Open Access
Add to BibSonomyAbstract
Paddy rice is one of the most significant food sources and an important part of the ecosystem. Thus, accurate monitoring of paddy rice growth is highly necessary. Leaf nitrogen content (LNC) serves as a crucial indicator of growth status of paddy rice and determines the dose of nitrogen (N) fertilizer to be used. This study aims to compare the predictive ability of the fluorescence spectra excited by different excitation wavelengths (EWs) combined with traditional multivariate analysis algorithms, such as principal component analysis (PCA), back-propagation neural network (BPNN), and support vector machine (SVM), for estimating paddy rice LNC from the leaf level with three different fluorescence characteristics as input variables. Then, six estimation models were proposed. Compared with the five other models, PCA-BPNN was the most suitable model for the estimation of LNC by improving R2 and reducing RMSE and RE. For 355, 460 and 556 nm EWs, R2 was 0.89, 0.80 and 0.88, respectively. Experimental results demonstrated that the fluorescence spectra excited by 355 and 556 nm EWs were superior to those excited by 460 nm for the estimation of LNC with different models. BPNN algorithm combined with PCA may provide a helpful exploratory and predictive tool for fluorescence spectra excited by appropriate EW based on practical application requirements for monitoring the N status of crops.
© 2017 Optical Society of America
1. Introduction
Paddy rice is one of the most important food sources and is a significant part of the ecosystem. Approximately 30 million hectares were utilized to grow paddy rice in China per year, making it world’s leading producer of paddy rice [1]. At present, a large number of investigations have indicated that nitrogen (N) is one of the most important nutrient elements that improve cereal crop yield [2]. Although, China accounts for only 7% of the word’s cultivated land, it accounts for 35% of the world’s N fertilizer consumption. Thus, large amounts of N fertilizer are wasted, resulting in severe environmental pollution including nitrate leaching problems in soil and water eutrophication. In addition, recent studies demonstrated that flooded rice systems can emit both N2O and CH4 that is closely related to the dose of N fertilizer [3]. They account for 5%-19% of the annual global CH4 emission [1]. Therefore, precise estimations of leaf N content (LNC) is a brilliant strategy that matches the N fertilizer rate with paddy rice N demands in both spatial and temporal dimensions.
To accurately monitor the N status of vegetation, laser-induced fluorescence (LIF) technique was proposed [4], who utilized ultraviolet (UV) light as excitation light source to acquire chlorophyll fluorescent characteristics. In the plant, chlorophyll is closely related with photosynthesis of green plant, and then chlorophyll fluorescence is widely utilized to monitor the photosynthetic activity of plants and to evaluate the effect of various stress factors on it [5–7]. What’s more, relative investigations have been conducted on the effect of different EWs on the fluorescence spectra of vegetation [8–10], and different EWs were found to result in different fluorescence spectral shapes. The UV excitation light was found to be suitable for exciting blue-green fluorescence, whereas blue and green wavelengths were suitable for exciting the red and far-red fluorescence [11, 12]. Furthermore, a laser induced fluorescence transient (LIFT) technique to remotely measure photosynthetic properties in terrestrial vegetation was also proposed [13]. The fluorescence signals at certain wavelength varying with time were detected and were utilized to calculate the biochemical parameters of plant. It also can be utilized to monitor the other stress of plant. Compared with LIFT, LIF spectra included more spectral information and the different nutrient stress will result in different fluorescence intensity [14]. Because of its rapid, non-destructive, and highly sensitivity properties, LIF technology was extensively employed to monitor the N fertilizer level of crops [8,15–17].
In addition, multivariate statistical methods such as, back-propagation neural networks (BPNN), support vector machine (SVM), and principal component analysis (PCA) have also been employed for the quantitative remote sensing of the biochemical content of vegetation, which can be greatly improved in recent years [18]. BPNN can describe complex and intricate relationships between spectral information and various crop conditions [19]. SVM is a classical supervised learning algorithm with a strong theoretical foundation in statistical theory and can convert low-dimensional characteristics to high-dimensional characteristics to recognize complex targets [20]. By analyzing major attributes, PCA has become a useful tool for the induction of data dimensionality. It uses few variables obtained by a linear combination of the original data to explain the most important variable information. Thus, BPNN, SVM, and PCA have been frequently used in the statistical analyses of many research fields [21, 22].
However, comparative studies on the combination of these algorithms with PCA for estimating paddy rice LNC from leaf level based on LIF technology are still lacking. In addition, few investigations have been conducted on the estimation of LNC based on fluorescence spectra excited by different EWs, and the comparisons among the predictive capacity of different fluorescence parameters for the estimation of LNC from leaf scale in paddy rice are rare. Therefore, this study aims to: (1) compare the predictive ability of BPNN and SVM algorithms based on different models, such as FP-BPNN and FP-SVM models using fluorescence characteristic peaks (FP), FS-BPNN and FS-SVM models utilizing fluorescence spectra (FS), PCA-BPNN and PCA-SVM models based on principal component (PC) scores; (2) discuss the performance of different fluorescence parameters (FP, FS, and PCA) for estimating LNC in paddy rice; and (3) analyze the capacity of fluorescence spectra excited by different EWs for monitoring paddy rice LNC.
