Rice Leaf Chlorophyll Content Estimation Using UAV-Based Spectral Images in Different Regions

: Estimation of crop biophysical and biochemical characteristics is the key element for crop growth monitoring with remote sensing. With the application of unmanned aerial vehicles (UAV) as a remote sensing platform worldwide, it has become important to develop general estimation models, which can interpret remote sensing data of crops by different sensors and in different agroclimatic regions into comprehensible agronomy parameters. Leaf chlorophyll content (LCC), which can be measured as a soil plant analysis development (SPAD) value using a SPAD-502 Chlorophyll Meter, is one of the important parameters that are closely related to plant production. This study compared the estimation of rice ( Oryza sativa L.) LCC in two different regions (Ningxia and Shanghai) using UAV-based spectral images. For Ningxia, images of rice plots with different nitrogen and biochar application rates were acquired by a 125-band hyperspectral camera from 2016 to 2017, and a total of 180 samples of rice LCC were recorded. For Shanghai, images of rice plots with different nitrogen application rates, straw returning, and crop rotation systems were acquired by a 5-band multispectral camera from 2017 to 2018, and a total of 228 samples of rice LCC were recorded. The spectral features of LCC in each study area were analyzed and the results showed that the rice LCC in both regions had signiﬁcant correlations with the reﬂectance at the green, red, and red-edge bands and 8 vegetation indices such as the normalized difference vegetation index (NDVI). The estimation models of LCC were built using the partial least squares regression (PLSR), support vector regression (SVR), and artiﬁcial neural network (ANN) methods. The PLSR models tended to be more stable and accurate than the SVR and ANN models when applied in different regions with R 2 values higher than 0.7 through different validations. The results demonstrated that the rice canopy LCC in different regions, cultivars, and different types of sensor-based data shared similar spectral features and could be estimated by general models. The general models can be implied to a wider geographic extent to accurately quantify rice LCC, which is helpful for growth assessment and production forecasts.

Spectral cameras can obtain both spectra (usually in the range of visible to near infrared) and images of targets. Strictly calibrated spectra are closely related to plant paper, the estimation of rice LCC using UAV hyperspectral and multispectral images in Ningxia and Shanghai, China, was studied to (1) investigate the spectral response characteristics of rice LCC in different geographic regions and by different sensors, and (2) establish general estimation models of rice LCC for different geographic regions and sensors.

