1. Introduction

The photosynthetic status relies highly on the contents of biochemical parameters, especially the leaf nitrogen content (LNC), which is an important indicator for the photosynthetic status of crops [1]. Thus, accurately estimating the LNC is important for the precise characterization of the photosynthetic status. Through indirectly analyzing the relationship between LNC and the specific meteorological variables, such as CO₂ flux [2] and climatic warming [3], researcher tracked the photosynthetic ability of crops. The analysis method was passive and active remote sensing technologies over a wide scale. In this article, we utilized a simple ratio index technology derived from active remote sensing methods, hyperspectral LiDAR (HSL) and laser-induced chlorophyll fluorescence (LIF), to estimate rice LNC.

Some estimations on the relationship between LNC and CO₂ flux were made by Kergoat et al. [2]. They found that the LNC significantly influenced the canopy light use efficiency (LUE,$\epsilon$) and canopy photosynthesis rate of vegetation. These studies were usually based on crop reflectance spectra which were used to derive some metrics such as vegetation indices (VIs or spectral indices). The relationship between VIs and LNC has been widely applied on remotely sensing vegetation, particularly the simple ratio indices in the red and near-infrared region of the reflectance spectrum, RI = R_red /R_nir [4, 5]. Rouse et al. [6] proposed another simple index called the normalized difference vegetation index (NDVI). However, RI or NDVI measures the optical properties as the “greenness” of land cover, and it does not consider the inherent photosynthetic process of vegetation. By contrast, the LUE is determined by the photosynthetic efficiency. Cannell et al. [7] reviewed several cases and found that $\epsilon$ was not a constant and it varied from 0.8 g of biomass per MJ solar radiation to 2.1 g biomass per MJ solar radiation. This variation is not observable in short hours through the reflectance spectra as the leaves remain green, but photosynthesis is reduced because of stress factors. Over a long time scale, a change in NDVI is not evident unless these stress factors cause premature senescence. A saturation problem exists in these VIs: the variation in greenness of vegetation leaf becomes constant when the leaf is green enough [8]. Thus, reflectance-based methods are not sensitive for the accurate estimation of LNC in early growing stages [9] and photosynthesis.

The LIF spectrum is an attractive and early indicator for detecting the nutrient stress of crops. It is highly related to the biochemical contents of plants during photosynthesis. In general, light illuminated onto the leaf surface is scattered and absorbed by chlorophyll in chloroplasts via the photosynthetic system II (PS-II). Subsequently, residual light energy is dissipated in the form of heat, one of which is Chlorophyll fluorescence (CF) [10], to prevent the oxidative damage of chloroplasts. Maxwell et al. [11] found that CF was highly sensitive to stress from the surrounding environment, such as extreme temperature, direct sunshine, and water or nutrient deficiency. Therefore, CF has been widely utilized to indicate a plant’s photosynthetic ability and evaluate various stress effects on photosynthesis in plants [12]. Yang et al. [13, 14] used the CF spectra to estimate rice LNC, and got some useful results based on principal component analysis. Wu et al. [15] and Patterson et al. [16] developed mathematical models of fluorescence in turbid media, and they reported that fluorescence was simply related to reflectance. Thus, the combination of reflectance and LIF spectra has great potential for the estimation of crops LNC. Du et al. [17] accurately estimated rice LNC with several regression methods by selecting more than eight wavelengths from HSL data and four peaks from the LIF spectrum. Yang et al. [18] also performed similar work using reflectance collected by ASD FieldSpec Pro FR (Analytical Spectral Devices, Inc., Boulder, USA) and LIF data. However, both of them combined reflectance and LIF spectra disconnectedly, and there were numerous variables for regression model in the study of Du et al.. Profio et al. [19] used a ratio of fluorescence to reflectance at the same wavelength, and they concluded that this ratio was proportional to the fluorophore concentration and the quantity $(1-\gamma )/(1+\gamma )$, where $\gamma$was the diffuse reflectance of tissues. Lichtenthaler et al. [20] pointed that CF indices at 440, 520, 690, and 740 nm were closely related to stress conditions. For high light exposure, drought, and extreme temperature, the fluorescence ratios at red (690 nm) and far-red (740 nm) increased significantly [21], which could be used as a field- or laboratory-based diagnostic tool for stress detection. In the present study, a combined ratio index of reflectance to fluorescence at red (690 nm) and far-red (740 nm) which was derived from HSL and LIF data was used to estimate rice LNC, and the performance of HSL or LIF data were also compared separately.

