Chlorophyll Content Retrieval of Rice Canopy with Multi-spectral Inversion Based on LS-SVR Algorithm

Jin Si-yu, Su Zhong-bin, Xu Zhe-nan, Jia Yin-jiang, Yan Yu-guang, and Jiang Tao

College of Electrical and Information, Northeast Agricultural University, Harbin 150030, China

Abstract: To monitor growth and predict the yield of rice over a large area, the chlorophyll contents in the rice canopy were estimated using the unmanned aerial vehicle (UAV) remote sensing technology. In this work, multi-spectral image information of the rice crop was obtained using a 6-channel multi-spectral camera mounted on afixed wing UAV, which wasflown 600 m above the ground, between 11: 00-14: 00 on a sunny day in summer. The measured chlorophyll values were collected as sample sets. The s-REP index was screened out to estimate chlorophyll contents through the analysis of six kinds of spectral indexes of chlorophyll estimated capacity. An inversion model of the chlorophyll contents was then built using the least square support vector regression (LS-SVR)algorithm, with calibration and prediction R-square values of 0.89 and 0.83, respectively. Finally, remote sensing mapping for a UAV image of the Fangzheng County Dexter Rice Planting Park was accomplished using the inversion model. The inversion and measured values were then compared using regressionfitting. R-square and root-mean-square error of thefitting model were 0.79 and 2.39,respectively. The results demonstrated that accurate estimation of rice-canopy chlorophyll contents was feasible using the LS-SVR inversion model developed using the s-REP vegetation index.

Key words: remote sensing, chlorophyll, rice, UAV, multi-spectral, inversion, LS-SVR

Introduction

Extant studies have shown that there exists a good correlation among chlorophyll contents and plant nitrogen status, development stage, and photosynthetic capacity, which are good indicators of the strength of plant photosynthetic capacity, nutritional and physiological status of crops, and ageing process of plants. As such, determination of plant chlorophyll contents is important in precision agriculture and yield assessment (Yuan et al., 2007). Over the past decade, significant research on accurate estimation of plant chlorophyll contents has been conducted by using remote sensing spectroscopy and many noteworthy achievements have been made in the sensitive band selection and modeling (Curran et al.,1990; Filella et al., 1994). Daughtry et al. (2000)analyzed a variety of spectral indices including MCARI/OSAVI to estimate chlorophyll concentration accuracy. An inversion model of chlorophyll contents was established in the leaves of spruce canopy based on a large amount of experimental data by Zhang et al (2008). In addition to the analyses of existing indices, some researchers developed new indices to estimate the chlorophyll contents for specific objects of their respective studies. The first derivative ratio FD (730/525) and normalized value FD (730-525)/(730+525) of wheat spectra in 730 and 525 nm were constructed, and a more accurate result was obtained using an inversion model of the chlorophyll contents (Liang et al., 2012; Song et al.,2005). In recent years, the use of unmanned aerial vehicles (UAVs) has rapidly increased in agricultural activities. Data mosaicking technology has matured in view of the high volume of remote sensing images captured by small UAV (Wang and Chen, 2009;Luan et al., 2007). A small UAV is very suitable for research in the field of precision agriculture because of its ability to capture high-resolution images at low cost and with low risk. Although studies on chlorophyll content estimation using spectral indices have been previously conducted in considerable details, at present, most studies remain in the feature spectrum phase. Furthermore, there is few research on chlorophyll content estimation based on UAV remote sensing images.

In this study, multi-spectral images were obtained from a remote sensing system installed on the fixedwing UAV (TY-06) equipped with a 6-channel multispectral camera (Micro-MCA6 Snap). These images were used to develop a variety of spectral indices and establish an inversion model with the most optimal index to achieve the objective of accurate chlorophyll content estimation of the rice canopy. The maximum likelihood classification method was then used to extract the rice area. A remote-sensing map of the chlorophyll contents in rice was drawn to provide accurate and timely information support for growth diagnosis and management of rice crops to facilitate precise management of rice production.