2. Materials and experiment
2.1 Study areas and site description
The experiment was conducted at the experimental station of the Huazhong Agricultural University in Wuhan, Hubei Province, China. The latitude of this area ranges from 29°58′ N to 31°22′ N, and the longitude ranges from 113°41′ E to 115°05′ E. The experimental area is a typical subtropical monsoon climate with abundant rainfall; the area is sunny and hot during summer and cold during winter. Thus, the area is very suitable for paddy rice cultivation and is well known as one of the most important grain production bases in China.
The paddy rice cultivar used was Yangliangyou 6. The grain were sown on 30 April 2015 and were transplanted to the field on 27 May. During the entire growth period, four N fertilization levels of urea (0 kg/ha, 120 kg/ha, 180 kg/ha, and 240 kg/ha) were utilized in these experimental fields. The N fertilization was divided into three splits, namely, 60% at seeding, 20% at tillering, and 20% at shooting. The experimental field is a complete block design with six replications for each treatment. Paddy rice samples were collected on four dates (20, 22, 24 and 26 July 2015), which corresponded to the tillering stage.
2.2 Laser-induced fluorescence (LIF) system
The LIF system was built in our laboratory and consisted of three parts, namely, the excitation light source, optical receiver assembly and the data collection, as well as the treatment part. The UV excitation light source is a 355 nm laser emitted by a neodymium-doped yttrium aluminum garnet laser. The 460 and 556 nm lasers were made by Spectra-Physics. The three EWs were relatively easy to obtain and represented the UV, blue, and green excitation light. In addition, the spectral range associated with the LNC of the plant was mainly between 650 nm to 800 nm. Thus, 355, 460, and 556 nm lasers were used to excite the leaf fluorescence of paddy rice. The LIF system works as follows. The laser light was transmitted from the beam expander and reflected by the L1 and L2, then the excitation light was transmitted perpendicularly to the targets. The fluorescence signal was collected by using Maksutov-Cassegrain telescope. Then, a single-mode optical fiber with a diameter of 200 µm was utilized to transmit the fluorescence collected between the telescope and spectrograph (Princeton Instrument SP2500i with the spectral resolution of 0.5 nm). The excited fluorescence then entered the spectrometer and was measured by utilizing the intensified charge coupled device (ICCD) camera. The data was stored in a personal computer for post-processing. In this system, an additional long-pass filter (Semrock BLP01-355R-25 with the edge of 361 nm and the 93% transmittance at 364.9-900 nm for 355 nm excitation light; and LP02-633RE-25 with the edge of 633 nm and the 93% transmittance at 636.9-1427.4 nm for 460 and 556 nm excitation light sources) was positioned behind the telescope and was used to eliminate the reflected light from the laser entering the optical fiber. In this study, the sampling interval of the fluorescence spectra was 0.5 nm. For different EWs, the normalized fluorescence intensity varied with emission wavelength as shown in Fig. 1.
Fig. 1 The normalized leaf fluorescence spectra of paddy rice excited by three different EWs with the changes in LNCs: (a) 355 nm; (b) 460 nm; (c) 556 nm.
2.3 Measurements of fluorescence spectra
The fluorescence spectra excited by different EWs were measured using the LIF system. In order to eliminate the influence of the geometry of the optical fiber and system, all fluorescence spectra were normalization processing. The fluorescence spectra excited by 355 nm EWs ranged from 360 nm to 800 nm and were normalized to 1 at 460 nm [Fig. 1(a)]. The fluorescence spectra excited by 460 nm laser ranged from 640 nm to 790 nm and were normalized to 1 at 685 nm [Fig. 1(b)]. When the 556 nm laser was used as excitation light source, the fluorescence spectra range changed from 640 nm to 800 nm and was normalized to 1 at 735 nm [Fig. 1(c)]. Based on the analysis of fluorescence parameters and on previous researches [23], the following fluorescence parameters can be used to estimate the LNC: F735/F460, F685/F460 and F735/F685 (F460: fluorescence intensity at 460 nm; F685: fluorescence intensity at 685 nm; F735: fluorescence intensity at 735 nm; F735/F460: fluorescence intensity ratio at 735 nm divided by that at 460 nm; F685/F460: fluorescence intensity ratio at 685 nm divided by that at 460 nm; F735/F685: fluorescence intensity ratio at 735 nm divided by that at 685 nm) in this study.
2.4 Measurement of leaf nitrogen content
Paddy rice leaves were destructively sampled by randomly cutting six leaves with three replicates for each experimental field, and the second leaves from the top were fully expanded. These leaves were sealed in plastic sacks, stored in an ice chest, and immediately transported to the laboratory for fluorescence spectra measurements. Afterward, all samples were shortly delivered to the Wuhan Academy of Agricultural Science and Technology for the LNC measurements. In this study, Kjeldahl approach was used to evaluate the LNC [24].