Compared with the other multivariate statistical methods, researchers utilized artificial neural networks (ANNs) to predict vegetation biochemical parameters and crop yields [22, 23] neither considering the assumptions of statistical frequency distribution of data, nor caring about the measurement scales of the features used in analysis. In using ANNs, traditional empirical risk minimization (ERM) was employed to improve the generalization performance; this step was the key difference from other regression methods [24, 25]. One of the common ANN models in regression was radial basic function neural networks (rbf-NNs), which was implemented in this work. The aim of this study was to investigate whether the ratio that combined HSL and LIF data could improve the rice LNC, and then compare their performance in the LNC analysis based on the rbf-NN model.

2. Materials and methods

2.1 Sampling areas

The experimental samples were collected from a main rice-producing area located in the Jianghan Plain of China (29°58′ N to 31°22′ N, 113°41′ E to 115°05′ E) during the rice-growing seasons of 2014 and 2015. This sampling area belongs to the subtropical monsoon zone. Its climate is mild and has a mean precipitation of 856–1070 mm and a temperature of 15.5 °C [26].

In 2014, the variety of rice was Yongyou 4949 which was grew in Suizhou City, China. The cultivars were seeded on April 27 and then transplanted to the field on June 1. Six levels of urea fertilizer were used in the experimental fields (0, 189, 229.5, 270, 310.5, and 351 kg/ha), and three repetitions were performed for each cultivation condition. All of the cultivation subareas were fertilized as follows: 30% as basic fertilization, 20% in booting stage, 25% in tillering stage, and 25% in heading stage. The leaf samples were collected on July 15 2014 and August 1 2014, which corresponded to the booting and heading stages, respectively.

In 2015, the the variety of rice was Yangliangyou 6 which was grew in the experimental base of Huazhong Agricultural University in Wuhan City, China. The rice seedlings were seeded on April 30 and transplanted to the field on May 27. During the cultivation period, four urea levels were applied (0, 120, 180, and 240 kg/ha), and the seedlings were fertilized with three splits: 60% at seeding, 20% at tillering, and 20% at shooting. Three repetitions were also performed for each identical cultivation condition. Leaf samples were collected on July 21, 22, 24, and 26 of 2015, which corresponded to the tillering stage.

2.2 Spectra measurement by HSL and LIF LiDAR system

The spectra (both reflectance and fluorescence) for each sample were measured at three different rice leaf positions. At the same measurement position, five replications were performed, which were then averaged as the characteristic spectra for one leaf. During the experiment, the incident laser spots resided onto the leaf surface perpendicularly and completely.

The HSL system used in this work was detailed in the study of Du et al. [27]. The system emitted a wide-band laser with a frequency of 20–40 kHz and a pulse duration of 1–2 ns. The laser spot illuminated targets with a diameter of approximately 10 mm, and its divergence angle was less than 3 mrad. The wavelength of this HSL system ranged from 538 nm to 910 nm with a spectral resolution of 12 nm. Before collecting the spectra of samples, a reference spectrum with the white panel (10 cm × 10 cm, >99% reflectance, Spectralon, Labsphere, Inc., North Sutton, NH, USA) was initially captured and denoted by R_ref. Furthermore, the reflectance spectrum for each measurement, denoted by$R(\lambda )$, was calculated as

The LIF system used in this work was consisted of an ultraviolet excitation light source (355 nm, 20 Hz), Maksutov–Cassegrain telescope, spectrograph (Princeton Instrument SP2500i), and ICCD camera. The fluorescence spectrum range was 360–800 nm with a resolution of 0.5 nm. Further information can be found in the study of Yang et al. [28]. Both of these LiDAR systems have the same optics design. The emitted laser (including the continuum laser and 355 nm) were reflected by two parallel mirror (M1 and M2) which had a angle of 45° with the direction of the emitted laser, and then illuminated onto a rotator. The rotator was located on a 2-D turntable which could scan the targets easily. The returns of laser could be collected and transformed by the spectrograph and the corresponding sensors through an optical fiber. The simple diagram of the LiDAR system is showed in Fig. 1.