Field Design and Plant Material
The field experiments were conducted in Yesheng, Qingtongxia, Ningxia (38°07′28″ N, 106°11′37″ E) and Zhuanghan, Fengxian District, Shanghai (30°33′25″ N, 121°23′17″ E), China, respectively (Figure 1a). Yesheng, Ningxia has a temperate arid climate, which is denoted by BSk in the Köppen climate classification system. The average annual temperature is about 8.5 °C, the annual precipitation is about 220 mm, and the annual sunshine hours are between 2800 to 3000 h. The landform type is the Yellow River alluvial plain, and the soil type is anthropogenic alluvial soil. The Yellow River water is introduced for irrigation. The land is fertile and suitable for rice growth. The rice variety planted in the experimental field of Ningxia was Ningjing No. 37. The rice was grown under three nitrogen (N) regimes in combination with 4 biochar (B) levels. Nitrogen regimes were: (ⅰ) 0 kg ha −1 (N0), (ⅱ) 240 kg ha −1 (N1), and (ⅲ) 300 kg ha −1 (N2). Biochar levels were: (ⅰ) 0 kg ha −1 (B0), (ⅱ) 4500 kg ha −1 (B1), (ⅲ) 9000 kg ha −1 (B2), and (ⅳ) 13,500 kg ha −1 (B3). The phosphate and potash fertilizers, which were applied equally to all plots were P2O5 (90 kg ha −1 ) and K2O (90 kg ha −1 ), respectively. Treatments were arranged in a split-plot experimental design with three replicates. The size of each plot was 14 m × 5 m. Yesheng, Ningxia has a temperate arid climate, which is denoted by BSk in the Köppen climate classification system. The average annual temperature is about 8.5 • C, the annual precipitation is about 220 mm, and the annual sunshine hours are between 2800 to 3000 h. The landform type is the Yellow River alluvial plain, and the soil type is anthropogenic alluvial soil. The Yellow River water is introduced for irrigation. The land is fertile and suitable for rice growth. The rice variety planted in the experimental field of Ningxia was Ningjing No.37. The rice was grown under three nitrogen (N) regimes in combination with 4 biochar (B) levels. Nitrogen regimes were: (i) 0 kg ha −1 (N0), (ii) 240 kg ha −1 (N1), and (iii) 300 kg ha −1 (N2). Biochar levels were: (i) 0 kg ha −1 (B0), (ii) 4500 kg ha −1 (B1), (iii) 9000 kg ha −1 (B2), and (iv) 13,500 kg ha −1 (B3). The phosphate and potash fertilizers, which were applied equally to all plots were P 2 O 5 (90 kg ha −1 ) and K 2 O (90 kg ha −1 ), respectively. Treatments were arranged in a split-plot experimental design with three replicates. The size of each plot was 14 m × 5 m.
Zhuanghang, Shanghai is located in the alluvial plain of the Yangtze River Delta with a flat terrain and intersecting river networks and has a subtropical marine monsoon climate (Cfa in Köppen-Geiger classification). The average annual temperature is about 15.8 • C, the annual precipitation is about 1221 mm, and the annual sunshine hours is about 1920 h. The soil type is paddy soil. These climatic and soil conditions make it a traditional rice-growing area. The rice variety planted in the experimental field in Shanghai was Huhan No.61. There were 36 split plots. The size of each plot was 7 m × 8 m. Three different experiments were designed in these plots. For 18 plots, 6 N fertilizer treatments were set up with 3 replicates as follows: (i) no N fertilizer (N0), chemical N fertilizer at the rate of (ii) 100 kg ha −1 (N100), (iii) 200 kg ha −1 (N200), (iv) 300 kg ha −1 (N300), and combinations of chemical and organic N fertilizers at the rates of (v) 200 kg ha −1 (ON200), and (vi) 300 kg ha −1 (ON300). For the other 12 plots, treatments were performed in triplicate as follows: (i) no fertilizer application (CK), (ii) conventional inorganic fertilizer application with 200 kg N ha −1 (CF), (iii) the same total pure nitrogen application as in CF plus 3000 kg ha −1 straw returning to the field (CS), and (iv) the same total pure nitrogen application as in CF plus 1000 kg ha −1 straw-derived biochar returning to the field (CB). For the rest 6 plots and 3 CF plots, 3 different crop rotation systems