*Article*

## **Using Digital Cameras on an Unmanned Aerial Vehicle to Derive Optimum Color Vegetation Indices for Leaf Nitrogen Concentration Monitoring in Winter Wheat**

**Jiale Jiang 1,2,3,4, Weidi Cai 1,2,3,4, Hengbiao Zheng 1,2,3,4, Tao Cheng 1,2,3,4, Yongchao Tian 1,2,3,4, Yan Zhu 1,2,3,4, Reza Ehsani 5, Yongqiang Hu 6, Qingsong Niu 6, Lijuan Gui 6 and Xia Yao 1,2,3,4,\***


Received: 29 September 2019; Accepted: 13 November 2019; Published: 14 November 2019

**Abstract:** Commercially available digital cameras can be mounted on an unmanned aerial vehicle (UAV) for crop growth monitoring in open-air fields as a low-cost, highly e ffective observation system. However, few studies have investigated their potential for nitrogen (N) status monitoring, and the performance of camera-derived vegetation indices (VIs) under di fferent conditions remains poorly understood. In this study, five commonly used VIs derived from normal color (RGB) images and two typical VIs derived from color near-infrared (CIR) images were used to estimate leaf N concentration (LNC). To explore the potential of digital cameras for monitoring LNC at all crop growth stages, two new VIs were proposed, namely, the true color vegetation index (TCVI) from RGB images and the false color vegetation index (FCVI) from CIR images. The relationships between LNC and the di fferent VIs varied at di fferent stages. The commonly used VIs performed well at some stages, but the newly proposed TCVI and FCVI had the best performance at all stages. The performances of the VIs with red (or near-infrared) and green bands as the numerator were limited by saturation at intermediate to high LNCs (LNC > 3.0%), but the TCVI and FCVI had the ability to mitigate the saturation. The results of model validations further supported the superiority of the TCVI and FCVI for LNC estimation. Compared to the other VIs derived using RGB cameras, the relative root mean square errors (RRMSEs) of the TCVI were improved by 8.6% on average. For the CIR images, the best-performing VI for LNC was the FCVI (R<sup>2</sup> = 0.756, RRMSE = 14.18%). The LNC–TCVI and LNC–FCVI were stable under di fferent cultivars, N application rates, and planting densities. The results confirmed the applicability of UAV-based RGB and CIR cameras for crop N status monitoring under di fferent conditions, which should assist the precision managemen<sup>t</sup> of N fertilizers in agronomic practices.

**Keywords:** leaf nitrogen concentration; leaf nitrogen accumulation; unmanned aerial vehicle (UAV); digital camera; vegetation indices

## **1. Introduction**

Nitrogen (N) is a component of many important compounds in plants, and thus plays an important role in plant growth [1,2]. Plant growth dominantly depends on the N supply [3]. A deficiency in N would reduce crop photosynthesis, whereas higher rates of N fertilization do not necessarily improve crop yield and can lead to serious water pollution [4–6]. Furthermore, leaf N concentration (LNC) is related to the photosynthetic capacity of leaves, and thus allows N fertilizer applications and grain quality to be modeled [7,8]. Therefore, timely quantification of LNC is a prerequisite for fertilization guidance and environmental quality [9,10].

Unmanned aerial vehicle (UAV) platforms have become a promising approach in precision agricultural assessments because they enable the non-destructive measurement of crop growth status, with a very high spatio-temporal resolution [11,12]. Due to their advantages of low cost, light weight, convenient operation, and simple data processing, digital cameras have been commonly deployed on UAVs in crop phenotype research [13]. Compared to other sensors, digital cameras can operate successfully in a range of working environments. Given that adequate image exposure can be set based on the weather conditions, data can be collected under both sunny and cloudy conditions [13]. Therefore, color images can be instantly acquired for researchers and farmers to monitor crop growth status [14].

