Introduction

High productivity with enhanced nutrient-use efficiency is one of the important desirable agronomic, economic, and environmental goals in crop husbandry (Haque, 2006). In most crops, nitrogen (N) is the most limiting nutrient for crop production (Ågren et al., 2012), and its efficient use is important for economic and ecological sustainability (Mălinaş et al., 2022). Wheat (Triticum aestivum L.) is a high N-demanding crop, where insufficient application of it may result in reduced production, whereas excessive fertilization implies wasting resources and an increase in environmental pollution (Good & Beatty, 2011). Nitrogen fertilization of wheat can account for up to 30% of the total production cost and around 50–75% of the N applied to the field is not used by the plant and is lost by leaching into the soil (Baresel et al., 2008; Hirel et al., 2011). Therefore, nitrogen is a key element required for plant growth, and is one of the most important yield-limiting nutrients in crop production in all agroecological regions of the world (Guo et al., 2019; Zhu et al., 2007). Therefore, rapid and accurate detection of the wheat nitrogen status is of great significance for growth diagnosis and precision management to achieve higher yield and better quality while also minimizing adverse environmental impacts (Yang et al., 2019).

At present, the implementation of remote sensing (RS) technology in precision agriculture provides new opportunities for the non-destructive real-time diagnosis of plant nitrogen status (Blekanov et al., 2023; Sahoo et al., 2015) and precision nitrogen management over a large area (Rabatel et al., 2017). Vegetation indices (VIs) derived from multiple spectral bands are now widely used in developing N monitoring models for various crops (Basyouni et al., 2015; Jiang et al., 2020; Ranjan et al., 2012; Xu et al., 2023). This technology contributed to the development of RS-based variable rate N recommendation approaches, which mainly use the remotely sensed crop growth information as the chief attribute of N fertilization (Späti et al., 2021).

Many researchers have attempted to derive the crop biophysical and chemical parameters using (VIs) derived from field spectral measurements (Dehghan-Shoar et al., 2023), and sensors on satellites (Jamali et al., 2023) and unmanned aerial vehicles (UAVs) (Buthelezi et al., 2023). Even though satellite-based multispectral images are useful to aid N management, farmers are still reluctant to adopt RS technologies. Satellite images normally have lower spatial and temporal resolution and, in some areas, they can be distressed by cloud cover during the image acquisition (Hunt et al., 2005). UAVs have tremendous potential for high-resolution requirements for site-specific management of weed control (Anderegg et al., 2023; Esposito et al., 2021; Jurado-Expósito et al., 2021) and N recommendation (Jiang et al., 2023; Maresma et al., 2016).

The advent of hyperspectral UAV sensors with narrow band widths made a breakthrough in crop monitoring by providing an accurate prediction of plant N levels and thereby N stress estimation. Leaf nitrogen content (LNC) of wheat was predicted by using first derivative normalized difference nitrogen index (FD-NDNI) and the first derivative ratio nitrogen vegetation index (FD-SRNI) with the help of random forest (RF) (Liang et al., 2018); the transformed chlorophyll absorption in the reflectance index/optimized soil-adjusted vegetation index (TCARI/OSAVI) on using partial least squares regression (PLSR) and modern machine learning (ML) methods, including artificial neural network (ANN), RF, and support vector machine (SVM) (Wang et al., 2021). Among the various non-parametric modeling algorithms, the RF approach performed best for estimating LNC in winter wheat using modified renormalized difference VI (RDVI) from UAV Multispectral Images (Zheng et al., 2018). Several other researchers also proposed UAV-based multispectral and hyperspectral VIs for LNC prediction in crops using different machine learning approaches (Fu et al., 2022; Jiale Jiang et al., 2019; Li et al., 2018; Liu et al., 2017; Osco et al., 2020; Xu et al., 2023; Zhang et al., 2022).