were applied with 3 replicates: (i) single rice rotation (R), (ii) rice-Chinese milk vetch rotation (RC), and (iii) rice-winter wheat rotation (RW), which were applied in 3 CF plots, and 200 kg N ha −1 were applied to the rice. The definitions of acronyms used here are listed in Table 1. The hyperspectral images of the rice canopy were captured in Ningxia. The sensor used here was a Cubert S185 hyperspectral imager (Cubert GmbH, Ulm, Germany) mounted on an octocopter UAV. The spectral range of S185 is 450-950 nm with 125 bands, and the spectral sampling interval is 4 nm. It acquires a hyperspectral image with 50 by 50 pixels for each band. The flight height was 70 m, resulting in a spatial resolution of 0.53 m. The forward and side overlaps were set at 85% and 80%, respectively. One frame of hyperspectral image was captured every second for about 12 min in each flight mission. Five flight missions were conducted in the rice field at different growth stages on 19 July 2016, 16 August 2016, 9 July 2017, 10 August 2017 and 11 September 2017. All flight missions were performed between 11:00 and 13:00 in cloudless weather.
The multispectral images of the rice canopy were captured in Shanghai. The sensor used here was a Micasense RedEdge 3 multispectral camera (Micasense Inc., Seattle, WA, USA) mounted on a DJI M600 Pro UAV (SZ DJI Technology Co., Shenzhen, China). RedEdge 3 acquires five-band images with 1280 by 960 pixels at five discrete wavelength ranges: blue (475 ± 20 nm), green (560 ± 20 nm), red (668 ± 20 nm), red-edge (−717 ± 10 nm), and near infrared (NIR) (840 ± 40 nm). The flight height was 100 m, yielding a spatial resolution of 0.06 m. The forward and side overlaps were set at 85% and 80%, respectively. One multispectral image was captured every second for about 15  Spectral images were radiometrically calibrated according to the reflectance of the reference panels on the ground. The reference panel of S185 (Cubert GmbH, Ulm, Germany) is a white board with 100% reflectance by the size of 50 cm × 50 cm. The reference panel of RedEdge 3 (Micasense Inc., Seattle, WA, USA) is a gray board with 50% reflectance and a size of 15.5 cm × 15.5 cm. The hyperspectral images of the whole rice field in Ningxia were generated by mosaicking the images within the aerial survey area using Agisoft PhotoScan Professional (Agisoft LLC, St. Petersburg, Russia). The multispectral image of the whole rice field in Shanghai was generated by mosaicking the images within the aerial survey area using Pix4DmapperPro (PIX4D, Lausanne, Switzerland). Geometric correction and projection transformation were applied to the mosaicked image according to the ground The number of GCPs was eight in Ningxia and ten in Shanghai. The plot boundaries were drawn according to the images in ArcGIS and buffered by 0.5 m to create the plot boundary files for spectra extraction. The mean spectra of all pixels inside the buffered plot boundary were calculated for each plot using ENVI 5.6 software (Harris Geospatial Solutions, Inc., Broomfield, CO, America).
Vegetation indices are usually calculated by linear or nonlinear combination of reflectance data at different bands and designed to maximize sensitivity to the vegetation characteristics while minimizing confounding factors such as soil background reflectance, directional, or atmospheric effects, which will cause fluctuation and noise in reflectance [35,36].
In this study, we tested dozens of broad-band vegetation indices that were commonly used for LCC estimation. Eight of them, which are applicable for both the hyperspectral imager (S185) and multispectral camera (RedEdge 3), were selected for this study ( Table 2). In the calculation of vegetation indices for the hyperspectral images by S185, we chose the spectral reflectance at the wavelength of 475 nm for the blue band, 560 nm for the green band, 668 nm for the red band, 717 nm for the red-edge band, and 840 nm for the NIR band. Table 2. Vegetation indices used in this study.