 

Fig. 1 Simple diagram of HSL and LID LiDAR system. M1 and M2 are plane mirrors which are to make the emitted laser be coaxial with the receiving module of LiDAR. The returned signals are collected by a grating spectrograph.

Download Full Size | PDF

2.3 LNC analysis with Kjeldahl methods

After the measurements of reflectance and fluorescence spectra, all samples were immediately weighed as fresh matter and then cut into pieces. Part of this fresh matter was dried at 105 °C for 0.5 h and then at 70 °C until a constant weight was obtained and denoted as dry matter. Part of the dry matter was used for LNC analysis with the Kjeldahl method [29]. LNC was expressed as milligrams of nitrogen per gram of leaf dry matter.

2.4 Data analysis methods

The reflectance and fluorescence spectra were measured with active remote sensing technologies. Several overlaps existed in the reflectance and fluorescence spectra, especially in the red and near-infrared bands. These bands were closely related to the LNC of vegetation. Thus, characteristics on the fluorescence spectrum at ~685 and ~740 nm and the reflectance spectrum at 694 and 742 nm were combined to analyze rice LNC. In this work, a simple reflectance/fluorescence ratio index (Eq. (2)) was applied for analyzing rice LNC. The results were then compared with the ratio index using only one type of data, such as reflectance or fluorescence spectrum. The same method was performed on another spectral index, the normalized difference spectral index (NDSI) as well.

In each growing season, 80% of the data sets was utilized to train the model of rbf-NNs, and the remaining 20% was utilized for testing. The rbf-NNs in this study was consisted of an input layer, nine hidden layers, and an output layer implementing default training and transfer function. The ANNs were invented based on the thinking process of the human brain. By learning from the input information x_i, the synaptic weights w_i can be freely changed and modified. All of the inputs u_i (Eq. (4)) can then be transformed as the output value y_i by a nonlinear activation function f (Eq. (5))

3. Results

3.1 Spectra measured by HSL and LIF LiDAR system

The characteristic spectrum of reflectance and fluorescence could be collected with the HSL and LIF LiDAR system. The reflectance spectra differed within 500–900 nm depending on the LNC levels because of the high absorption of various leaf pigments [30]. The fluorescence spectra exhibited two main peaks at ~685 and ~740 nm, which were attributed to chlorophyll a and b, respectively [31]. By observing the samples, results in Fig. 2 show that the reflectance intensity is lower at 500-670 nm than > 700 nm and the fluorescence intensity appears two obvious peaks, which is consistent with the optical properties in theory. Figure 2 shows the spectrum collected with HSL and LIF LiDAR system under different nitrogen status. The thin lines represent a nitrogen level of 3.4 mg/g and the thick lines represent the lower nitrogen level (3 mg/g). The reflectance spectrum covers a range of 538–802 nm (the blue line); the characteristic spectrum of fluorescence gained with LIF LiDAR system exhibits double peaks (the red line). Besides, the spectrum intensity are different under different nitrogen levels. Concretely, the reflectance intensity is positively related to nitrogen level at a range of < 680 nm (or < 670 nm, called “red-edge”), while this relationship becomes negative when > 700 nm. In contrast, the fluorescence intensity at the characteristic double-peaks increases along with the increase of nitrogen levels.

 

Fig. 2 Spectrum collected with HSL at a range of 538–802 nm (blue line). The superimposed (red line) is the characteristic fluorescence spectrum gained with the LIF system with double peaks at ~685 and ~740 nm. Thin lines represent a nitrogen level of 3.4 mg/g, and the thick lines represent the lower nitrogen level of 3 mg/g.