3. Analytical methods
3.1 Principal component analysis
PCA is one of the most powerful statistical multivariate analysis technologies whose main objective is to reduce the dimensionality of the spectra and extract the most important characteristic variables by analyzing internal correlation of data suitable for spectral analysis. PCA computes new variables called principal components (PC) which are obtained as linear combinations of the original variables. It can employ fewer new variables than the original ones to work out the difficult analysis [25]. In this study, this is because the spectra often contain large amounts of redundant, irrelevant information. Therefore, PCA was utilized to analyze the fluorescence spectra.
3.2 Support vector machine
SVM is a classical supervised learning algorithm with a strong theoretical foundation in statistical theory and can convert low-dimensional characteristics to high-dimensional characteristics for complex targets recognition [20]. The SVM exhibits special advantages in heterogeneous classes for small samples as well as nonlinear and high dimensional cases. It has displayed an excellent generalization performance (accuracy on test sets) in practice. The detailed description of SVM can be referenced [18, 26]. Kernel function is a critical factor for SVM analysis. In this study, radial basis function was utilized and can be presented as follows:
where, is a kernel parameter, denotes the test data and represents the training inputs.
3.3 Back-propagation neural network
BPNN is an efficient tool for the prediction of nonlinearities and comprise individual processing units called neurons that resemble neural activity [21]. All independent neurons can act together, and this allows them to analyze and solve a wide variety of complex tasks simultaneously [22]. BPNN is one of the most frequently utilized supervised training algorithms. It has been implemented in pattern recognition, agricultural research, biological, and classification [22, 27], and a brief introduction to BPNN can be referenced [22, 28]. Therefore, the BPNN model was used to estimate the paddy rice LNC based on fluorescence characteristics excited by different EWs in this study.
3.4 Performance analysis parameters
MATLAB 2009a was utilized to process these fluorescence spectra data. Before analysis, wavelet transform was utilized to eliminate the noises of the spectra. The wavelet transform is similar to the Fourier transform with a completely different merit function. The ability of the wavelets to provide multiresolution, low entropy, and makes them an ideal tool for studying fluorescence spectra [29]. The fluorescence characteristics of each EW were randomly divided into two parts, namely, 70% as the train set and another 30% as the validation set for LNC prediction. The coefficient of determination (R2), root mean square error (RMSE), and relative error (RE) in the prediction were employed to analyze the performance of the model. The RMSE and RE can be written as follows:
where n corresponds to the number of samples, Pi demotes predicted values, and Mi represents measured values. represents the mean of the measured values, and RE is the relative difference between the predicted and observed value. High R2 and low RMSE and RE denote high precision and accuracy of the model to predict the LNC [30].4. Results and discussion
4.1 Fluorescence spectra
The fluorescence spectra excited by different EWs were shown in Fig. 1.
All of the fluorescence spectra exhibited two main fluorescence peaks at 680-690 nm and at 730-740 nm. The center wavelengths of the two fluorescence peaks were 685 and 735 nm. Based on previous investigations [31], the fluorescence peak at 685 nm was attributed to chlorophyll a of Photosystem II, whereas the 735 nm was responsible for the antenna chlorophyll of Photosystems I and II. In addition, the fluorescence spectra excited by 355 nm laser also displayed a fluorescence peak at 460 nm and a peak shoulder at 525 nm. Nicotinamide adenine dinucleotide (NADPH) and riboflavin contributed to the fluorescence peak and the peak shoulder, respectively. As shown in Fig. 1, different EWs result in different fluorescence spectral characteristics [8–11]. The possible interpretation is that the fluorescence is re-absorbed on its way towards the leaf surface, which has been investigated in detail [32]. With the excitation light going from blue to red (440-635 nm), the band at 685 nm is reduced with respect to the 735 nm one. Since, for the wavelengths employed, transmittance of excitation inside the leaf increases from blue to green light, green-excited fluorescence receives contributions from deeper layers than the blue-excited fluorescence [33]. Therefore, the re-absorption process by the chlorophyll pigments in the upper layer leaf cells to the emission fluorescence spectra between 730 and 740 nm was much lower than that between 680 and 690 nm [34]. Chlorophyll content is associated with the re-absorption process influenced by the LNC. Thus, the re-absorption increased along with LNC. According to previous investigations [15, 23], the LNC was closely related to the fluorescence peaks (685 and 735 nm). Figures 1(a) and 1(b) demonstrate that the intensity of the fluorescence peaks increases as LNC increases. In Fig. 1(c), the fluorescence peak at 685 nm showed the opposite changing trend compared with that in Figs. 1(a) and 1(b) [23]. Thus, these fluorescence parameters can be employed to estimate LNC in paddy rice.
4.2 Relationship of fluorescence parameters to paddy rice LNC
As shown in Fig. 1, the different EWs resulted in different fluorescence spectral shapes and the different LNCs resulted in different fluorescence characteristics at fluorescence peaks (685 and 735 nm). Then, the correlation between these fluorescence parameters and the LNC was established and was shown in Fig. 2.
Fig. 2 The relationship between LNC and fluorescence parameters excited by different EWs. (n = 432). The red solid line represents linear regression.