A consumer-grade RGB camera is an "o ff-the-shelf" device, with red, green, and blue channels. Because each pixel value in color images can be calculated from the digital number (DN) values of specific bands, color indices can be extracted to accentuate a particular vegetation greenness and identify the vegetation features [14,15]. Hunt et al., [16] used the normalized green–red di fference index (NGRDI) from RGB images to estimate the biomass of corn, alfalfa, and soybean, and found a linear correlation between the NGRDI and biomass. Kawashima and Nakatani [17] used a video camera to analyze the color of wheat leaves for estimating chlorophyll content. Woebbecke et al. [18] investigated the capability of several color indices to distinguish vegetation from the background, and found that the excess green vegetation index (ExG) could provide a near-binary intensity image outlining a plant region of interest. Moreover, color indices from RGB cameras containing a large amount of information regarding crop status can be used to estimate the vegetation fraction, plant height, biomass, and yield [19–21]. However, many vegetation indices (VIs) proposed for crop status monitoring contain the near infrared (NIR) bands [22–24]. Therefore, RGB cameras with a Bayer-pattern array of filters have been modified by replacing either the blue or red channel with a NIR channel to obtain color near-infrared (CIR) images [25]. Based on a newly-developed digital CIR camera system, Hunt et al., [25] found a strong correlation between the green normalized di fference vegetation index (GNDVI) and leaf area index (LAI) in winter wheat. This CIR camera system has also been used to assess winter crop biomass [26]. Four VIs, the normalized di fference vegetation index (NDVI), enhanced NDVI (ENDVI), GNDVI, and ExG, derived from UAV-based RGB and CIR images, have been shown to be reliable to assess experimental plots [27].

Previous studies have indicated that it is feasible to estimate crop growth status using RGB and CIR images, but few studies have investigated their usability for N status monitoring [12,28]. Firstly, the capability of digital cameras to monitor wheat LNC at di fferent growth stages remains poorly understood. Since the composition of canopy components (e.g., leaves, tassels) and background materials (e.g., soil) varied sharply during whole growth stages of winter wheat [29], the performance of digital cameras on estimating LNC is necessary to be tested across di fferent growth stages. Secondly, it is crucial to investigate the saturation problem of VIs, with varying LNC values. Because the application of N fertilizer has increased recently in China, it is important to e ffectively monitor LNC under middle to high application levels. Thirdly, the capability of digital cameras for LNC estimation under di fferent conditions is unclear. The relationships among VIs and LNC in cereal crops have been investigated in terms of the mechanisms involved based on hyperspectral remote sensing. Many studies have proposed e ffective VIs that can be adjusted to variations in growth stage [30] and geographic

location [31], and that can reduce the effects of the soil background [32]. Therefore, there is also a need to assess the capability of digital cameras for estimating wheat LNC under different conditions.

The overall objective of this study was to evaluate whether digital cameras mounted on UAVs could be applied to monitor LNC in winter wheat. Five typical VIs derived from RGB images and two widely used VIs derived from CIR images were selected to estimate LNC. Additionally, we developed the true color vegetation index (TCVI) and the false color vegetation index (FCVI) from RGB and CIR images, respectively. Experiments with different wheat varieties, planting densities, and N application rates were conducted in the field to: (1) quantify the relationship between LNC and the VIs from digital imagery at different growth stages, (2) evaluate the saturation sensitivity of the VIs under various LNC levels, and (3) validate the applicability of the LNC estimation models under different treatments.

#### **2. Materials and Methods**

#### *2.1. Study Site and Experimental Design*

The study site was located in Rugao City, Jiangsu Province, China (120◦45 E, 32◦16 N), as shown in Figure 1. The regional annual precipitation of this area is around 927.53 mm, with an average annual temperature of 16.59 ◦C. Two field experiments using winter wheat (*Triticum aestivum* L.) were designed that included three N application rates, two planting densities (D), and two varieties (V) in two growing seasons (Table 1). In each experiment, a split design was used with three replications and there were 36 plots, each with a plot area of 35 m<sup>2</sup> (Figure 1). The basal fertilizer included 120 kg/ha P2O5 and 120 kg/ha K2O and there were three N application rates (0, 150, and 300 kg/ha as urea) applied at the end of October 2013/2014. Compound fertilizer was applied in early March 2014/2015, including N applications at the three different rates. The N fertilizers were applied in 50% as basal fertilizer at the sowing day and in 50% at the jointing stage. All other agronomic managemen<sup>t</sup> was undertaken according to local wheat production practices.

**Figure 1.** The study site of field experiments with two winter wheat varieties (V1, V2), two planting densities (D1, D2) and three nitrogen application rates (N0, N1, N2) in Rugao City, Jiangsu Province, China.