Download Full Size | PDF

3.2 Rice LNC analysis with spectral indices derived from HSL and LIF data

Two kinds of spectral indices were developed to estimate rice LNC: one was the ratio index calculated with Eq. (2) (defined as an operator R(,)), and the other was in the form of NDSI defined as an operator N(,) in Eq. (3). To examine the relationship between LNC and these two kinds of indices, rbf-NNs regression was carried out in this work. The value of R² and RMSE of the ANN models are shown in Fig. 2. In this section, ratio indices were developed using reflectance spectra at 694 (H685) and 742 nm (H740) and fluorescence spectra at ~685 (F685) and ~740 nm (F740).

As shown in Fig. 2, all of the values of R² are not high, whereas the RMSE values are sometimes higher than R², which means that these separate indices are slightly related to rice LNC. Besides, indices using the reflectance spectra from the HSL system are more closely related to rice LNC than that using fluorescence spectra. Moreover, the best prediction model using indices is based on R(H685,H740) in the same year, although the rice samples are collected in different growing stages shown in Figs. 3(a) and 3(b). In 2015, the variable of the best prediction model of rice LNC estimation is N(H685,H740), whose R² exceeds 0.5 but RMSE is approximately 0.25 shown in Fig. 3(c).

 

Fig. 3 Comparison of the determination coefficient (R²) and RMSE using separate ratio indices for rice LNC estimation based on the data collected on (a) July 15 2014, booting stage (b) August 1 2014, heading stage, and (c) 2015.

Download Full Size | PDF

3.3 Rice LNC analysis with combined ratio indices

In this section, two kinds of combined ratio indices were utilized to estimate rice LNC. Figures 4(a), 4(c), and 4(e) showed the results that concluded with the reflectance to fluorescence ratio. Figures 4(b), 4(d), and 4(f) indicated the results of using ratio indices developed like NDSI based on reflectance and fluorescence at ~685 and ~740 nm, respectively. In different growing years, the performance of these ratio indices in LNC estimation of rice varied. Comparatively speaking, the ratios of reflectance to fluorescence has a higher R² than ratio indices developed like NDSI, except for the samples in 2015, which has the best LNC prediction model in the order of R(H740, F740) > N(H685, F685) > R(H685, F685) > N(H740, F740). In addition, the ratios of reflectance to fluorescence at ~740 nm (R(H740, F740)) are highly related to rice LNC with a R² of > 0.7 in all growing years. The fusion of HSL and LIF data significantly improves the LNC estimation accuracy of rice.

 

Fig. 4 Correlation between the measured LNC and predicted LNC estimated with combined ratio indices and ANNs in different years. (a), (b), (d) and (e): 2014; (c), and (f): 2015; (a), (c) and (e): ratio of reflectance to fluorescence; (b), (d) and (f): ratio indices developed like NDVI using reflectance and fluorescence at ~685 and ~740 nm, respectively. The dotted red line represents the 1:1 line.

Download Full Size | PDF

When using ratio indices developed from HSL and LIF data separately to estimate rice LNC, the R² values of regression model were < 0.6, and the RMSE were even higher than R² shown in Fig. 3. While Fig. 4 showed that the ratio indices combining reflectance and fluorescence variables, such as R(H740, F740), could predict LNC of rice with a R² of > 0.7. It could be said that ratio indices of reflectance to fluorescence was a relatively effective indicator index for LNC estimation of rice.

4. Discussion

In the visible and near-infrared bands, reflectance and fluorescence spectrum of vegetation leaf have special characteristics because of the absorption effect of chlorophyll a and b. These characteristics have been detected and observed by the LiDAR systems (HSL and LIF) described in this paper accurately, thus, they are the effective and reliable tools of remotely sensing nitrogen status of rice leaf.