Figure 2 demonstrates that the fluorescence parameters excited by different EWs displayed a closely positive linear correlation with LNC, which is consistent with findings of previous investigations [35, 36]. Thus, LIF technology can be utilized to monitor paddy rice LNC. When 355, 460 and 556 nm lasers were used as excitation light sources, the R2 of the linear regression statistics between the fluorescence characteristics F735/F685 and LNC were 0.70, 0.81, and 0.85, respectively [Figs. 2(c), 2(d) and 2(e)]. Meanwhile, when the 556 nm laser was used as the excitation light source, the correlation between the F735/F685 and LNC was better than that 460 and 355 nm excitation light sources. The may have resulted from the penetrating power of the 556 nm EW, which was stronger than that of the 460 and 355 nm EWs inside the leaf [34]. Furthermore, relative studies demonstrated that 556 nm EW is more suitable for chlorophyll fluorescence than 355 nm and 460 nm EWs.
In addition, the fluorescence spectra excited by the 355 nm laser contained more spectral information than those excited by 460 and 556 nm lasers. Thus, the fluorescence characteristics of F735/F460 and F685/F460 can also be utilized as indicators of nutrient stress [37]. The R2 of the linear regression statistics among the F685/F460, F735/F460, and LNC were 0.78 and 0.84 [Figs. 2(a) and 2(b)]. The correlation between the F735/F460 and LNC was superior to that between F685/F460 and LNC. Because the re-absorption process at 684-695 nm was stronger than that between 730 and 740 nm [35]. However, the details of the causes still need to be further studied in the future.
4.3 The PCA of fluorescence spectra
PCA was utilized in this study to analyze fluorescence spectra. The fluorescence spectra contained abundant verbose information which may influence on the estimation of LNC. Apparently, when the number of PCs exceeded three, the increase of the explained variance with additional PC was reduced to less than 1%. Therefore, the first three PCs were used for further study in this paper, and their explained variables and cumulative variances for different EWs were listed in Table 1.
Table 1. The percentage of explained variance for the first three PCs. EV: Explained variance; CV: Cumulative variance.
As shown in Table 1, 97.92%, 92.81%, and 94.84% of the total variance contained in the fluorescence spectra can be explained utilizing the first three PCs corresponding to 355, 460 and 556 nm EWs, respectively. In addition, other PCs included less spectral information that can be ignored. To better comprehend the efficiency of PCA in describing the LIF spectra, the loading plots of the first three PCs were shown in Fig. 3.
Fig. 3 The loading weights of the first three principal components with different EWs: (a) 355 nm; (b) 460 nm; (c) 556 nm.
Figure 3 displays that the first three PCs contained the main fluorescence spectral information, which is closely correlated with the LNC. PC1 was mainly attributed to the characteristics of fluorescence peaks at 685 and 735 nm, whereas PC2 and PC3 mainly complemented other fluorescence characteristics. In the 355 nm EW [Fig. 3(a)], PC2 and PC3 were attributed to the wavelengths 440, 480, 540, 680 and 740 nm. The 460 nm EW was responsible for the 670 700 730 and 740 nm [Fig. 3(b)]. The 556 nm EW corresponded to the 681, 683 and 690 nm wavelengths [Fig. 3(c)]. Thus, it further demonstrated that the first three PCs captured the most significant information in the fluorescence spectra, which can be applied in further analysis. Moreover, the factor scores calculated from the first three PCs were utilized as the input variables of PCA-BPNN and PCA-SVM models for predicting paddy rice LNC.
4.4 Estimation of LNC based on BPNN
To compare the predictive capacity of the different models and to analyze the effect of EWs on the estimation of LNC, the BPNN and SVM algorithms were utilized to inverse LNC based on different fluorescence parameters. The fluorescence spectra excited by each EW were randomly divided into two parts: 70% (n = 302) was served as the train set and the other 30% (n = 130) was utilized as the validation set. Based on different models, the relationship between the predicted and observed LNC were established and were illustrated in Fig. 4.
Fig. 4 The relationship between the predicted LNC using different fluorescence characteristics based on BPNN algorithm and observed LNC (n = 130) for different EWs. (a), (b), (c): 355 nm laser; (d), (e), (f): 460 nm laser; (g), (h), (i): 556 nm laser. (a), (d), (g): using fluorescence characteristic peaks as the input variables; (b), (e), (h): using fluorescence spectra as the input variables; (c), (f), (i): using the factor scores calculated from the first three PCs as the input variables. The dotted line denotes the 1:1 line and the red solid line represents linear regression.
As shown in Fig. 4, the predicted and observed LNC was nearly in accordance with the line of 1:1 (the dotted line). Ideally, the red solid line, which is the linear regression between the predicted and observed values, should be coincided with the 1:1 line. By comparing the R2, Fig. 4 demonstrates that FP-BPNN model for different EWs is superior to FS-BPNN model but is inferior to PCA-BPNN in monitoring paddy rice LNC. The new characteristics calculated by PCA served as input variables, which were better than that FP and FS based on the BPNN algorithm. The reason is that the correlation spectral variables, which are closely related to LNC, were selected from the fluorescence spectra by PCA. Thus, PCA has been extensively applied in spectral analyses and obtained encouraging results in the field of remote sensing [38, 39]. Furthermore, the experimental results showed that when the fluorescence spectra were served as input variables for FS-BPNN model, the estimation results were the worst among all the models. The reason is perhaps that the fluorescence spectra may have contained abundant verbose spectral information, which seriously affected the precision of inversion [21, 40].