**Table 1.** Experimental Designs and Sampling Dates.

#### *2.2. Data Acquisition and Processing*

#### 2.2.1. Color Images from Unmanned Aerial Vehicle (UAV)

In this study, an eight-rotor ARF-MikroKopter UAV (Figure 2a) was used as the platform for the UAV-camera system and its specifications are listed in Table 2. A Canon 5D Mark III (Canon Inc., Japan) commercial digital camera (Figure 2b) was mounted on the UAV and took RGB images in continuous mode. The CIR camera was modified from a Canon SX260HS camera (Canon Inc.) by changing the original red channel to a near-infrared channel. The main parameters of the two cameras are described in Table 3. An MC-32 remote control module and a ThinkPad laptop were used to control the autonomous UAV flight. During each flight, the camera was fixed on a two-axis gimbal, with the lens positioned vertically downward at 50 m altitude. Considering the lighting conditions, the exposure time and shutter speed were fixed for each campaign. UAV campaigns were conducted at noon in clear weather under stable light conditions. The spatial resolution of the RGB and CIR imagery was 3 cm. The acquisition dates of these images are listed in Table 4.

**Figure 2.** The unmanned aerial vehicle (UAV) camera system: (**a**) ARF-MikroKopter UAV, (**b**) normal color (RGB) camera (Canon 5D Mark III) and (**c**) color near-infrared (CIR) camera (Canon SX260HS).

**Table 2.** Specifications of ARF-MikroKopter Unmanned Aerial Vehicle (UAV).



**Table 3.** Main Parameters of Normal Color (RGB) and Near Infrared (NIR-G-B) Cameras.

**Table 4.** Acquisition Dates of Unmanned Aerial Vehicle (UAV)-Based Digital Camera Images.


Before image pre-processing, the original digital images were screened. We selected images with a heading overlap rate of ~70% and a side overlap rate of ~30%, and excluded images with excessive repetition. The selected images were then pre-processed, including a lens distortion correction, image mosaicking, image registration, and ortho-rectification. First, lens distortion was corrected based on the Brown Model and the correction coefficients were calculated by an Agisoft Lens. Second, image mosaicking was conducted in Photoscan (Airsoft LLC, Russia). Third, image registration was referred to ground control points (GCPs) in the experimental area (see Figure 1). The GCPs were painted on a road surface as black annuluses, with inner and outer diameters of 10 and 50 cm, respectively. The geographic coordinates of the GCPs were determined using a real time kinematic (RTK) GPS system, with an error less than 2 cm in the horizontal direction and less than 3 cm in the vertical direction. Finally, ortho-rectification was automatically performed using Photoscan. Figure 3 shows the processed images from RGB and CIR cameras at the four growth stages.

**Figure 3.** The digital images from normal color (RGB) (**<sup>a</sup>**–**d**) camera and color near-infrared (CIR) (**<sup>e</sup>**–**h**) at the booting (**<sup>a</sup>**,**<sup>e</sup>**), heading (**b**,**f**), anthesis (**<sup>c</sup>**,**g**) and filling (**d**,**h**) stages. Red-green-blue and near infrared (NIR)-red-green channels are presented as RGB for RGB and CIR cameras.

#### 2.2.2. Determination of Leaf N Concentration (LNC)

Common measures of crop LNC are either area-based (LNCarea, g/m2) or mass-based (LNCmass, %). LNCarea is the N mass per unit leaf area and LNCmass is the ratio of N mass to leaf dry mass. These two measures can be converted between each other through the leaf mass per area (LMA), i.e., LNCarea = LNCmass × LMA [33]. Therefore, LNCarea should be determined based on both LNCmass and LMA measures, while LNCmass can be measured directly in the laboratory. Given its strong connection with photosynthetic capacity [34] and its widespread use in fertilization managemen<sup>t</sup> [35,36], LNCmass has received more attention and has been estimated from remotely sensed data more often than its counterpart LNCarea [37–40]. The term LNC hereafter refers to LNCmass.

Ground destructive samplings were taken at critical growth stages of winter wheat (Table 1) on the same dates as the UAV campaigns. Thirty hills of wheat plants were randomly cut above the ground surface of each plot and separated into leaves, stems, and panicles. All leaves were oven-dried at 105 ◦C for 30 min and then at 80 ◦C until a constant weight. We then grounded and stored all samples in plastic bags for chemical analysis. The LNC was determined based on the micro-Keldjahl method [41] with SEAL AutoAnalyzer 3 HR (SEAL Analytical, Ltd., German).

