Monitoring the nitrogen nutrition status of rice plants using spectral 1 and image technologies

17 Background: We aimed to investigate methods to estimate the nitrogen (N) nutrition status of rice 18 plants using data obtained using a digital camera and a spectroradiometer. The overall aim was to 19 compare the advantages and potential of image technology and spectral technology to monitor rice 20 N indexes accurately, inexpensively, and in real time to optimize fertilization strategies. Realizing 21 the technical selection of definite spectrum or image diagnosis aiming at different rice nitrogen 22 nutrition indexes. We conducted field trials of rice plants grown with different levels of N fertilizer 23 in 2018 to 2019. Spectral information and images of the rice canopy were obtained, various image 24 and spectral characteristic parameters were selected to construct models to estimate rice N status. 25 Results: The determination coefficients of the models constructed using the ratio vegetation index 26 (RVI [800,550] ) and cover canopy (CC) as dependent variables were most significant. Among the 27 models using spectral parameters, those constructed using RVI [800,550] to estimate rice N indexes had 28 the obviously coefficient of determination ( R 2 ) values, which were 0.69, 0.58, and 0.65 for the 29 models to estimate leaf area index(LAI), aboveground biomass(AGB), and plant N 30 accumulation(PNA). As for image parameter, those using CC to predict rice N indexes showed the 31 highest R 2 values (0.76, 0.65, and 0.71 for the models to estimate LAI, AGB, and PNA, respectively) 32 ( P < 0.01). The model using the spectral parameter RVI [800,550] had a good fit and stability in 33 estimating plant nitrogen accumulation ( R 2 = 0.65, root mean square error (RMSE) = 1.35 g · m -2 , 34 relative RMSE (RRMSE) = 14.05%), and the model using the image parameter CC had a good fit 35 in predicting leaf area index ( R 2 = 0.76, RMSE = 0.28, RRMSE = 7.26%) and aboveground biomass 36 ( R 2 = 0.65, RMSE = 22.03 g · m -2 , RRMSE = 7.52%). Different detection technology should be 37 adopted for different rice varieties and Conclusions: Spectral and image parameters can be used as technical parameters to estimate rice N 39 status. The spectral parameter RVI [800,550] can be used to accurately estimate plant nitrogen 40 accumulation, and the image parameter CC can be used to accurately estimate leaf area index and 41 aboveground biomass.

the amount of N fertilizer applied to rice crops accounts for 37% of the N fertilizer used globally.  were 0.54, 0.55, and 0.55, respectively (P < 0.01) (Fig.2). 128 Take DVI[800,720] as an example, there were significant differences between conventional rice and 129 hybrid rice in the application of spectral parameters to predict rice nitrogen status. The relationships 130 between DVI[800,720] and the N indexes of rice were all polynomial functions, the accuracy of 131 monitoring rice N indexes by DVI [800,720] in Zhongjiazao 17 was higher than that in hybrid rice 132 Changliangyou173 (Fig.3). The relationships between CC and N nutrition indexes of rice were all polynomial functions. The 139 R 2 values for models using CC to predict LAI, AGB, and PNA were 0.76, 0.65, and 0.71, 140 respectively (P < 0.01). LAI, aboveground biomass and plant nitrogen accumulation of rice 141 increased with the increase of CC, while the correlation coefficients between NRI, Hue and rice N 142 indexes were not significant(R 2 <0.5), the average correlation coefficient of NRI model was 0. 16,143 and that of hue was 0.10. As for conventional rice and hybrid rice in the application of CC to predict 144 rice nitrogen status, the average coefficient of models based on CC in hybrid rice Changliangyou173 145 was 0.74, which was higher than that in conventional rice Zhongjiazao17. 146 It can be seen from the above that the models based on RVI [800,550] and CC had good prediction 147 effect for rice N indexes, and there were significant differences among different gene varieties.