In addition, the 355 nm EW (R2 = 0.80) outperformed the 460 and 556 nm EWs (R2 = 0.72, R2 = 0.75) in the monitoring of LNC on the basis of the FP-BPNN models [Figs. 4(a), 4(d) and 4(g)]. Because the fluorescence spectra excited by the 355 nm laser contained more fluorescence characteristic peaks (F735/F460, F685/F460 and F735/F685). For FS-BPNN models, the 556 nm EWs was better than the 355 and 460 nm EWs for the estimation of LNC in this study. However, when 355 and 556 nm lasers were served as excitation light sources displayed the approximate accuracy (R2 = 0.89, R2 = 0.88) during LNC estimation on the basis of PCA-BPNN model. Thus, the fluorescence spectra excited by 355 nm EWs combined with PCA-BPNN (R2 = 0.89) was superior to other models for the estimation of LNC with high R2 based on BPNN method. The reason is that the correlation spectral variables, which are closely related to LNC, were selected from the fluorescence spectra by PCA. Furthermore, the 556 nm EW coupled with PCA-BPNN model also exhibited promising potential for the estimation of paddy rice LNC (R2 = 0.88) due to the 556 nm excitation light penetrated deeper in the leaf tissue than the other EWs [11].
4.5 Estimation of LNC based on SVM
Using SVM method to estimate the LNC, the relationships between the predicted LNC based on different models and the observed LNC are shown in Fig. 5.
Fig. 5 The relationship between the predicted LNC using different fluorescence characteristics based on SVM algorithm and the observed LNC (n = 130) for different EWs. (a), (b), (c): 355 nm laser; (d), (e), (f): 460 nm laser; (g), (h), (i): 556 nm laser. (a), (d), (g): using fluorescence peaks as the input variables; (b), (e), (h): using fluorescence spectra as the input variables; (c), (f), (i): using the factor scores calculated from the first three PCs as the input variables. The dotted line denotes the 1:1 line and the red solid line represents linear regression.
Figure 5 shows the relationship between the predicted LNC using different models based on SVM method and observed LNC. The red solid lines show the linear regression between the predicted and observed values and the dotted lines represent the 1:1 line. By comparing the R2, the experimental results demonstrated that PCA-SVM model for the three EWs was superior to FS-SVM and FP-SVM models for the estimation of LNC. For the 355, 460, and 556 nm EWs, the R2 were 0.86, 0.80 and 0.85, respectively. Thus, among these fluorescence characteristics based on SVM algorithm, the new variables calculated by PCA were the most suitable for the inversion of LNC. The reason may be that these new variables calculated by the linear combination of fluorescence spectra have represented the most original variations, and the abundant verbose spectral information were eliminated [40, 41]. Based on the SVM algorithm, the FS was a better input variable than the FP for FS-SVM model when 460 and 556 nm lasers were served as excitation light sources. However, the opposite was observed for 355 nm EW.
For different EWs, Fig. 5 illustrates that when FP-SVM model and PCA-SVM model were utilized to estimate LNC, the 355 nm EWs (R2 = 0.74 and 0.86) outperformed the 460 and 556 nm EWs (R2 = 0.66, 0.80 and R2 = 0.68, 0.85) based on SVM algorithm. The fluorescence spectra excited by the 355 nm laser contained more fluorescence characteristic, which are closely linked to LNC [35, 42]. For FS-SVM, 556 nm was the optimal EW for the estimation of LNC in this study. Possibly, the 556 nm excitation light penetrated deeper in the leaf tissue than the other EWs [11, 32, 33]. Furthermore, related studies demonstrated that 355 and 460 nm EWs are more suitable for the blue-green fluorescence, whereas 556 nm is more suitable for the excitation of the chlorophyll fluorescence [11, 12].
4.6 Performance analysis of models
To comprehensively analyze the effect of the EWs, fluorescence parameters and different algorithms on the estimation of LNC based LIF technology, R2, RMSE, and RE in prediction based on different models were analyzed and are listed in Table 2.
Table 2. The performance analysis of different models for leaf nitrogen content (n = 130) based on coefficient of determination (R2), root mean square error (RMSE), and relative error (RE) in the prediction with different excitation wavelengths (EWs).
Table 2 shows the performance analysis of models based on BPNN and SVM algorithms. Table 2 demonstrates that the BPNN and SVM algorithms displayed promising potential for estimating LNC based on LIF technology. By comparing these results with higher R2 and lower RMSE and RE, the BPNN was found to be better than SVM for the estimation of LNC in this study. In addition, the variables calculated by PCA (PCA-BPNN and PCA-SVM models, the maximum R2 reached was 0.89 and 0.86, respectively) were superior to other fluorescence parameters. Moreover, the 355 and 556 nm lasers were better than that the 460 nm laser for the estimation of paddy rice LNC [11]. Therefore, the proper EW combined with the appropriate multivariate analysis should be considered based on the practical application requirements. In this study, 355 and 556 nm EWs combined with PCA-BPNN model were more suitable for the estimation of LNC with high R2 and low RMSE and RE.