Vegetation Index Acronym Equation Reference
Normalized difference vegetation index R represents the reflectance value in specified bands.

Rice LCC Measurement
Twenty flag leaves of rice away from the edge were selected randomly in each plot for LCC measurment by the SPAD-502 (Minolta Camera Co., Osaka, Japan) right after the flight mission on the day of every UAV survey. The average SPAD value of the 20 leaves was calculated as the LCC of a given plot, and the SPAD value of each plot was recorded as a sample. Finally, a total of 180 samples in Ningxia and 228 samples in Shanghai were recorded.

Correlational Analysis
Correlational analysis was employed to investigate the response of spectral reflectance and vegetation indices to rice LCC. The Pearson's correlation coefficients (r) between LCC and the spectral parameters (reflectance and vegetation indices) in Ningxia and Shanghai was calculated using Equation (1): where n is the sample size, x i and y i are the individual sample i, x i is the mean value of all x samples, and analogously for y i . An absolute value of r closer to 1 indicates a better correlation between x and y.

Regression Analysis
Three groups of datasets were established from the collected data: Ningxia, Shanghai, and Ningxia-Shanghai. The Ningxia group contained 180 samples. The Shanghai group contained 228 samples. The Ningxia-Shanghai group, which was the combination of the Ningxia and Shanghai groups, contained 408 samples. Each dataset was split into two parts of sub-datasets, with 2/3 for model calibration (Cal) and the rest, 1/3 for validation (Val). The Ningxia-Shanghai calibration (Ningxia-Shanghai_Cal, 272 samples) sub-dataset consisted of the Ningxia calibration (Ningxia_Cal, 120 samples) and Shanghai calibration (Shanghai_Cal, 152 samples) sub-datasets. The Ningxia-Shanghai validation (Ningxia-Shanghai_Val, 136 samples) sub-dataset consisted of the Ningxia validation (Ningxia_Val, 60 samples) and Shanghai validation (Shanghai_Val, 76 samples) sub-datasets.
For each of the three groups of datasets, rice LCC estimation models were established by taking the vegetation indices as independent variables and using three methods: PLSR, SVR, and Artificial Neural Network (ANN). PLSR is a multiple linear regression method integrating principal component analysis, correlation analysis, and canonical correlation analysis. PLSR helps to build stable models by effectively eliminating multiple correlations between the independent variables and extracting the composite variables, which are most explanatory of the dependent variables [44][45][46][47]. SVR is a machine-learning regression algorithm based on statistical learning theory. By using kernel functions, the input data is mapped into a higher-dimensional space, in which the optimal regression models are constructed [48]. ANN is a nonlinear machine learning algorithm that mimics the structure and function of biological neural networks [49].
In this study, the number of components for the PLSR model was determined when the overall mean predicted R 2 , calculated based on the predicted residual sum of squares (PRESS) reached a maximum in the LeaveOne-Out Cross Validation (LOOCV) scheme. For SVR models, the Gaussian function was adopted as the kernel, and the grid search method and cross validation were used to determine the parameters. For ANN models, the multilayer feedforward neural network (including the input layer, hidden layer, and output layer) was used to build the SPAD estimation models, and the training algorithm was chosen as Levenberg-Marquardt and the hidden layer number was set as 5.
The coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute percentage error (MAPE) were used to evaluate the predictive performance of each model. The equations of R 2 , RMSE, and MRE are presented in Equations (2)-(4): where n is the number of samples, y i andŷ i is the measured and predicted value of sample i, y i is the average value of all samples. A higher R 2 value (close to 1) with a lower RMSE and MAPE (close to 0) indicate a better model accuracy. Two kinds of validation were applied to test the prediction accuracy and generalization ability of the different models. First, the models were validated within the groups in which they were built. Second, the models were validated interactively, that is to say, validating the models built based on the Ningxia (Shanghai) group using data from the Shanghai (Ningxia) group, by which the most stable modelling method could be revealed. All these statistical analyses were performed using the MATLAB R2018a software (MathWorks, Natick, MA, USA). Table 3 shows the summary statistics for the measured LCC of rice leaves in Ningxia and Shanghai. The statistical characteristics of rice LCC showed similar tendencies in these two study areas.

Response of Spectral Reflectance and Vegetation Indices to Rice LCC
The correlation between the spectral reflectance at each band and rice LCC in Ningxia and Shanghai showed a similar pattern in the wavelength range of 450-720 nm with significant negative correlations (|r| > 0.4), as shown in Figure 2. The highest correlations appeared in the red bands in both Ningxia and Shanghai, with r = −0.88 at 670 nm for Ningxia and r = −0.80 at 668 nm for Shanghai. At the NIR range, the correlations between spectral reflectance and LCC were slightly positive for the rice in Ningxia and significantly positive for the rice in Shanghai with r = 0.44. Table 3 shows the summary statistics for the measured LCC of rice leave and Shanghai. The statistical characteristics of rice LCC showed similar te these two study areas.

Response of Spectral Reflectance and Vegetation Indices to Rice LCC
The correlation between the spectral reflectance at each band and rice LCC and Shanghai showed a similar pattern in the wavelength range of 450-720 n nificant negative correlations (|r| > 0.4), as shown in Figure 2. The highest appeared in the red bands in both Ningxia and Shanghai, with r = −0.88 at Ningxia and r = −0.80 at 668 nm for Shanghai. At the NIR range, the correlatio spectral reflectance and LCC were slightly positive for the rice in Ningxia and s positive for the rice in Shanghai with r = 0.44.  Table 4, all of the eight vegetation indices were significantl with rice LCC with an absolute r value higher than 0.4 in both Ningxia an  Table 4, all of the eight vegetation indices were significantly correlated with rice LCC with an absolute r value higher than 0.4 in both Ningxia and Shanghai. NPCI had a high negative correlation with LCC. The other seven vegetation indices had positive correlations with LCC. NDVI and NPCI both responded to LCC better than the other vegetation indices in Ningxia with an r value of 0.75 and −0.81, respectively. The r values of RVI, GNDVI, and RENDVI were between 0.6 and 0.7, while those of GRVI, REVI, and MTCI were below 0.6. For rice in Shanghai, the r values were all above 0.7, which indicated a good response of all of the vegetation indices to LCC.