On account of potential applications of UAV hyperspectral remote sensing in the field of agriculture and the importance of N management practices, the present study emphasizes the performance of various hyperspectral vegetation indices for wheat leaf N estimation. The study conducted involves an exhaustive comparative assessment of the VIs for wheat leaf N assessment and was performed in multiple stages, (1) UAV-based hyperspectral data collection and pre-processing (2) selection of the most significant indices for leaf nitrogen prediction and (3) prediction of leaf N through ANN using selected indices as input.

Dataset

Experimental site and field sampling

An experiment was conducted on the Research Farm of ICAR-Indian Agricultural Research Institute, New Delhi with wheat crop (variety-HD 3059) during the rabi (winter) season 2021-22. The field is situated between 28°38′29.99″N to 28°38′26.58″N latitudes and 77°9′0.99″ E to 77°9′5.09″E longitudes. The experiment was conducted in a split-plot design with three irrigation types as the main plot: I1-soil moisture sensor-based irrigation, I2-CWSI-based irrigation, and I3-conventional irrigation. Under each irrigation level, there were five levels of nitrogen application i.e., 0, 50, 100, 150, and 200 kgha−1corresponding to N0 to N4, respectively as sub-plot treatment. As a basal dose, one-third of recommended N and full recommended phosphorus (P) and potassium (K) were applied as urea (46% N), diammonium phosphate (18% N and 46% P2O5), and muriate of potash (60% K2O), respectively. The remaining N was applied later in two equal split doses during the first and second irrigations. There were three replications R1, R2, and R3. The wheat crop was sown on 13th December 2021 and was harvested on 15th April 2022. The size of the sub-plot and main plot were kept as 93.6 m2 (7.2 × 13 m) and 468 m2 (7.2 × 65 m), respectively. The location map of the study area and its experimental layout is presented in Fig. 1. The study area experiences a semi-arid climate with warm long summers lasting from April to August, monsoon in July-September, and mild winter in October. The soil is sandy loam (Typic Haplustepts) having a medium to angular blocky structure and non-calcareous texture. The soil is slightly alkaline with a mean pH of 8.0, mean electric conductivity of 0.25dSm−1 and organic carbon of 0.57%. The average nitrogen-phosphorous-potassium (NPK) content of the soil is estimated as 217.1, 52.6, and 508.0 kgha−1, respectively.

Fig. 1
figure 1

Study area map showing the Geographic location of the experimental plots with different fertilization treatments. a, b The location of the ICAR- IARI campus and experimental wheat plots. c Wheat field layout

Full size image

Hyperspectral image acquisition using UAV

The hyperspectral images of the wheat crop were captured using a hyperspectral camera (Nano Hyperspec of Headwall Photonics Inc., Bolton, MA, USA) mounted on a hexacopter (Fig. 2). The UAV data acquisition was conducted on 7th February 2022 during day time between 11:00 hrs to 12:00 hrs on a sunny day within a wind speed of 2–3 ms−1, relative humidity ranging between 39–41%, and a temperature of 21 °C. The specifications of the imaging sensor used are summarized in Table 1. The flight path planning for the survey has been done using UgCS Mission planning software (UgCS, 2017) which allows the remote monitoring of the flying route, speed of the drone, height above the ground, and overlapping.

Fig. 2
figure 2

a Drone with hyperspectral camera facing nadir, b image data acquisition in wheat field

Full size image
Table 1 Major specifications of UAV hyperspectral sensor
Full size table

Methodology

Preprocessing of hyperspectral image data

The pre-processing of the acquired hyperspectral image data involved (1) conversion of digital numbers (DN) to radiance, (2) calculation of reflectance values, (3) image orthorectification (4) generation of ortho-mosaicked image for analysis (5) spectral smoothening and (6) crop area segmentation.