In this study, the effect of EWs, fluorescence characteristics and multivariate analysis methods on the estimation of LNC was analyzed in detail. The potential of the fluorescence spectra combined with multivariate analysis for the quantitative monitoring of LNC was demonstrated. However, this preliminary study only compared the EWs, and fluorescence parameters with the help of multivariate analysis for the estimation of LNC based on LIF technology, and some limitations should be considered. For the BPNN, the optimal network architecture and training approaches have not been discussed and established. The results obtained using BPNN method were trained on the basis of our experiences in analyzing restricted data combinations. Ideally, the best results should be compared by changing the numbers of hidden neurons and network architectures. Thus, further studies should be done on the effectiveness of the BPNN for estimating LNC based LIF technology. Moreover, the effect of different kernel functions of SVM on the estimation of LNC should also be analyzed in future studies. In order to improve the robustness of this model and to obtain a solid conclusion, although different nitrogen treatments were set in this study, different paddy rice cultivars, whole growth seasons of vegetation, different genotypes, and other crops should also be considered in future studies [25].
5. Conclusion
In conclusion, this investigation demonstrated that fluorescence spectra excited by different EWs with the help of multivariate analysis displayed different predictive abilities for estimating LNC from leaf scale in paddy rice. The leaf fluorescence spectra of paddy rice excited by different EWs were measured and exhibited different spectral characteristics. BPNN and SVM combined with PCA were utilized to analyze these fluorescence spectra. The experimental results demonstrated that these spectral characteristics, which were calculated by PCA, served as input variables for BPNN and SVM model and were better than FP and FS for estimating LNC based on LIF technology. Among the six different models, PCA-BPNN was the optimal model for estimating LNC with high R2 and low RMSE and RE in this study. For the 355, 460, and 556 nm EWs, R2 was 0.89, 0.80 and 0.88, respectively. In addition, the 355 and 556 nm EWs were superior to 460 nm when the same model was utilized to predict LNC. This investigation showed the potential of the LIF technology combined with PCA as well as BPNN and SVM algorithms for the estimation of LNC from leaf level. This research also provided reference for the application of multivariate analysis on fluorescence spectra. Although the 355 nm EW outperformed 556 nm for the estimation of LNC, the outstanding results of PCA-BPNN model indicated that the fluorescence spectra excited by 556 nm EW with the help of multivariate analysis also have promising potential for the estimation of LNC based on LIF technology. To obtain a solid conclusion, however, more investigations are still needed to be done with more algorithms, rice cultivars, and other crops in the following works.
Funding
This work was supported by National Natural Science Foundation of China (Grant No. 41601360), Fundamental Research Funds for the Central Universities (Grant No. 2042016kf0008), Natural Science Foundation of Hubei Province (Grant No. 2015CFA002) and Open Research Fund of State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (Grant No.15R01).
Acknowledgments
The authors wish to thank College of Plant Science & Technology of Huazhong Agricultural University for providing the experimental samples. We wish to thank the Wuhan Academy of Agricultural Science & Technology for providing the LNCs of samples.
References and links
1. X. Liang, H. Li, S. Wang, Y. Ye, Y. Ji, G. Tian, C. van Kessel, and B. Linquist, “Nitrogen management to reduce yield-scaled global warming potential in rice,” Field Crops Res. 146, 66–74 (2013). [CrossRef]
2. S. S. Malhi, R. Lemke, Z. Wang, and B. S. Chhabra, “Tillage, nitrogen and crop residue effects on crop yield, nutrient uptake, soil quality, and greenhouse gas emissions,” Soil Tillage Res. 90(1-2), 171–183 (2006). [CrossRef]
3. P. Nugroho, M. Shimizu, H. Nakamato, A. Nagatake, S. Suwardi, U. Sudadi, and R. Hatano, “Nitrous oxide fluxes from soil under different crops and fertilizer management,” Plant Soil Environ. 61, 385–392 (2015).
4. E. W. Chappelle, F. M. Wood, J. E. McMurtrey, and W. W. Newcomb, “Laser-induced fluorescence of green plants. 1: A technique for the remote detection of plant stress and species differentiation,” Appl. Opt. 23(1), 134–138 (1984). [CrossRef] [PubMed]
5. H. M. Kalaji, A. Jajoo, A. Oukarroum, M. Brestic, M. Zivcak, I. A. Samborska, M. D. Cetner, I. Łukasik, V. Goltsev, and R. J. Ladle, “Chlorophyll a fluorescence as a tool to monitor physiological status of plants under abiotic stress conditions,” Acta Physiol. Plant. 38(4), 1–11 (2016). [CrossRef]
6. M. Brestic, M. Zivcak, H. M. Kalaji, R. Carpentier, and S. I. Allakhverdiev, “Photosystem II thermostability in situ: environmentally induced acclimation and genotype-specific reactions in Triticum aestivum L,” Plant Physiol. Biochem. 57, 93–105 (2012). [CrossRef] [PubMed]
7. Z. Tuba, D. K. Saxena, K. Srivastava, S. Singh, S. Czobel, and H. M. Kalaji, “Chlorophyll a fluorescence measurements for validating the tolerant bryophytes for heavy metal (Pb) biomapping,” Curr. Sci. 98, 1505–1508 (2010).