#### 2.2.3. Derivation of Vegetation Indices (VIs)

The main VIs derived from digital camera images are summarized in Table 5. For RGB cameras, five widely used color indices were chosen in this study: the NGRDI [42], Kawashima index (IKaw) [17], red green ratio index (RGRI) [43], visible atmospherically resistance index (VARI) [42], and ExG [18]. For the CIR camera, we used the GNDVI [44] and enhanced normalized di fference vegetation index (ENDVI), because GNDVI is related to chlorophyll concentration, while the ENDVI was recommended by the company that manufactured the modified cameras (www.maxmax.com) and also can monitor vegetation vigor [27]. The VIs could be categorized into two groups according to the number of channels used. Some were constructed by two channels, such as NGRDI, IKaw, RGRI, and GNDVI. The others (i.e., VARI, ExG, and ENDVI) were established using three channels.


**Table 5.** Formulas and References of Possible Vegetation Indices (VIs).

1 TCVI is the True Color Vegetation Index derived from RGB images. 2 FCVI is the False Color Vegetation Index from CIR images. They are newly conducted in this paper.

Color vegetation index (CVI) derived from digital imagery is often calculated as ratios of DN values. However, the previous VIs listed in Table 5 were proposed without considering background material, such as soil in the field. Previous studies have suggested that the influence of the soil background can be reduced by adjusting the VIs with a term representing soil brightness [22,28]. Based on the ratio form and an equivalent soil adjustment, the commonly used CVIs from digital cameras can be expressed as:

$$CVI = \frac{(1+L)(a\_1R + a\_2G + a\_3B)}{(a\_4R + a\_5G + a\_6B + 255 \ast L)}\tag{1}$$

where *G* and *B* are the DN values of the green and blue channels, respectively; *R* represents the red component of RGB imagery or the near-infrared component of CIR imagery; *ai* is the coe fficient of each channel; and *L* is a soil background adjustment parameter.

To explore the capability of digital cameras for estimating LNC in wheat, we constructed new CVI from both RGB and CIR images. They were determined by optimizing the values of *ai* and *L* based on the cost function *J* defined as:

$$J = 1 - \frac{\sum\_{i=1}^{n} \left( LN\mathbb{C}\_{m,i} - LN\mathbb{C}\_{p,i} \right)^2}{\sum\_{i=1}^{n} \left( LN\mathbb{C}\_{m,i} - \overline{LN\mathbb{C}\_{m,i}} \right)^2} \tag{2}$$

where *LNCm* is the measured LNC. *LNCp* was predicted by the best fitted function of CVI and *LNCp* is the average value of *LNCp*. According to the imaging principle of a Bayer filter, we set the value of *ai* as an element of {−2, −1, 0, 1, 2}. Based on the setting rules of a soil background adjustment parameter in [22], *L* ranges from 0 to 1, with intervals of 0.1. The value of *L* varies by the amount or cover of green vegetation: in very high vegetation regions, *L* = 0; and in areas with no green vegetation, *L* = 1. Given that the maximum value of the DN was 255, *L* was multiplied by 255 in the denominator. All data from Exp. 2 were used in Equation (2). All possible combinations of variables (i.e., *ai* and *L*) were traversed, and the corresponding values of *J* were compared. The values of *ai* and *L* in Equation (1) that yield the best *J* were the optimal variables. Consequently, the TCVI from RGB images and the FCVI from CIR images were determined as follows:

$$TCVI = \frac{(1 + 0.4) \* (2R - 2B)}{(2R - G - 2B + 255 \* 0.4)} \tag{3}$$

$$FCVI = \frac{(1+0.5) \ast (2NIR + B - 2G)}{(2G + 2B - 2NIR + 255 \ast 0.5)} \tag{4}$$

#### *2.3. Data Analysis and Evaluation*

The channel information from digital cameras that constituted the VIs was first analyzed based on the data from Exp. 2. We analyzed how the DN values of di fferent channels changed with the changes of LNC in wheat. To compare and evaluate the performances of di fferent VIs for estimating LNC, the quantitative relationship between the VIs and LNC was analyzed at di fferent growth stages. For all growth stages of winter wheat, the LNC–VI models from both RGB and CIR images were calibrated and validated with a 10-fold cross-validation procedure using the data from Exp. 2. The whole dataset was randomly divided into three equal-sized sub-datasets with two sub-datasets used as the calibration (training) dataset and the rest as the validation (testing) dataset. The process was repeated 10 times [45]. For the CIR cameras, the estimation models were also validated independently with the data from Exp. 1. The performance of the di fferent VIs and models were evaluated using the determination coe fficient (R2) and the relative root mean square error (RRMSE).