Raw data conversion

The acquired raw data was converted to radiance followed by reflectance using Headwall SpectralView software version 3.1.4 (Headwall Photonics, Bolton, MA, USA). The dark signal needs to be captured during image data acquisition due to varying sensor characteristics, such as sensor settings and orientation. To obtain an absolute radiometric value from the sensor it has to be calibrated since its signal is subject to influences other than the impinging light. The raw hyperspectral data cubes use the dark reference of the sensor captured pre-flight. The radiance dataset of individual scan lines was processed using the white reference generated using the reflectance data from pixels of calibration tarp placed in the field during flight. The positioning information from Global Navigation Satellite System (GNSS) receivers and Inertial Measurement Unit (IMU) synchronized with data and the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model was used for orthorectification (Santos-Rufo et al., 2020). The geo-rectified hyperspectral data cubes were stitched together and validated by overlaying with Google Earth imagery. The signal noise present in the calibrated image has been removed by applying a Savitzky–Golay (SG) smoothening filter with a second-order polynomial and the frame lengths of 5, available in image processing software ENVI 5.5 (L3 Harris Geospatial, Boulder, CO, USA) (Ge et al., 2019).

Crop area segmentation

The wheat canopy from 45 subplots was extracted using a combined approach of thresholding of hyperspectral normalized difference vegetation index (hNDVI) (Rouse et al., 1974) and spectral angle mapper (SAM) binary masks (Rejith et al., 2020). The threshold for the hNDVI was selected from the histogram representing the frequency of vegetation index values corresponding to all 45 subplots of different N treatments. A threshold value was selected halfway between the vegetation peak and the neighbouring trough. The SAM classification was further carried out using the endmembers spectra computed manually by selecting the pixels of wheat, other vegetation features, and soil. The SAM measures the spectral angle between the endmember spectra and the image spectra in n-dimensional space. Lowering the angle suggests a closer match to reference spectra. The hNDVI mask was later multiplied with SAM binary mask to generate a primary mask image. Then the morphological erosion operator (3 × 3 matrix of ones) was applied to it to remove the mixed pixels from the canopy edges. Finally, this approach generates a complete mask that was used for concealing the wheat pixels from the neighbouring soil, shadow, and grass pixels in the hyperspectral image dataset. The flow chart showing the proposed methodology adopted for the present study is given in Fig. 3.

Fig. 3
figure 3

Flow chart showing the methodology adopted for the study

Full size image

Computation of vegetation indices

After an exhaustive review of the literature, a total of 61 vegetation indices (Table 2) were selected for leaf N assessment using imaging spectroscopy (Adak et al., 2021; Chandel et al., 2019; Liang et al., 2018; Pradhan et al., 2013; Ranjan et al., 2012). It is evident from previous studies that, indices-based approaches combined with prediction techniques have yielded optimum results for N assessment. The NDVITs constructed by spectral and textural information of UAV multispectral imagery performed well in wheat leaf nitrogen concentration (LNC) monitoring (Fu et al., 2022). Vegetation indices such as NDRE or RERVI composed of near-infrared and red edge bands of UAV—multispectral scanner (UAV-MSS) data showed better results for estimating crop nitrogen status in rice plants (Li et al., 2018). Four representative vegetation indices such as the Visible Atmospherically Resistant Index (VARI), Red edge Chlorophyll Index (CIred−edge), Green band Chlorophyll Index (CIgreen), Modified Normalized Difference Vegetation Index with a blue band (mNDblue) were derived from the multi-angular UAV-MSS images for estimating LNC, plant nitrogen concentration (PNC), leaf nitrogen accumulation (LNA), and plant nitrogen accumulation (PNA) of wheat canopies (Lu et al., 2019).

Table 2 List of spectral indices commonly used for the N estimation (Rλ= reflectance at λ wavelength)
Full size table

To compare the performance of different spectral indices for leaf N content at various N fertilizer treatments, twenty-five points were randomly selected within each subplot by excluding an inner buffer distance of 1m from the borders. This approach facilitates the exact measurements of crops and avoids the areas of bare soil without crops (Maresma et al., 2016). At last, the mean of the values extracted at each point within a subplot was used to determine the best vegetation index/es to predict LNC. Three sets of top-most leaf samples were collected from each plot for N determination and their mean was taken. The N concentration of leaf samples was estimated using the Micro-Kjeldahl method (Guebel et al., 1991; Ranjan et al., 2012).