8. S. Apostol, A. A. Viau, and N. Tremblay, “A comparison of multiwavelength laser-induced fluorescence parameters for the remote sensing of nitrogen stress in field-cultivated corn,” Can. J. Rem. Sens. 33(3), 150–161 (2007). [CrossRef]
9. J. Yang, W. Gong, S. Shi, L. Du, B. Zhu, J. Sun, and S. Song, “Excitation Wavelength Analysis of Laser-Induced Fluorescence LiDAR for Identifying Plant Species,” IEEE Geosci. Remote Sens. Lett. 13(7), 977–981 (2016). [CrossRef]
10. J. Yang, W. Gong, S. Shi, L. Du, J. Sun, and S. Song, “Laser-Induced Fluorescence Characteristics of Vegetation by a New Excitation Wavelength,” Spectrosc. Lett. 49(4), 263–267 (2016). [CrossRef]
11. J. Schweiger, M. Lang, and H. K. Lichtenthaler, “Differences in Fluorescence Excitation Spectra of Leaves between Stressed and Non-Stressed Plants,” J. Plant Physiol. 148(5), 536–547 (1996). [CrossRef]
12. S. Svanberg, “Fluorescence lidar monitoring of vegetation status,” Phys. Scr. T58, 79–85 (1995). [CrossRef]
13. Z. Kolber, D. Klimov, G. Ananyev, U. Rascher, J. Berry, and B. Osmond, “Measuring photosynthetic parameters at a distance: laser induced fluorescence transient (LIFT) method for remote measurements of photosynthesis in terrestrial vegetation,” Photosynth. Res. 84(1-3), 121–129 (2005). [CrossRef] [PubMed]
14. M. Živcak, K. Olsovska, P. Slamka, J. Galambošová, V. Rataj, H. Shao, and M. Brestič, “Application of chlorophyll fluorescence performance indices to assess the wheat photosynthetic functions influenced by nitrogen deficiency,” Plant Soil Environ. 60, 210–215 (2014).
15. J. Yang, W. Gong, S. Shi, L. Du, J. Sun, Y. Y. Ma, and S. L. Song, “Accurate identification of nitrogen fertilizer application of paddy rice using laser-induced fluorescence combined with support vector machine,” Plant Soil Environ. 61(11), 501–506 (2015). [CrossRef]
16. H. M. Kalaji, G. Schansker, M. Brestic, F. Bussotti, A. Calatayud, L. Ferroni, V. Goltsev, L. Guidi, A. Jajoo, P. Li, P. Losciale, V. K. Mishra, A. N. Misra, S. G. Nebauer, S. Pancaldi, C. Penella, M. Pollastrini, K. Suresh, E. Tambussi, M. Yanniccari, M. Zivcak, M. D. Cetner, I. A. Samborska, A. Stirbet, K. Olsovska, K. Kunderlikova, H. Shelonzek, S. Rusinowski, and W. Bąba, “Frequently asked questions about chlorophyll fluorescence, the sequel,” Photosynth. Res. 318, 1–54 (2016). [PubMed]
17. M. P. Cendrero-Mateo, M. S. Moran, S. A. Papuga, K. R. Thorp, L. Alonso, J. Moreno, G. Ponce-Campos, U. Rascher, and G. Wang, “Plant chlorophyll fluorescence: active and passive measurements at canopy and leaf scales with different nitrogen treatments,” J. Exp. Bot. 67(1), 275–286 (2016). [CrossRef] [PubMed]
18. Y. Ma and W. Gong, “Evaluating the performance of SVM in dust aerosol discrimination and testing its ability in an extended area,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 5(6), 1849–1858 (2012). [CrossRef]
19. M. Buscema, “Back propagation neural networks,” Subst. Use Misuse 33(2), 233–270 (1998). [CrossRef] [PubMed]
20. M. A. Hearst, S. T. Dumais, E. Osman, J. Platt, and B. Scholkopf, “Support vector machines,” Intelligent Syst. Appl. 13(4), 18–28 (1998). [CrossRef]
21. R. Zong, Y. Zhi, B. Yao, J. Gao, and A. A. Stec, “Classification and identification of soot source with principal component analysis and back-propagation neural network,” Aust. J. Forensic Sci. 46(2), 224–233 (2014). [CrossRef]
22. A. I. Samborska, V. Alexandrov, L. Sieczko, B. Kornatowska, V. Goltsev, D. C. Magdalena, and H. M. Kalaji, “Artificial neural networks and their application in biological and agricultural research,” J. Nano PhotoBioSciences 2, 14–30 (2014).