Materials and Methods

Research location

The study was performed in Fangzheng County in the south-central region of the Heilongjiang Province,which is located in the middle reaches of the south bank of the Songhua River, the Zhangguangcai Offshoot of the Changbai Mountain West Lingbei Segment, and Dragonfly River. Geographic longitudinal and lateral coordinates of the location are 128°13'41"-129°33'20" and 45°32'46"-46°09'00",respectively. The region has a cold temperate continental monsoon type with an average annual rainfall of 579 mm, and the average number of annual sunshine hours is 4 446 h. The growth period offield crops in this region is from May to September with the total seasonal sunshine hours of 1 178 h and a sunshine percentage of 54% received for an average duration of 8 h per day.

The aerial region was mainly in Dexter Rice Planting Park in Fangzheng Country. Different varieties of rice were planted in the experimental area from May 6 to 10, 2015. The entire crop appeared to be in good condition, and all the paddies were in the early stages offilling. In the park, thefield was neatly divided into blocks, with convenient transportation to help remote sensing image mosaicking and ground data acquisition. Meanwhile, there were various complete material information resources, which provided a reliable guarantee for later application research in agricultural monitoring.

Air data acquisition

The data were obtained from a UAV remote sensing experiment conducted at the Founder Rice Research Institute on August 13, 2016 between 11: 00 a.m. and 14: 00 p.m. It was a sunny day, with a gentle breeze.Afixed-wing UAV TY-06 equipped with a 6-channel multi-spectral camera was used as the remote sensing platform and operating at an altitude of 600 m above the ground, with longitudinal and side overlaps of 75%and 60%, respectively. A total of 703 orthographic multi-spectral images were captured with an impressive useful image count of 696. Specifications of the multi-spectral camera used in the experiment are listed in Table 1.

Ground data acquisition

Sixty-four sampling areas were set up in the test region(Fig. 1). These were further divided into 44 training areas where the samples were used for training the inversion model and 20 test areas for testing the accuracy of the model. Fifty samples were recorded in each sampling area. Ground data for the study were collected on all the days among August 13-15, 2016 using instruments, such as the plant nutrient analyzer TYS-4N and a hand-held GPS. Chlorophyll values were measured using the plant nutrient analyzer,which recorded the soil plant analysis development(SPAD) values in the middle of the rice canopy leaf as the chlorophyll values of the samples. Sampling cell coordinates were recorded by the hand-held GPS unit.

Table 1 Specifications of Micro-MCA6 Snap Camera

Fig. 1 Demographic of the study area and sampling site distribution

Data preprocessing

The multi-spectral images were subjected to procedures, such as image stitching and rectification,using Pix4Dmapper software. ENVI5.1 software was then used for subsequent geometric correction and radiometric calibration. Geometric correction was processed using 30 ground control points and coordinates near the research institute. The geometric correction error was confirmed to be less than 0.5 pixel by checking the image. The radiometric calibration employed a pseudo-standard ground object radiometric calibration method, which adjusted multispectral images by measuring the reflectance of the ground calibration (reflectivity=3% and 48%), and the multi-spectral images demonstrated a relatively realistic reflectivity thereafter.

Data analysis

Numerous studies have demonstrated that there was a strong correlation between certain features of vegetation spectral bands and photosynthetic pigments(Blackburn et al., 1999; Sims and Gamon, 2002;Haboudane et al., 2002; Viña et al., 2011; Peng and Gitelson, 2012; Kong et al., 2014; Zaman-Allah et al.,2015). In the visible region (400-700 nm) of the electromagnetic spectrum, the spectral reflectance of the leaves was mainly determined by the presence of chlorophyll because there was a negative correlation between the two. The "green peak" at 550-560 nm was often used as an interference indicator to eliminate or weaken any effective radiations induced by nonphotosynthetic substances. In the vicinity of 670 nm,the red light absorption peak of chlorophyll was often used to characterize changes in chlorophyllsensitive contents. In the near-infrared region (700-1 000 nm), the boundary of the red color near 730 nm also reflected sensitive chlorophyll information with increased reflectance (Filella and Penuelas, 2007). So far, researchers had developed certain spectral indices,such as NDVI705, RVI, REP, PRI, and MCARI/OSAVI that could sensitively characterize plant chlorophyll contents.