Selection of suitable indices

Selection of optimal spectral bands through R2

The R2 value for all 61 derived indices was calculated for the growth stage under study. R2 value equal to or higher than 0.7 was considered as optimum for N assessment. In this approach, an index-by-index linear correlation matrix was computed to interpret redundant information. Further, Pearson correlation coefficient (R) values have been squared (R2) to remove inverse correlation.

Selection of optimal spectral bands through PPR-VIP statistics

Projection pursuit regression (PPR) is a multivariate statistical technique comprising of three-layered architecture with input, hidden, and output layers. Unlike the linear regression approach, which assumes a direct linear combination of the predictor variables accumulating the cost of inaccurate predictions; PPR is a non-parametric regression algorithm. The algorithm helps in finding the most suitable model for prediction problems. Projection ‘directions’ or ‘strengths’ between input and hidden neurons; hidden and output neurons and the activation function are calculated by minimizing the error using least square criteria. The mathematical formulation of the model is given by Eq. 1.

$$L= \sum _{i=1}^{q}{[{y}_{i}-\sum _{k=1}^{m}{\beta }_{ik}{f}_{k}{(\alpha }_{k}^{T}x)]}^{2}$$
(1)

where, \({\alpha }_{k}^{T}=[{\alpha }_{k1},\dots \dots ,{\alpha }_{kp}]\) is the projection directions given by weights between the input and the hidden layers; \({\beta }_{k}=[{\beta }_{1k},\dots \dots ,{\beta }_{qk}]\) is the projection directions given by weights between the hidden layers and the output layer; \({f}_{k}\) is the activation function; \({y}_{i}\) is the response variable which is modelled as a weighted linear combination ofthe activation function. PPR learns layer by layer and calculates the weights by least square estimation.

Variable importance Projection (VIP) is a statistical technique for synthesizing the contribution of predictors and response variables. This measure is highly dependent on the accuracy and importance of the fitted model (here, PPR). It summarizes the contribution of a variable to the model. The indices with the highest VIP values (centers of the peaks observed) have been selected by keeping the threshold value as 0.5. The equation is given as:

$$\sqrt{n\frac{\sum _{j=1}^{a}{b}_{j}^{2}{t}_{j}^{T}{t}_{j}{\left(\frac{{w}_{kj}}{\left|\left|{w}_{j}\right|\right|}\right)}^{2}}{\sum _{j=1}^{a}{b}_{j}^{2}{t}_{j}^{T}{t}_{j}}}$$
(2)

where \(b\) is the regression coefficient, \({w}_{j}\) is the weight vector, \({t}_{j}\) is the score vector for the \({k}^{th}\) element.

Selection of optimal spectral bands through VIF statistics

The Variance Inflation Factor (VIF) helps in measuring the multicollinearity in regression algorithms. It quantifies the inflation of variance of estimated coefficients. Inflation factors were calculated by regressing the predictor with every other predictor variable considered for the model followed by the computation of R-squared values. The VIF for the \({i}^{th}\) predictor is given by:

$$VIF= \frac{1}{1-{R}_{i}^{2}}$$
(3)

where \({R}_{i}^{2}\) is the r-square value obtained.

Artificial neural network (ANN)

ANN mimics the functionality of a human brain where neurons work as the most fundamental unit to perform complex tasks. It is a three-layered architecture including input, hidden, and output layers, along with an activation function that ignites the most significant neuron as the output (Fig. 4). The fully connected structure is managed by weights from the input to the hidden layer followed by the hidden to the output layer (Fig. 5). ANN being a machine learning measure, can learn from the data and predict accordingly. The data traverses through the hidden layer which transforms the data into a usable form to generate the result, that is it helps in extracting useful information from raw data by generalizing the unknown facts.