$$R^2 = 1 - \frac{\sum\_{i=1}^{n} (O\_i - P\_i)^2}{\sum\_{i=1}^{n} (O\_i - \overline{O\_i})^2} \tag{5}$$

$$RRMSE = \sqrt{\frac{1}{n} \sum\_{i=1}^{n} \left(O\_i - P\_i\right)^2} \times \frac{1}{\overline{O\_i}} \times 100\% \tag{6}$$

where *n* is the number of samples; *Oi* is the observed LNC value; and *Pi* is the estimated value. The saturation sensitivity of the VI versus LNC was evaluated using the index of noise equivalent (NE LNC) [46,47].

$$\text{NEALNC} = \frac{\text{RMSE} \| VI \text{vs.LNC}}{d(VI) / d(L\text{NC})} \tag{7}$$

where *RMSE*{*VIvs.LNC*} is the root mean square error (RMSE) of the best fit function and the actual LNC value, and *d(VI)*/*d(LNC)* is the first derivative of VI with respect to LNC. A higher NE LNC indicates a lower sensitivity to LNC. The accuracies of the optimal estimation models from RGB and CIR images were compared under di fferent treatments (i.e., di fferent varieties, N application rates, and planting densities) using RRMSE.

## **3. Results**

#### *3.1. Changes of Digital Number (DN) Values in Di*ff*erent Channels*

Figure 4 shows the changes of DN values in di fferent channels with the variation in LNC for images from the RGB and CIR cameras. The DN values of all channels from the RGB camera decreased when the LNC increased to 3.0%, and then became flatter as the LNC continually increased (Figure 4a). Conversely, the DN values of the near infrared channel from the CIR camera increased as the LNC increased (Figure 4b). The variations of the DN values in the blue and green channels from both cameras were similar, but the values were di fferent. For the RGB camera, the values of the green channel were higher than those of the blue channel. For the CIR camera, the DN values of the green and blue channels were similar.

**Figure 4.** Changes of digital number values in di fferent channels of (**a**) normal color (RGB) images and (**b**) color near-infrared (CIR) images with leaf N concentration (LNC) in winter wheat.

#### *3.2. Leaf N Concentration (LNC) Estimation Model Constraction and Validation*

3.2.1. Quantitative Relationships between Leaf N Concentration (LNC) and Vegetation Indices (VIs) at Di fferent Growth Stages

Figure 5 shows the relationship between LNC and the VIs determined from both RGB and CIR images at di fferent growth stages of winter wheat. For the VIs derived from RGB images, the results were quite di fferent (Figure 5a–f). When the LNC increased, the NGRDI, VARI, and ExG increased, while the IKaw, RGRI, and TCVI decreased. The exponential relationship between the ExG and LNC was weak at all four stages (Figure 5e). The relationship between the NGRDI and LNC (Figure 5a) was similar to that between the VARI and LNC, but the fitting curves of the VARI and LNC at the booting and heading stages were stronger than those of the NGRDI and LNC (Figure 5d). Unlike the other VIs, the relationship between the TCVI and LNC was almost identical at all four stages (Figure 5f). For the VIs derived from CIR images, the GNDVI, ENDVI, and FCVI all increased as the LNC increased (Figure 5g–i). The results for the GNDVI and FCVI at each growth stage were similar (Figure 5g,i). The sample points for the relationship between the ENDVI and LNC were distributed along the fitting curves at the booting, anthesis, and filling stages, but at the heading stage the results were scattered (Figure 5h). At intermediate to high LNC levels, the NGRDI, RGRI, VARI, and GNDVI were not sensitive to changes in LNC > 3.0%, especially at the heading stage.

**Figure 5.** Leaf N concentration (LNC) (%) plotted against different vegetation indices (VIs) at different growth stages of winter wheat. (**a**) normalized green-red difference index (NGRDI), (**b**) Kawashima index (IKaw), (**c**) red green ratio index (RGRI), (**d**) visible atmospherically resistance index (VARI), (**e**) excess green vegetation index (ExG), (**f**) true color vegetation index (TCVI), (**g**) green normalized difference vegetation index (GNDVI), (**h**) enhanced normalized difference vegetation index (ENDVI), (**i**) false color vegetation index (FCVI). Statistics are given in Table 6.