Since the position of the red edge was usually defined as the inflection point of the red infrared slope, an accurate determination required a large number of spectral measurements in very small bands in this region. Clevers (1994) developed a method to get the red edge location using the reflectivity at 670,700, 740, and 780 nm. In this method, the inflection point was approximated by fitting a curve to fewer measurements. This method could be described as the followings:

REP=700+40[(ρ670+ρ780)/2-ρ700]/(ρ730-ρ710) (1)

Studies had shown that the first derivative spectra of rice canopy had obvious "double peaks" in the range of 680-760 nm, of which the main peak was concentrated near 730 nm and the second peak was concentrated near 718 nm. The position of red edge of rice canopy reflectance spectra was between 724 and 733 nm, during the entire growth period (Xie et al.,2010). In this study, the Micro-MCA6 Snap Camera captured the Red Band at 680, 710, 730 nm and Near Red Band at 800 nm. Based on the above researches,the following s_REP calculation method was proposed in this study:

For this study, six spectral indices were been selected, which had clear physical counterparts and a high degree of recognition (Table 2).

Table 2 Formulae of vegetation indexes for retrieving chlorophyll

LS-SVR algorithm

Support vector machines (SVM), proposed by Vapnik,were based on the principle of minimum structural risk in statistical learning theory (Wu and Sun, 2006).Least squares support vector machines (LS-SVMs)employed a modified SVM algorithm with good stability and fast calculation speed. Compared to other SVM algorithms, the parameters to be selected in LSSVR were few. Amongst these parameters, selection of the kernel function type, penalty coefficient (C),and kernel function parameter (G) had a significant effect on the inversion model. In the proposed study,the model optimization algorithm was selected based on the above advantages.

lnversion workflows

The workflow of the chlorophyll content retrieval process for rice is illustrated in Fig. 2. In the first step, various types of spectral indices were calculated using the pre-processed UAV images. The optimum index was then calculated by comparing and analyzing R2and the RMSE values, and the inversion model was established by least squares support vector regression (LS-SVR). In the final step, target-area mapping was achieved in the UAV image by using the inversion model established, and an accuracy test was conducted by comparing the mapping results with ground truth data (Qin et al., 2016; Wang et al.,2014). Data processing and remote sensing mapping were completed using ENVI5.1, MATLAB 2014, and ARCGIS10.1 software platforms.

Results

Correlation between spectral index and canopy leaf chlorophyll contents

To select the best index, the model coefficient (R²) and root mean square error (RMSE) values were compared and analyzed for different inversion models. The results of this comparison are summarized in Table 3.In decreasing order of estimation index accuracy,the candidate inversion models could be arranged as s-REP, RVI, NRI, GNDVI, NPCI, and NDVI. All the models were based on linear equation, level of significance less than 0.01 (p-value less than 0.01). Compared to other indexing models, values of R2and RMSE for the proposed three-band index (s-REP)model (0.85 and 3.20, respectively), demonstrated higher level of precision. Meanwhile, the constructed NDVI spectral index model demonstrated low inversion accuracy.

Fig. 2 Flow chart of rice chlorophyll content inversion

Table 3 Chlorophyll content inversion models and their spectral index evaluation indicators

Chlorophyll content prediction power of each index

When vegetation indices were used to estimate crop physical and chemical parameters, there were some common problems encountered in the sense that some indices became saturated by changes in measured parameters (such as chlorophyll contents), became increasingly less sensitive to such changes, and eventually lost response. To test the ability of each index to reliably predict the chlorophyll contents in different samples, the samples were divided into three subsets according to the level of chlorophyll present.Subset a was the low-content sample set (SPAD＜35),subset b was the medium-content sample set (35≤SPAD＜46), and subset c was the high-content sample set (SPAD≥46). Various indexing models were then used to predict RMSE values. A useful evaluation index was one with consistent prediction results for each subset.