Fig. 4
figure 4

Basic architecture of artificial neural network (ANN)

Full size image
Fig. 5
figure 5

Single-layer perceptron

Full size image

For every input \({x}_{i,}\) the corresponding weight value \({w}_{i}\) is multiplied indicating the strength of the connection. The one with a higher influence on the output value is triggered.

$$\sum =\left({x}_{1}\times {w}_{1}\right)+\left({x}_{2}\times {w}_{2}\right)+\dots \left({x}_{n}\times {w}_{n}\right)$$
(4)

where, \(x=[{x}_{1},{x}_{2},{x}_{3}\dots {x}_{n}]\) and \(w=[{w}_{1},{w}_{2},{w}_{3}\dots {w}_{n}]\) are the row vectors belonging to input and the weights respectively. Therefore, the dot product is given by:

$$x \cdot w = \left( {x_{1} \times w_{1} } \right) + \left( {x_{2} \times w_{2} } \right) + \ldots \left( {x_{n} \times w_{n} } \right)$$
(5)
$$\sum = x\cdot w$$
(6)

Bias acts as an offset for the activation function and produces the output value.

$$z=x\cdot w+b$$
(7)

This intermediate value generated is passed to a non-linear activation function which governs the learning speed of the network. There exists a multitude of activation functions, depending upon the problem. One of the simplest ones is the sigmoid activation function, \(\hat{y}\) gives the predicted value and \(\sigma\) is the activation function, and it is computed as,

$$\hat{y}={\upsigma }\left(\text{z}\right)=\frac{1}{1+{e}^{-z}}$$
(8)

The training mechanism includes back-propagating the error by computing the gradient values with respect to the weight. The mean square error is calculated by the difference between the actual values (\(y_{i}\)) and the predicted values (\({\hat{y}}_{i}\)) of error,

$${MSI}_{i}={({y}_{i}-{\hat{y}}_{i})}^{2}$$
(9)

The cumulative error for the entire training set is termed as loss function, which is calculated as,

$$MSE= \frac{1}{n}\sum _{i=1}^{n}{({y}_{i}-{\hat{y}}_{i})}^{2}$$
(10)

Later, the weights can be optimized and hyper-parameters such as the minimum error, number of epochs, and learning rates can be fixed. A training: validation: testing ratio of 70:15:15 is used to measure the prediction accuracy of the ANN model.

Results and discussions

Variability in the measured leaf nitrogen concentration (LNC)

The LNC is one of the major indicators used to describe the leaf nitrogen status in different crops. Figure 6a shows the variation in LNC at different levels of nitrogen and water stresses. Figures 6b, c are the maps showing the nitrogen stress levels and LNC. As an observation, for every irrigation level, the LNC shows an increasing trend with a decrease in nitrogen stress. The average value of LNC corresponding to N0 is 4.30% and for N4 treatment is 5.28%. The LNC ranges from 4.19 to 5.12% in I1, 4.22 to 5.48% in I2, and 4.48 to 5.24% in I3 irrigation treatments. No significant change was observed with varying levels of irrigation. A similar trend is detected for the wheat plant in the case of LNC at different water and nitrogen stresses for the tillering stage (Ranjan et al., 2012). At the recommended value of 150 kgha−1 (N3), the LNC shows a maximum value of 4.90% in I1 and 4.69% in I2. When it comes to zero where no N application was done, the LNC varies between 4.19% in I1 to 4.48% in I3. The variation in irrigation levels affects more on plant nitrogen accumulation (PNA) than LNC (Ranjan et al., 2012). The overall results of LNC indicate that the different N stress levels play an important role in the N status in the leaves of wheat plants.

Fig. 6
figure 6

a LNC variation at different levels of nitrogen and water stresses; b graded N levels; and c measured leaf nitrogen

Full size image

Optimal indices for leaf N prediction

The pre-processed hyperspectral imagery of the wheat plots is shown in Fig. 7a. The wheat canopy area of the study site was successfully segmented using a combined approach of hNDVI threshold and SAM binary masks as expounded in the previous section. The crop area of 45 wheat plots is shown in Fig. 7b. These data were used for generating 61 vegetation indices for analyzing the plant N status using LNC.