23. A. A. Gitelson, C. Buschmann, and H. K. Lichtenthaler, “The chlorophyll fluorescence ratio F 735/F 700 as an accurate measure of the chlorophyll content in plants,” Remote Sens. Environ. 69(3), 296–302 (1999). [CrossRef]
24. Y. C. Tian, X. Yao, J. Yang, W. X. Cao, D. B. Hannaway, and Y. Zhu, “Assessing newly developed and published vegetation indices for estimating rice leaf nitrogen concentration with ground- and space-based hyperspectral reflectance,” Field Crops Res. 120(2), 299–310 (2011). [CrossRef]
25. Q.-X. Yi, J.-F. Huang, F.-M. Wang, X.-Z. Wang, and Z.-Y. Liu, “Monitoring rice nitrogen status using hyperspectral reflectance and artificial neural network,” Environ. Sci. Technol. 41(19), 6770–6775 (2007). [CrossRef] [PubMed]
26. M. K. Gill, T. Asefa, M. W. Kemblowski, and M. McKee, “Soil moisture prediction using support vector machines,” J. Am. Water Resour. Assoc. 42(4), 1033–1046 (2006). [CrossRef]
27. V. Goltsev, I. Zaharieva, P. Chernev, M. Kouzmanova, H. M. Kalaji, I. Yordanov, V. Krasteva, V. Alexandrov, D. Stefanov, S. I. Allakhverdiev, and R. J. Strasser, “Drought-induced modifications of photosynthetic electron transport in intact leaves: analysis and use of neural networks as a tool for a rapid non-invasive estimation,” Biochim. Biophys. Acta 1817(8), 1490–1498 (2012). [CrossRef] [PubMed]
28. L. E. Keiner and X.-H. Yan, “A neural network model for estimating sea surface chlorophyll and sediments from thematic mapper imagery,” Remote Sens. Environ. 66(2), 153–165 (1998). [CrossRef]
29. N. Agarwal, S. Gupta, A. Pradhan, K. Vishwanathan, and P. K. Panigrahi, “Wavelet transform of breast tissue fluorescence spectra: a technique for diagnosis of tumors,” IEEE J. Sel. Top. Quantum Electron. 9(2), 154–161 (2003). [CrossRef]
30. W. Feng, X. Yao, Y. Zhu, Y. Tian, and W. Cao, “Monitoring leaf nitrogen status with hyperspectral reflectance in wheat,” Eur. J. Agron. 28(3), 394–404 (2008). [CrossRef]
31. N. Tremblay, Z. Wang, and Z. G. Cerovic, “Sensing crop nitrogen status with fluorescence indicators. A review,” Agron. Sustain. Dev. 32(2), 451–464 (2012). [CrossRef]
32. M. E. Ramos and M. G. Lagorio, “True fluorescence spectra of leaves,” Photochem. Photobiol. Sci. 3(11-12), 1063–1066 (2004). [CrossRef] [PubMed]
33. G. Agati, “Response of the in vivo chlorophyll fluorescence spectrum to environmental factors and laser excitation wavelength,” Pure Appl. Opt.: J. Euro. Opt. Soc. A 7(4), 797–807 (1998). [CrossRef]
34. D. L. Farkas, A. S. Gouveia-Neto, J. E. A. Silva, E. B. Costa, L. A. Bueno, L. M. H. Silva, M. M. C. Granja, M. J. L. Medeiros, T. J. R. Câmara, L. G. Willadino, D. V. Nicolau, and R. C. Leif, “Plant abiotic stress diagnostic by laser induced chlorophyll fluorescence spectral analysis of in vivo leaf tissue of biofuel species,” Int. Soc. Opt. Photonics 7568, 75680G (2010).
35. Z. Malenovský, K. B. Mishra, F. Zemek, U. Rascher, and L. Nedbal, “Scientific and technical challenges in remote sensing of plant canopy reflectance and fluorescence,” J. Exp. Bot. 60(11), 2987–3004 (2009). [CrossRef] [PubMed]
36. C. Buschmann, “Variability and application of the chlorophyll fluorescence emission ratio red/far-red of leaves,” Photosynth. Res. 92(2), 261–271 (2007). [CrossRef] [PubMed]
37. F. Heisel, M. Sowinska, J. A. Miehé, M. Lang, and H. K. Lichtenthaler, “Detection of nutrient deficiencies of maize by laser induced fluorescence imaging,” J. Plant Physiol. 148(5), 622–631 (1996). [CrossRef]
38. H. Shahabi, B. B. Ahmad, M. H. Mokhtari, and M. A. Zadeh, “Detection of urban irregular development and green space destruction using normalized difference vegetation index (NDVI), principal component analysis (PCA) and post classification methods: a case study of Saqqez City,” Int. J. Phys. Sci. 7, 2587–2595 (2012).
39. R. Cavalli, G. Licciardi, and J. Chanussot, “Archaeological structures using nonlinear principal component analysis applied to airborne hyperspectral image,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 6(2), 659–669 (2013). [CrossRef]
40. N. Neeti and J. R. Eastman, “Novel approaches in Extended Principal Component Analysis to compare spatio-temporal patterns among multiple image time series,” Remote Sens. Environ. 148, 84–96 (2014). [CrossRef]
41. H. Abdi and L. J. Williams, “Principal component analysis,” Wiley Interdiscip. Rev. Comput. Stat. 2(4), 433–459 (2010). [CrossRef]
42. J. Yang, S. Shi, W. Gong, L. Du, Y. Y. Ma, B. Zhu, and S. L. Song, “Application of fluorescence spectrum to precisely inverse paddy rice nitrogen content,” Plant Soil Environ. 61(4), 182–188 (2015). [CrossRef]