RMSE values of each index for different sample subsets are depicted in Fig. 3. For the models generated by different indexes, each subset's RMSE prediction varied with the changes in chlorophyll contents.The following conclusions could hence be drawn.

(1) For Sample a, prediction results of RVI model were the most accurate with a RMSE value of 2.45.For Samples b and c, prediction results of s-REP model were the most accurate with corresponding RMSE values of 1.00 and 0.96.

(2) For RVI model had the smallest range of 0.964,which showed that the estimation results of RVI to each sample subset had better consistency. The index NDVI, GNDVI and NPCI had poor estimation accuracy for sample subset a, and their RMSE was greater than three. The range of NDRE index was 1.004, the second was only to the RVI index. The index s_REP had less estimation accurate than that of RVI in sample subset a, but it was the best for sample subset b and c.

(3) There was no significant difference in the performance of the six indices for Sample c.

Index model test

To investigate the accuracy of each index inversion model, a validation set (20 samples) was used to test the models on the basis of different spectral index variables. Fig. 1 showed the spatial distribution of the measured and predicted values, predictive correlation coefficients (predicted R and P-R²), and RMSE values of the predictions made. The equations were tested at the 0.01 significance level. The results demonstrated that the s-REP had the highest prediction accuracy,whereas the prediction accuracy of two indices—NDVI and NPCI—was found to be relatively poor.Thus, it was once again verified that these indices were not suitable for chlorophyll inversion of rice at the earlyfilling stage.

From the above analysis, it could be concluded that the s-REP model, having higher accuracy and better robustness, was the optimal index for characterizing the rice-canopy chlorophyll contents at the early stages of grainfilling.

Fig. 3 RMSE values of prediction results for sample subsets with different chlorophyll concentrations

Table 4 Chlorophyll content detection models and their spectral index evaluation indicators

LS-SVR model structure and test

To further improve the inversion accuracy, the LSSVR algorithm was used to optimize the proposed model. The parameters of LS-SVR model were set as the followings. The kernel type was the radial basis function kernel, the penalty coefficients C and radial basis function kernel parameter G were determined by cross validation,and the remaining parameters were set to the default values in LS_SVR (Gu et al., 2010).

To test the accuracy of the LS-SVR model, the measured values werefitted with estimated results of the training set. The results are shown in Fig. 4, with an R² value of 0.8976, which was higher than that for the corresponding linear model (0.8580). At the same time, the RMSE value of 2.389 was found to be lower than that of the linear model (2.794). Moreover, the slope of 0.8976 was closer to 1, which was contrary to the linear-model case. This indicated that the estimation results of the LS-SVR model more closely approximated the measured values.

To verify the predictive power of the LS-SVR model,the measured values were fitted with the estimated results of the test set. The results are depicted in Fig. 5.The LS-SVR model was found to possess a more concentrated regression line, a slope close to unity for the regression equation, a smaller offset, higher determination coefficient, and lower RMSE. This indicated that the LS-SVR model not only had betterfitting accuracy, but also better predictive ability.

Multi-spectral remote sensing mapping of rice chlorophyll contents

A variety of geometric objects appear in the multispectral images captured by UAV. First, for the production of more accurate maps of chlorophyll in rice, the crop growth was classified according to the likelihood classification method of supervised classification in ENVI. The remote sensor mapping of the chlorophyll contents in rice was then accomplished using s-REP index model. Finally, by combining the growth of rice grain filling and measured data characteristics, SPAD values of rice were divided into five grades and the chlorophyll retrieval of rice was performed in ARCGIS. Using the geographical coordinates, chlorophyll contents at any location in the rice covered area (Fig. 6) could be estimated. The overall growth of rice was good for the year. Level differences, however, demonstrated that there still existed some insufficiencies in field management measures for rice crops and that there was still some potentials for improvement.

Fig. 4 Measured value versusfitted value of linear and LS-SVR models in calibration set

Fig. 5 Measured value versus predicted value of linear and LS-SVR models prediction set

Accuracy test of remote sensing mapping

In order to assess the precision of UAV remote sensor image mapping, an accuracy test involving the use of ground truth data acquisition was conducted.Regression fitting was carried out using ground measurements and the corresponding point in the remote sensing map of rice chlorophyll contents. The results indicated high similarity between the two sets of data.