Fig. 7
figure 7

a UAV hyperspectral imagery of the experimental plot; b masked vegetation

Full size image

All the indices summarized in Table 2 are sensitive to plant N and hold significance for predicting the LNC using the feature importance score. A large number of vegetation indices were used to explore the potential of optimizing feature variables combined with machine learning models for predicting N content. The graph showing the feature importance scores for all 61 vegetation indices derived from R2 and VIP is shown in Fig. 8. The feature importance scores of R2 and VIP help in ranking the 61 indices for selecting the best ones. The GMI1 performed the best with an R2 score of 0.81 while the NPQI performed the least with a R2 score of 0.01. In the case of VIP, the best index is MSAVI with a performance score of 0.72 and the lowest performance was shown by BNI.

Fig. 8
figure 8

Variable importance ranking for vegetation indices derived from R2 and VIP

Full size image

The selection of important spectral variables for calculating VIF scores was done based on threshold criteria, R2 > 0.7, and VIP > 0.5. On applying these thresholds, 39 and 15 indices were generated from R2 and VIP respectively. Then the VIF technique was applied to these two sets of data for generating the optimal indices. The VIF generates 15 and 13 indices each from these two sets, out of which 13 were found to be common and they were selected as the most suitable indices for N prediction. The GI and TVI were excluded from the final list of suitable indices since they were only obtained from the VIP calculation. The VIF score of the 13 suitable indices is shown in Fig. 9.

Fig. 9
figure 9

Graph showing the VIF scores obtained for selected indices

Full size image

Among the selected 61 indices, 13 indices showing the best correlation with measured leaf N are taken for Leaf N prediction. They are NDCI, GMI2, RI_half, PSSRa, SR_900, RI_1db, SR705, Ctr1, RI_3db, RVI3, GMI1, ZM, and MSAVI. The selected indices belong to Red-edge spectral region (0.7 and 0.8 µm) which is found sensitive toward LNC (Raj, 2021). The spectral regions 350–710 nm and 740–1100 nm are equivalently considered highly sensitive to LNC at different N stresses (Ranjan et al., 2012). In literature, the N deficiency in corn canopies was successfully derived by an increase in red reflectance and a decrease in NIR reflectance (Walburg et al., 1981). The red (671 nm), NIR (780 nm) and the spectral region between 550 and 710 nm are relatively sensitive to plant N levels(Stone et al., 1996; Zhao et al., 2005). Since the visible bands are strong absorbents of chlorophyll especially in red and blue regions (Hansen & Schjoerring 2003; Pablo J. Zarco-Tejada et al., 2001), the optimal indices generated from the present study contain bands from visible and red-edge regions thus proving the ability to evaluate crop growth and N status. The vegetation indices developed by selecting spectral bands from visible regions show strong absorption by chlorophyll and hence it proved to be very useful in predicting the nitrogen status in plants (Ranjan et al., 2012). The indices such as Ctr1, GMI-1, GMI-2, NDCI, SR705, ZM and PSSRa are used for examining the Chlorophyll content in plants (Broge & Leblanc 2001; Gitelson & Merzlyak 1994). The RVI3 indicates the leaf nitrogen accumulation (Zhu et al., 2008). Further, MSAVI, RI-1dB, RI-3db, and RI-half are used for analyzing the vegetation sensitivity of plants. And finally, the indices such as SR_900 are used for measuring above-ground green biomass (Rouse et al., 1974). The N along with other macro-nutrients such as phosphorous (P) from soil are significant in the synthesis of chlorophyll during photosynthesis (Fredeen et al., 1990). So, N is one of the major constituents of chlorophyll which is closely associated with leaf color, crop yield, and growth (Fageria et al., 2010). Thus it is demonstrated that the optimal indices selected from the present study show a strong relationship with the leaf N content in plants. Figure 10 and 11 show the maps of optimal vegetation indices and Table 3 summarizes the characteristics by which the suitable indices are influenced. The maximum and minimum values corresponding to each index are specifically seen in plots with N4 and N0 treatments. The R2 score of these 13 indices ranges from 0.66 to 0.81, which implies that their index values are best correlated with LNC. The highest VIF score was shown by both ZM and RI_half while the least score was shown by NDCI.