3) Regression validation 155
To test the accuracy of the models, those based on the spectral parameter RVI[800,550] and the image 156 14 parameter CC were tested and evaluated using data from experiment 3 obtained at the jointing stage 157 (Fig. 7, Fig.8). The RMSE, RRMSE, and r 2 values were calculated to evaluate the accuracy and 158 stability of the models. The result showed that the r 2 from RVI[800,550] regression equations were 159 0.51,0.47 and 0.86 respectively, and that the RMSE values were 0.77, 42.18 and 1.35, respectively 160 ( Fig. 7a, 7b, 7c). As shown in Fig. 8

Comparison of Methods to Estimate Rice N Status 176
In recent years, accurate and non-destructive spectral and image techniques have been developed 177 for the real-time monitoring of crop growth and N nutrition [21,22]. However, few studies have 178 compared and contrasted models constructed using data obtained using these two techniques. 179 Hyperspectrometry has many advantages, including precise measurements and abundant spectral 180 data [23,24]. Single bands readily become saturated, it is better to use data from two or more bands 181 as spectral parameters to create models to estimate the biochemical parameters of vegetation [25]. can describe plant color [28], which can reflect its nutrient status, especially N content and 198 absorption. Several studies have shown that RGB color space parameters extracted from vegetation 199 canopy images can be used to predict vegetation yield and nutrient status [15,19,20]. Among the 200 models constructed with image parameters in this study, those constructed using NRI were unstable, 201 possibly because the parameters of NRI were obtained by extracting RGB values from images. 202 These values can be affected by the time and the weather when the image was acquired. The model 203 constructed using CC had a good fitting effect. The R 2 values of the models using CC to estimate 204 LAI, AGB, and PNA were 0.76, 0.66, and 0.71, respectively, consistent with the conclusion that CC 205 is a reliable parameter to estimate vegetation N content [29]. The CC value is obtained by removing 206 the influence of soil and water in the image. Compared with other image parameters, CC is obtained 207 more easily and is not affected by weather or light intensity. 208

Advantages and Disadvantages of Models using Spectral and Image Parameters 209
The results of previous studies indicated that the booting stage is the peak period of rice plant growth, 210 when the LAI is the highest. The booting stage is considered as the best time and cut-off point for 211 estimating rice yield using remote sensing. However, some other studies have found that the early 212 heading stage is the best time to use the spectral index RVI and color indexes to estimate rice LAI 213 [30,31]. In our study, through the correlation analysis of spectral parameters and image parameters 214 with the nitrogen nutrition index of the whole growth period of rice, the parameters with larger 215 correlation value were selected for modeling. According to the practice of fertilization in the double 216 cropping rice region of southern China, the last fertilizer, panicle fertilizer, must be applied before 217 booting stage to supply the nutrition needed after booting. Therefore, in order to achieve accurate 218 fertilization before booting, the data of jointing stage were used for modeling. The results of the 219 comparative analysis of the constructed models ( In addition, different rice varieties also had influence on model construction. The difference in 231 18 nitrogen nutrition diagnosis between hybrid rice and conventional rice may also be related to 232 nitrogen use efficiency. Previous studies have shown that the nitrogen accumulation in hybrid rice 233 is significantly higher than that in conventional rice as the nitrogen supply level increases [4]. Peng 234 et al. [32]indicated that the application ratio of panicle fertilizer should be increased to promote 235 nutrient absorption and accumulation in the middle and late growth stage of hybrid rice. There was 236 a significant correlation between vegetation reflection and nitrogen accumulation, which could be 237 analyzed using multi-term linear regression method [33], consistent with this study. Moreover, the 238 correlation between crop population reflection spectrum and nitrogen accumulation was better than 239 that between digital image and nitrogen accumulation. The two-band combination has advantages 240 in the inversion of nitrogen accumulation. 241 An effective strategy to optimize N use for rice should be suitable for the methods used by farmers, 242 while taking account of factors such as cultivars that affect the N requirements of rice and the 243 efficiency of its use. There are still many uncertain factors in remote sensing of crop N status. In 244 this study, we did not consider the effects of several imaging factors (shooting angle, storage format, 245 shooting time, and camera resolution). To obtain a reliable and universal model, it is necessary to 246 further standardize imaging factors, test varieties, growth period, and test points, and to integrate 247 soil and climate data. This will improve the accuracy of models so that they can be used to quickly 248 diagnose the nutrient status of field crops and establish a tailored fertilization system. 249