The reflectance is an important spectral property utilizing the ability of vegetation photosynthesis. Thus there are many spectral indices centered in the range of visible and near-infrared wavebands based on the reflectance spectrum. Another spectral property on vegetation photosynthesis is chlorophyll fluorescence, which expresses the balance between light absorption and light application in photosynthesis together with the non-photosynthesis quenching process [32]. Although the principle of reflectance and fluorescence differs, both of them are the indicators of the vegetation photosynthesis, and the complementarity of these two spectral properties makes it effective and robust in LNC estimation. In this paper, we try to combined these two spectral properties to estimate LNC of rice. By experiment and calculating, we find that rice LNC can be estimated by combined spectral indices with satisfactory accuracy, especially those calculated based on the ratio of reflectance to fluorescence. The ratio R(H740,F740) is thought to have a relatively higher sensitivity to rice LNC, which is validated by our study with relatively high R² and low RMSE than the other ratio indices shown in Figs. 4(a), 4(c) and 4(e). The ratio R(H685,F685) shows a lower R² and RMSE than the ratio R(H740,F740), however, it is higher than those that derived from reflectance or fluorescence separately (with a R² of < 0.6 in Fig. 3, while > 0.65 in Fig. 4), which indicates that the accuracy of LNC estimation can be improved by combining the reflectance and fluorescence variables in a form of ratio indices.

Most studies demonstrated that a reflectance index, called the photochemical reflectance index (PRI) calculated with Eq. (7) was closely related to LUE of vegetation in leaf and crop scales [2]. This index can be utilized to detect the energy changes in photosynthesis, which is based on a narrow waveband centered at 531 nm and a reference waveband at 570 nm. We also developed a similar index denoted by N(H685, F685) and N(H740, F740). However, it displays poor relationship with rice LNC, and even has weaker performance in LNC estimation than the ratio indices using the HSL or LIF data separately. This scenario may be due to the magnitude of reflectance and fluorescence differs considerably.

Previous studies indicated that the fusion of different data, such as reflectance, HSL, and LIF data, could improve the estimation accuracy of vegetation parameters [33–35]. However, no literature has evaluated the fusion of HSL and LIF data in a ratio form for predicting the LNC of rice. In the study of Du et al. [17], they used different regression methods, including rbf-NNs, to estimate the rice LNC by combining reflectance and fluorescence variables. Except for the different forms of data fusion, they found that the value of R² could be more than 0.9 using rbf-NNs, however, there were more inputs of regression models (more than 10). The number of inputs in this paper was two, maybe this was why the R² values were not particularly strong enough for rice LNC prediction. In this study, two kinds of ratio indices derived from HSL and LIF data were used as the inputs of the multivariate nonlinear regression models. Except for the mentioned ratio indices, we also calculated another complex index, denoted by R(H685, F685)*R(H740, F740), whose R² could be more than 0.85 (0.894 for 2014 first experiment, 0.895 for 2014 second experiment 2014 and 0.879 for 2015 experiment). This index had four input variables being same as R(H685, H740)*R(F685, F740) which was independent of system measurement errors and was useful to improve the prediction ability of LiDAR system in LNC estimation. Thus, our proposed method combining ratio of HSL-LIF-derived metrics based on different data sets was relatively effective for estimating rice LNC. In the future, further analysis may consider improving the data fusion approach to estimate vegetation parameters more effectively. The models aimed at expressing the light use process of vegetation may be helpful for this purpose, such as the PROSPECT model [36] in leaf scale and a leaf and canopy reflectance model, including the fluorescence model developed by Zarco-Tejada et al. [37].

This study estimated the LNC of rice based on the leaf samples collected from the fields. However, the HSL and LIF LiDAR system is an active remote sensing method with advantage of targeted detection or large-scale monitoring, such as in canopy scale. Future work may investigate the biochemical and biophysical compositions of vegetation canopy or the structural parameters of crop canopy in fields.

5. Conclusion

This study aimed to compare the performance of ratio indices derived from HSL and LIF spectra in LNC estimation of rice, as well as explore the potential of the combined ratio index combining reflectance and fluorescence based on the rbf-NN model. The variables at 694 and 742 nm from the reflectance spectra were more strongly related to LNC estimation of rice than the fluorescence variables. However, by combining the reflectance and fluorescence variables in the form of ratio indices, especially the ratio of reflectance to fluorescence, the accuracy of rice LNC estimation was significantly improved. This research revealed that implementing reflectance from the HSL system combined with the LIF LiDAR system could be an essential and valuable method for remotely sensing the LNC of rice accurately.