Estimation Models of Rice LCC
Estimation models of rice LCC were constructed based on the calibration sub-datasets using PLSR, SVR, and ANN methods for Ningxia, Shanghai, and Ningxia-Shanghai, respectively. As shown in Table 5, all models yielded good performances, with R 2 values higher than 0.8, RMSE lower than 3.5, and MAPE lower than 10%. The models of the Shanghai group performed better than those of the Ningxia and Ningxia-Shanghai groups, with RMSE values lower than 2.2 and a MAPE below 5.4%. The models of Ningxia, in which the RMSE was higher than 2.8 and MAPE was above 8.2%, were less accurate compared with the other two groups. The performances of Ningxia-Shanghai models were at moderate levels. Two machine learning methods, SVR and ANN, were superior to PLSR, with an R 2 value between 0.88 and 0.90 and a smaller RMSE and MAPE.

Validation of Models within the Group
Validation sub-datasets for each group were used to evaluate the prediction accuracy of the LCC estimation models for their group. As shown in Table 6 and Figure 3, the models for Shanghai yielded the highest prediction accuracy, with an RMSE between 2.08 and 2.24 and MAPE between 5.11% to 5.72%. The models of Ningxia turned out to be lower in prediction accuracy, with an RMSE between 3.45 and 4.00 and MAPE larger than 10.1%. In terms of modeling methods, SVR and ANN models had better predictive ability than the PLSR model in each group, with an R 2 value between 0.84 and 0.86 and lower values of RMSE and MAPE. Scatter plots between measured and predicted LCC along the 1:1 line indicated a trend that lower LCC were overestimated by all the models in the Ningxia and Ningxia-Shanghai groups (Figure 3). of RMSE and MAPE. Scatter plots between measured and predicted LCC along the line indicated a trend that lower LCC were overestimated by all the models in the Ning and Ningxia-Shanghai groups (Figure 3).

Interactive Validation of Models
In order to test whether the UAV-based rice LCC estimation models built in one study area could be used in another region with a different sensor, here we use all the data in the Ningxia group (Ningxia_All) to validate the rice LCC estimation models built based on the Shanghai group and all the data in the Shanghai group (Shanghai_All) to validate the models constructed based on the Ningxia group. The results are given in Table 7 and  in the Ningxia group (Ningxia_All) to validate the rice LCC estimation models built bas on the Shanghai group and all the data in the Shanghai group (Shanghai_All) to valid the models constructed based on the Ningxia group. The results are given in Table 7 a Figure 4. For both Ningxia and Shanghai, the PLSR models outperformed the SVR a ANN models in predictive accuracy and stability, with R 2 values of 0.74 and 0.

Discussion
In this study, we found that despite the differencse in the region, variety, and sens the correlation between rice LCC, spectral reflectance, and vegetation indices in Ning and Shanghai had similarities. This result was in accordance with the previous findin