Thefitting model yielded an R2value of 0.79 and an RMSE of 2.39, illustrating that remote sensor mapp-ing of the chlorophyll contents in rice had higher precision. The inversion model based on s-REP was,therefore, applicable in the field of remote sensor analysis.

Fig. 6 Spatial distribution map of race chlorophyll contents

Fig. 7 Measured values of chlorophyll content versus inversion values

Discussion

Through the analysis presented in this paper, it was found that different spectral indices had considerable influence on inversion results. By comparing the same with previous analyses, it was noticed that the same spectral index had considerable influence on the precision of rice chlorophyll retrieval at different growth stages of the plant. The index with the highest inversion accuracy was the RVI (the two-band spectral index), with C-R2and P-R2values (0.84 and 0.72,respectively) slightly lower than the s-REP (threeband index). The corresponding RMSEC and RMSEP values were 3.25 and 2.46, respectively. It was unclear whether the accuracy of spectral indices constructed by the three-band index was generally higher than that of those constructed using two-band indices. In other words, the S-REP index might have a higher accuracy in the inversion of the rice-canopy spectrum and it was therefore necessary to conduct an in-depth analysis in this regard.

As for the models generated by different indices,each subset's prediction of RMSE varied with the changes in chlorophyll contents. For SPAD＜35, the inversion accuracy of each index inversion model was relatively low. For 35＜SPAD＜44, s_REP had the best estimation of chlorophyll contents. When SPAD＞44,the accuracies did not show a significant difference. It could be seen that in the case of the chlorophyll content saturation, the correlation of the different vegetation indices and chlorophyll contents was almost the same.Chlorophyll contents changed with the growth of rice.In general, when the contents of chlorophyll increased,the red edge shifted in the longwave direction; when the contents of chlorophyll decreased, the red edge shifted in the shortwave direction (Chen et al., 2010;Xie et al., 2010). Therefore, further researches were needed to investigate the applicability of s-REP index to other growth periods of rice and other crops.

Although the test results were obtained in a field test, they showed the impact of a single factor. In actual field production, various factors, such as surrounding conditions, moisture conditions, and management techniques might affect the spectral characteristics of rice. In the future, it was needed to improve the accuracy and universality of model estimation through the extensive examination and improvement of different ecological points, productivity levels, and cultivation conditions so as to promote direct application in rice growth monitoring. The multi-spectral camera used in this study obtained the most effective 6-channel spectral information bands for target factors, having more effective bands compared to an ADC camera. Therefore, the low-altitude remote sensing system had greater practicability and broad prospects in terms of crop yield estimation, regional disaster assessment of agricultural production, and crop growth in largefields.

Conclusions

In this paper, an inversion model of rice chlorophyll was established using the spectral index s-REP based on UAV spectral images and ground measured data. The chosen spectral index demonstrated C-R²and P-R² values of 0.89 and 0.83, respectively, and corresponding RMSEC and RMSEP values were determined to be 2.59 and 2.16. Thus, the proposed model possessed high accuracy and good predictive ability. Analysis showed that s-REP was the preferred index for inversion of chlorophyll in rice, because it could estimate the chlorophyll contents in the initial filling of rice more accurately and could effectively avoid the effect of sample range and other factors on the estimation of chlorophyll contents. Remote sensor mapping of UAV images using the s-REP spectral inversion model was completed. The fitting model, obtained by comparing the mapping results and the ground-measured values (with R2=0.79 and RMSE=2.39), demonstrated that the remote-sensing estimation results were accurate and could provide scientific basis of rapid and nondestructive testing for the rice.

The research presented in this paper was limited to the rice-growth information of a specific growth stage. In future studies, rice crop models would be established using a vegetation index to analyze the relationship between crop quantity and the onset of large-scale rice (such as rice blast) to build an early warning system for related diseases of rice.