**Table 6.** Relationship between leaf N concentration (LNC) and vegetation indices (VIs) at different growth stages of winter wheat. *x* and *y* are VI and LNC, respectively. R2 is the determination coefficient. Relative root mean square error (RRMSE) (%) is the relative root mean square error. The bold values mean the most accurate result of normal color (RGB) and color near-infrared (CIR) cameras for each term.


To quantitatively analyze the ability of the VIs to estimate LNC at di fferent growth stages, the R<sup>2</sup> and RRMSE values at each growth stage and over all four stages are listed in Table 6. Generally, the results at the filling stage were much worse than at the other three stages, with an RRMSE higher than 15%. In the first three stages, the RRMSE values of most VIs were around 10%. However, the performances of the ExG and ENDVI were much poorer than those of the other VIs, especially at the heading and filling stages. The results for the TCVI and FCVI were not remarkably better than the other VIs at each growth stage, but they were obviously better for all four stages combined. For the RGB camera, the IKaw performed best at the booting and heading stages, and the VARI had the highest R<sup>2</sup> and lowest RRMSE values at the anthesis stage. Although the TCVI results were not the most accurate in the first three stages, they were superior to those of the other VIs derived from RGB images at the filling stage. For the CIR camera, the performance of the ENDVI was much poorer than that of the GNDVI and FCVI. The FCVI performed best at each stage, although the RRMSE of 9.5% at the booting stage was slightly higher than the value of 9.1% for the GNDVI. Compared to the VIs from RGB images, the FCVI results were less accurate than those of the IKaw at the first two stages, but were the most accurate at the anthesis and filling stages. For all four stages combined, the performance of the TCVI was remarkable, with an R<sup>2</sup> value of 0.852 and RRMSE of 12.1%, followed by the FCVI with an R<sup>2</sup> value of 0.792 and RRMSE of 14.0%.

#### 3.2.2. Validation of the Leaf N Concentration (LNC) Models for Wheat

The validation statistics for the LNC models constructed from the di fferent VIs are presented in Table 7. According to the results of the 10-fold cross-validation, the estimation models constructed from the TCVI had the highest accuracy. Compared to the other VIs from the RGB camera, the RRMSEs were improved by 6.32% to 15.76% for LNC based on the TCVI. For CIR images, the statistical values of the FCVI were the best for LNC, followed by the GNDVI. The independent validation indicated that the FCVI was capable of producing accurate LNC estimations at all stages.


**Table 7.** Validation statistics for the leaf N concentration (LNC) estimation models from di fferent vegetation indices (VIs). The bold values mean the most accurate result of normal color (RGB) and color near-infrared (CIR) cameras for each term.

#### *3.3. Saturation Sensitivity of Vegetation Indices (VIs) at Di*ff*erent Leaf N Concentrations (LNCs)*

As shown in Figure 6, the saturation of the selected VIs was tested under the LNC range of 0 to 5%. The NE LNC of ExG increased sharply as the LNC increased. The ExG was the VI with the fastest and largest increase, followed by the ENDVI. When LNC < 3%, the NE LNC of the IKAW, RGRI, and VARI were similar. However, when LNC > 3%, the NE LNC of the IKAW increased more slowly than that of the RGRI and VARI. Compared to the other VIs, the NE LNC of the TCVI and FCVI were

much lower, with that of the TCVI being the smallest. The NE LNC of the TCVI was still the lowest of all the VIs at high LNCs.

**Figure 6.** Sensitivity of the selected vegetation indices (VIs) to leaf N concentration (LNC).

#### *3.4. Applicability of the Leaf N Concentration (LNC) Models under Di*ff*erent Treatments*

According to the results in Table 6, the optimal LNC estimation models for the whole season from RGB and CIR images were constructed by the TCVI and FCVI, respectively. Table 8 shows a comparison of the optimal estimation models from RGB and CIR images under the different treatments. Generally, the RRMSE values of LNC estimations under the various treatments were lower than 15%.