Discussion
In this study, we found that despite the differencse in the region, variety, and sensor, the correlation between rice LCC, spectral reflectance, and vegetation indices in Ningxia and Shanghai had similarities. This result was in accordance with the previous findings by researchers on the spectral characteristics of rice LCC or chlorophyll content using visible and near-infrared spectrometers [50][51][52][53], which indicated that the spectral responses of rice LCC or chlorophyll content shared the same pattern, that is, the spectral reflectance and rice LCC had significant negative correlations at the wavelength range of 450-720 nm. The eight vegetation indices were found to be closely related to rice LCC, which was in accordance with the findings by Xie [50]. On the other side, differences in spectral reflectance in the NIR range between Ningxia and Shanghai can be observed. The most possible reason was that the Shanghai group contained more data collected in the maturation stage, when the correlation between LCC and spectral reflectance in the NIR range was higher than at other growth stages [54]. Similar phenomena were also found in other crops such as wheat, maize, and potatoes [25,55,56]. Since the field remote sensing data collection largely depended on weather conditions, the growth stages when the field campaigns were carried out were not unified in different study areas and years. The difference between Ningxia and Shanghai groups in growth stage also affected the response of vegetation indices and model accuracy.
Theoretically, the reflectance of the same target measured under standard conditions by different spectral sensors, which are strictly calibrated should be coincident. However, in the actual measurement operation, the spectral reflectance value is affected by bandresponse functions, spatial resolution, and environmental factors such as atmospheric transmissivity, solar zenith angle, solar declination angle, and soil background, which would cause a systematic difference [1,57]. Vegetation indices, which are designed to be particularly sensitive to vegetative covers, are less affected by those factors [58,59]. The rice fields in Ningxia and Shanghai are at different latitudes, with different soil backgrounds and solar illumination geometry. The spectral response functions and spatial resolution of S185 and Redege 3 also vary. In order to minimize the error in spectral reflectance caused by the environment and sensor type and make the variables available for both regions, vegetation indices rather than spectral reflectance were chosen as independent variables in the rice leaf SPAD estimation models.
The validation of rice LCC estimation models for the Ningxia-Shanghai group indicated that the models were capable of predicting LCC of rice in both Ningxia and Shanghai with relatively low errors, which answered the question of whether a certain crop trait could be predicted by a general model for different regions and sensors. In addition, the interaction validation of PLSR models of Ningxia and Shanghai in Section 3.4.2 also yielded good predicting results with MAPE less than 20%, which indicated that, based on the same spectral response pattern, a model of the same crop trait constructed for one region and spectral sensor using the PLSR method was still available for another region and sensor. It is impossible to build new models for every new region in the application of remote sensing monitoring of crop growth. In cases where no prior data is available, models transplanted from another region and sensors could be used to provide informative results.
The performances of models using different regression methods varied in validation steps. Models using two machine learning methods, SVR and ANN, achieved high accuracy when validated by the validation sub-dataset from the same group of the dataset (Section 3.4.1). However, the LCC prediction accuracy decreased dramatically when SVR or ANN models for one region were applied to the other, implying limited generalization ability (Section 3.4.2). PLSR models showed better prediction accuracy and stability in the interactive validation. Although machine learning is effective in data mining, the lack of interpretability may increase the uncertainty of the model [60]. Previous studies have suggested that the training data has a strong influence on the performance of machine learning methods, and extreme attention should be paid when applying regression models with machine learning algorithms to data collected under different conditions from the training data [61]. There are several possible explanations, as follows: The data from the two study areas had global features in common (for example, the same spectral response pattern as LCC), as well as local features of their own that were caused by the difference in location, cultivar, and sensor. Two machine learning methods, SVR and ANN, would make the most of both global and local features within the training dataset in the modeling procress to minimize prediction error. However, the difference in local features would lower the prediction accuracy when applying the machine learning-based models to another dataset. In contrast, the PLSR method mainly used global features in model construction, thus resulting in better generalization ability than SVR and ANN.

Conclusions
The application of UAV-based remote sensing to crop growth monitoring has gradually entered the practical stage. Therefore, it is of great importance to develop remote sensing models for crop monitoring, which are suitable for multiple regions and types of sensors. In this study, images of rice fields in two different regions were acquired by multispectral and hyperspectral cameras mounted on UAV platforms. Analysis of the spectral response of rice LCC in different regions showed a similar pattern: spectral reflectance at the green and red bands, as well as two vegetation indices (NDVI and NPCI), were highly correlated with the LCC. All the estimation models of rice LCC yielded good accuracy within the group where the training data came from. Models built with PLSR rather than SVR and ANN, which achieved outstanding performance in the interactive validation, had the potential to be used as general estimation models of rice LCC.
As far as we know, this was the first approach to building models based on UAV-based remote sensing data from different regions and sensors. It demonstrates that it is feasible to study general models for UAV-based retrieval of agronomic parameters. The outcomes of this study could be used by practitioners and organizations to develop sensors that can directly produce LCC or chlorophyll content distribution maps based on spectral images of rice fields, which is very intuitive and helpful for farmers.
This study involved two independent experiments, the conditions of which varied in many aspects. It is difficult to analyze the influence of specific factors on the LCC estimation. Therefore, this paper focused on finding the common points in the spectral response and estimation models of the two experiments. In the future, experiments with carefully controlled conditions should be designed for more in-depth studies on the influence of specific factors such as flight height, time, and weather.

Data Availability Statement:
The data presented in this study are available from the corresponding author upon reasonable request.