Conclusion 250
In this study, we constructed models to estimate rice N indexes with the image parameter and 251 the spectral parameter. We analyzed the accuracy and stability of the models to predict LAI, AGB, 252 and PNA. The results showed that the R 2 values of the models constructed with the image parameter 253 19 CC and the spectral parameter RVI[800,720] were very significant. Compared with other models, the 254 polynomial model constructed using CC to predict LAI ,AGB and the model constructed using 255 RVI[800,550] to predict PNA during the jointing stage had better prediction and test results. Our results 256 showed that image parameters can be used to estimate rice N status (especially LAI and AGB). We 257 conclude that image technology can be used as a low-cost, non-destructive, and rapid method to 258 monitor rice N status instead of spectral technology, which could be suitable for the methods used 259 by farmers. 260

Study Area and Experimental Details 262
Three independent experiments were performed in this study. 263

Field Data Collection 297
Repeated destructive sampling was carried out in each plot for Exp. 1 and Exp. 2 Three rice plants 298 from each experimental plot were randomly selected to determine LAI. For each sample, the green 299 leaves were separated from the stems, and the leaf area (LA) was immediately determined by 300 multiplying length by width. The LAI for each plot was calculated based on the planting densities. 301 After bagging, the plant samples were heated in an oven at 105 °C for 30 min, dried to constant 302 weight at 80 °C, and then weighed to determine the dry weight per unit area. Samples were crushed 303 before determining N content using the Kjeldahl method. The PNA value was calculated as follows: 304 Where LNC is leaf N content, LDW is leaf dry weight, SNC is stem N content, SDW is stem 307 dry weight, PNC is plant N content, and PDW is plant dry weight. Before sampling, images of the 308 rice canopy were obtained using a Canon EOS 100D digital camera (resolution, 72 DPI) (Canon, 309 Tokyo, Japan) . The camera lens was about 1.0 m away from the rice canopy at an angle of 60° 310 relative to the ground. The camera was set to auto mode to control the color balance automatically. 311 The images were stored in JPEG format with a resolution of 5184 × 3456 pixels. In the periods of rice growth, the canopy does not completely obscure the ground, so images contain 322 soil, water, and other non-canopy items. Consequently, it is necessary to segment and extract the 323 canopy part from the image. Image segmentation eliminates interference from non-canopy items so 324 that data for the crop canopy can be extracted and analyzed. We used the Otsu threshold 325 segmentation algorithm to segment images. This image segmentation method is based on the 326 difference of reflectance spectra between green vegetation and soil in the visible light region. 327 parameters, a variety of color parameters can be obtained. Table 3 showed that the references of 340 23 image parameters previous researchers used to indirectly characterize crop nitrogen nutrition. In this 341 study, eight color parameters including image R-G-B were selected. 342 2) Spectral data processing 344 Table 4 showed that the references of spectral reflectance parameters previous researchers used to 345 indirectly characterize crop nitrogen nutrition. 346 Normalized difference vegetation index NDVI (λ1, λ2) [35] Red edge position wavelength λrep [36] 3) Data analysis 348 In the models, the rice N nutrition index was set as the dependent variable, and image 349 parameters and spectral parameters were set as independent variables. The quantitative relationships (2) 359 .
(3) 360 In the above formulae, n is the number of samples tested for model test; Pi is the predicted value 361 of the model, is the average value of the predicted value; Oi is the measured value; and is 362 the average value of the measured values.