For the different wheat varieties, the RRMSE values of LNC estimations from RGB images were lower than those from CIR images, with the lowest RRMSE of 9.88% for Shengxuan 6. Under the different N application rates, the LNC models from RGB images performed better than those from CIR images. For both RGB and CIR images, the RRMSE values for N application rates of 150 kg/ha were lower than for the other N application rates. As planting density decreased, the accuracy of LNC estimation from RGB images decreased, while for CIR images it increased. The best results for the lowest planting density were obtained from CIR images, and the lowest RRMSE values for the highest planting density were obtained from RGB images.

## **4. Discussion**

#### *4.1. Performance of the Vegetation Indices (VIs) Derived from Digital Imagery in Estimating Wheat Leaf N Concentration (LNC)*

The estimation of VIs is the most common and simplest way to extract information on crop growth status from digital images [27]. Di fferent VIs calculated from di fferent wavelengths highlight various vegetation properties [48–50]. Due to the di fferent band combinations and di fferent formulas used, the accuracy of N status estimation varies between the di fferent VIs.

In this study, seven commonly used VIs (i.e., NGRDI, IKAW, RGRI, VARI, ExG, GNDVI, and ENDVI) and two new VIs (i.e., TCVI and FCVI) derived from both RGB and CIR images for estimating LNC were analyzed in wheat. As shown in Table 9, the numerator of the VIs from RGB images consisted of two channels, except ExG. Among them, the numerator of IKaw and TCVI contained the red and blue bands, while the red-green VIs (i.e., NGRDI, RGRI, and VARI) used the red and green bands as the numerator. The results in Figures 5 and 6 indicated that the ability to mitigate the saturation of the red-green VIs under high LNC levels was weakest for IKaw and TCVI. This might be because the red and blue bands are the chlorophyll and carotenoid absorption bands [51]. Given that crops respond to N status mainly by a change in chlorophyll concentration in the leaves [52], VIs using ratios or normalized di fferences of values acquired in the red and blue bands were significantly related to the N status of crops [53–55]. The GNDVI, which is calculated from the normalized di fference between values of the green and NIR bands, was also limited by the saturation under intermediate to high LNC levels (Figures 5g and 6). Unlike the GNDVI, the proposed FCVI acquired information in the blue band, which reduced the saturation. Although the ENDVI also used all three bands, its performance in estimating LNC was relatively poor. Therefore, we cannot conclude that a VI incorporating more bands is able to produce more reliable LNC results. The accuracy is dependent on both the band configuration and VI formulation.


**Table 9.** Coe fficients for Di fferent Vegetation Indices (VIs).

#### *4.2. Accuracy and Universality of Leaf N Concentration (LNC) Estimation Models in Wheat*

In this study, the LNC estimation models constructed with the TCVI and FCVI were sensitive under varying LNCs (from 0 to 5%) and had a better accuracy under the di fferent treatments. Moreover, they were generalizable from the booting to filling stages of wheat. As the growth stage progressed, dramatic changes in the composition of canopy components and background materials can occur. These changes pose a critical challenge for the timely monitoring of crop N status. For the late stages (booting to filling) in the reproductive period, a single LNC–TCVI or LNC–FCVI model could be fitted, with a high e fficiency and low error (Figure 5 and Table 6). Although ExG is mainly used to extract vegetation from di fferent backgrounds and has been widely cited [12,42,56], its performance for estimating N status was very poor, which was consistent with previous results [12]. Previous studies have suggested that adjusting VIs with a term representing soil brightness could reduce the e ffect of the soil background [22,28]. Given that a soil background adjustment parameter (*L*) was added to the TCVI and FCVI, they had the potential to reduce the significant e ffect of soil background during the early stages of the vegetative period.

Due to the limited experiment, we only made an independent validation for the LNC–VI models from CIR images based on Exp. 1, and the LNC–VI models from both RGB images were calibrated and validated with a 10-fold cross-validation procedure using the data from Exp. 2. Since the FCVI and TCVI were developed based on the relationships between CVI and LNC from Exp.2, there is a need to validate their performance with independent measurements under di fferent conditions (i.e., varied crop types, di fferent site conditions and other cameras). Although the coe fficients of FCVI and TCVI had been modified, the accuracy of LNC–TCVI or LNC–FCVI models was not always the highest (see Table 6). For example, the IKaw performed best at the booting and heading stages, and the VARI had the highest R<sup>2</sup> and lowest RRMSE values at the anthesis stage. However, these relationships may vary with crop type, site condition, and growth stage as influenced by the variation in physiological processes. Therefore, the universality of the LNC estimation models needs to be further verified during the early growth stages and under various conditions.