Fig. 10
figure 10

Vegetative Indices generated using UAV hyperspectral image. a Ctr1; b GMI-1; c GMI-2; d MSAVI; e NDCI; f PSSRa

Full size image
Fig. 11
figure 11

Vegetative Indices generated using UAV hyperspectral image (Continued). g RI-1dB; h RI-half; i RI-3db; j RVI3; k SR705; l SR_900; m ZM

Full size image
Table 3 Categorisation of selected indices based on different plant characteristics
Full size table

Leaf nitrogen prediction

ANN architecture with 13 input, 10 hidden neurons, and an output node has been constructed to train the network for the prediction of N content. The training: validation: testing ratio is 70:15:15 with an R-square value of 0.97, 0.84, and 0.86 respectively. The regression plots shown in Fig. 12 illustrate the network output plotted versus the targets as open circles. The best linear fit is indicated by a dashed line and the solid line represents the perfect fit.

Fig. 12
figure 12

Regression plots for N prediction using ANN

Full size image

Further, the best validation performance is obtained at 0.067 mean square error (MSE) at 5 epochs. The result seems reasonable as the test set error and the validation set error have similar characteristics therefore any significant overfitting has not occurred [Refer to Fig. 13a]. The error histogram is plotted between the observed and predicted N values, after training the neural network, where the y-axis represents the number of samples in a particular bin, as shown in Fig. 13b. From Fig. 13c, the final value of the gradient coefficient at epoch number 11 is 0.0114 approximated with minimum error, also the gradient value keeps on decreasing with increasing the epoch.

Fig. 13
figure 13

Performance evaluation using ANN for N prediction. a Best validation performance, b error histogram, c training set histogram

Full size image

The N-predicted map generated using the ANN is shown in Fig. 14. The predicted LNC ranges from 2.53 to 6.51%. The current approach of N prediction significantly reduces the salt and pepper noise pertaining to the raw image. The noise is still observed in the overlapping regions due to the averaging of pixel values during ortho-mosaicking. Also, less salt and pepper noise and misclassification were reported for random forest (RF) (Wang et al., 2021) and deep convolutional neural network (CNN) classification (Zhong et al., 2020)of UAV data. This method presented is an alternate approach for the UAV-based imaging spectroscopy for predicting wheat leaf nitrogen using spectral modelling technique (Sahoo et al., 2023).

Fig. 14
figure 14

Leaf N prediction using ANN. a Measured LNC; b Predicted LNC

Full size image

Conclusion

Leaf N assessment for the winter wheat crop concerning different nitrogen levels was successfully evaluated for the considered experimental field. A UAV-based hyperspectral push-broom scanner system seems to be capable of acquiring high-resolution spatial and spectral data for N studies in precision farming. However, the side overlap regions after mosaicking the scanlines is still an area of concern but later be tackled for the prediction of N values. The hyperspectral imagery of ultra-high spatial resolution was captured over the experimental wheat fields under varying irrigation and nitrogen treatments. After pre-processing and removal of spectral noises, the crop area is extracted for prediction of N values to avoid the intervening effect of other features such as impervious areas, soil, etc. Abrupt increase or decrease in reflectance values has been observed in the overlapping areas in multiple scan lines to generate the mosaiced image, which is decreased using ANN as a prediction model. About 61 vegetative indices have been computed, out of which 13 most suitable indices were extracted using a combined approach of R-Square and VIP followed by the VIF technique. These optimal indices were used in the ANN model for the prediction of leaf N values for the entire field. The prediction model shows an R-square value of 0.97, 0.84, and 0.86 for training, validation, and testing respectively, suggesting high prediction accuracy. The generated N map seems to be similar to the ground-truth values of N produced for every plot making the proposed approach viable for the same. The approach seems to be fast, accurate, and suitable for N prediction, further, it may be upscaled for the farmer’s field level with the availability of field data. The generated spatial map may be used by a sprayer drone for site-specific nitrogen application to promote sustainability and economical use. A similar technique may be deployed for other crop-related parameters, such as chlorophyll and leaf area index (LAI), etc. as well.