BP neural network algorithm. Furthermore, the results of the present study revealed that the estimation model based on the RBF neural network is superior BP neural network model [53].
Numerous studies have revealed that neural networks exhibit a great potential in learning and developing nonlinear complex relationship models, and they exhibit high tolerance for input objects . Neural networks can simulate heteroscedasticity better and have the ability to learn hidden relations in data without imposing any xed relations in the data [53,54].
The present study was conducted in experimental plots and the method used to determine NO 3 --N content in petioles caused considerable damage to plant tissues. Therefore, the number of samples analyzed was limited, which could in uence the accuracy of modeling. The sampling frequency from bud stage to owering and boll stage should be increased, and more data should be acquired to further improve the modeling capacity of the estimation model.

Comparison between WNN and RF
Most of the previous studies on crop physiological parameter estimation have demonstrated that the RF algorithm exhibits high accuracy and estimation ability, and confers the advantages of strong stability and high e ciency when compared with other modeling methods. Loozen et al. [55] used RF technology estimate the N content of a European forest canopy, which exhibited superior accuracy (R 2 = 0.62, RMSE = 0.18). To establish an e cient method for estimating winter wheat biomass, Yue et al. [56] used RF algorithm to develop a regression model of winter wheat biomass by combining spectrum, radar backscattering, vegetation index, and radar vegetation index, and the results revealed the potential application of stochastic forest algorithm in remote sensing to estimate winter wheat biomass. The RF regression algorithm has been demonstrated to result in over tting phenomenon and high test errors when compared with neural network algorithm [57]. RMSE and MAE values of WNN and RBF models based on the vegetation indices are lower than those of the RF model during model validation (Fig. 3).
The R 2 value of WNN based on blue edge parameters was higher than that of the RF model, and RMSE and MAE values were lower than those of the RF model (Fig. 4), which is consistent with previous ndings that the RF method exhibited weak prediction ability [58].

Conclusions
The present study analyzed and compared the performance of 'trilateral parameters' and vegetation       Figure 1 Study area location. a location of the study area; b eld experiment; c high yield veri cation cotton eld photographed by UAV. Note: The designations employed and the presentation of the material on this map do not imply the expression of any opinion whatsoever on the part of Research Square concerning the legal status of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. This map has been provided by the authors. Validation of predicted and measured values of NO3--N content in petioles by blue edge parameters