#### *4.3. Capability of Commercial Digital Cameras for Wheat Leaf N Concentration (LNC) Estimation*

The development of simple but e fficient methods to monitor crop growth across a wide range of LNCs is urgently needed for precision agriculture in China. This is because the application of N fertilizer has increased in recent years in an attempt to boost production. Therefore, the timely monitoring of crop N status under intermediate to high N application levels is essential to maximize yield. In addition, easy-to-operate and low-cost equipment is required to help the owners of large farms or smallholders with fertilization management. Due to the low price and convenient operation of digital cameras, they have potential application prospects. In this study, UAV campaigns were conducted at noon in clear weather under stable lighting conditions. However, radiation correction should be considered in the future work, especially when experiencing changeable light intensity (e.g., for rice monitoring in summer).

To explore the capability of commercial digital cameras to estimate LNC in wheat, we constructed new VIs (i.e., TCVI and FCVI) from both RGB and CIR images. Because the TCVI and FCVI were based on large amounts of field experimental data under di fferent conditions, they performed reliably in indicating the LNC in wheat. It is suggested that commercial digital cameras have the capability to derive optimum VIs for LNC monitoring in winter wheat. In addition, the performance of RGB images was generally better than that of CIR images (see Table 8). This study supported the widespread agreemen<sup>t</sup> that digital cameras are powerful tools for assessing crop growth status and further proved the applicability of UAV-based RGB and CIR cameras for the monitoring of crop N status.

## **5. Conclusions**

The applicability of digital cameras mounted on UAVs for monitoring the LNC of winter wheat was evaluated in this study. Seven commonly used VIs (i.e., NGRDI, IKaw, RGRI, VARI, ExG, GNDVI, and ENDVI) and two proposed VIs (i.e., TCVI and FCVI) were used to estimate LNC at di fferent growth stages of winter wheat. The performances of NGRDI, RGRI, VARI, and GNDVI were limited by saturation at intermediate to high LNCs (i.e., LNC > 3.0%). The accuracy of LNC estimation using ExG was the poorest among the VIs tested, while the optimal models were constructed by the TCVI and FCVI for all stages. The models were then cross-validated with datasets from di fferent cultivars, N application rates, and planting densities. Compared to the other VIs derived from the RGB camera, the RRMSE values of the TCVI were improved by 6.32% to 15.76% for LNC. For the CIR camera, the statistical values of the FCVI were the best for determining LNC ( R<sup>2</sup> = 0.756, RRMSE = 14.18%). The independent validation also indicated that the FCVI was capable of accurately estimating LNC at all growth stages. In summary, commercial digital cameras mounted on an UAV are feasible for monitoring wheat LNC at the farm-scale, especially under the high N fertilizer applications and different treatments typically used in fields in modern China.

In practical terms, commercial digital cameras are low-cost and easy to operate for researchers and farmers. Although the equipment is applicable, the LNC–VI relationship may vary with crop type, site condition, and growth stage, and may be influenced by variations in physiological processes. Therefore, additional calibrations are needed for different conditions before extending this method to other crops. For example, the use of UAV-based digital cameras for crop N status monitoring should be further investigated during the early stages of the vegetative period. Moreover, fluctuating ambient lighting conditions are an issue that should be addressed in future studies.

**Author Contributions:** X.Y. and T.C. conceived and designed the experiments. J.J. and H.Z. performed the experiments. J.J. analyzed the data. J.J. and X.Y. wrote the paper. All authors contributed to the interpretation of results and editing of the manuscript.

**Funding:** This work was supported by grants from the Key Projects (Advanced Technology) of Jiangsu Province (BE2019383), National Natural Science Foundation of China (31671582, 31971780), Jiangsu Qinglan Project, Jiangsu Collaborative Innovation Center for Modern Crop Production (JCICMCP), the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD), Qinghai Project of Transformation of Scientific and Technological Achievements (2018-NK-126); Xinjiang Corps Great Science and Technology Projects (2018AA00403), and the 111 project (B16026).

**Acknowledgments:** We would like to thank former graduate students Yong Liu and Ni Wang for their help with field data collection.

**Conflicts of Interest:** The authors declare no conflict of interest.
