Estimating Above-Ground Biomass of Maize Using Features Derived from UAV-Based RGB Imagery

: The rapid, accurate, and economical estimation of crop above-ground biomass at the farm scale is crucial for precision agricultural management. The unmanned aerial vehicle (UAV) remote-sensing system has a great application potential with the ability to obtain remote-sensing imagery with high temporal-spatial resolution. To verify the application potential of consumer-grade UAV RGB imagery in estimating maize above-ground biomass, vegetation indices and plant height derived from UAV RGB imagery were adopted. To obtain a more accurate observation, plant height was directly derived from UAV RGB point clouds. To search the optimal estimation method, the estimation performances of the models based on vegetation indices alone, based on plant height alone, and based on both vegetation indices and plant height were compared. The results showed that plant height directly derived from UAV RGB point clouds had a high correlation with ground-truth data with an R 2 value of 0.90 and an RMSE value of 0.12 m. The above-ground biomass exponential regression models based on plant height alone had higher correlations for both fresh and dry above-ground biomass with R 2 values of 0.77 and 0.76, respectively, compared to the linear regression model (both R 2 values were 0.59). The vegetation indices derived from UAV RGB imagery had great potential to estimate maize above-ground biomass with R 2 values ranging from 0.63 to 0.73. When estimating the above-ground biomass of maize by using multivariable linear regression based on vegetation indices, a higher correlation was obtained with an R 2 value of 0.82. There was no signiﬁcant improvement of the estimation performance when plant height derived from UAV RGB imagery was added into the multivariable linear regression model based on vegetation indices. When estimating crop above-ground biomass based on UAV RGB remote-sensing system alone, looking for optimized vegetation indices and establishing estimation models with high performance based on advanced algorithms (e.g., machine learning technology) may be a better way.


Introduction
Agriculture is facing a great challenge in how to meet the increasing demand for agricultural production with limited soil and water resources due to the increasing population and rapidly growing economy [1]. By 2050, the projected population and socio-economic growth will double current food demand. To meet this challenge in developing countries, grain production needs to increase by 40% and net irrigation water requirements by 40-50% [2]. Maize is one of the most important food crops and (2) establish maize AGB estimation models based on PH alone by using linear and exponential regression analyses; based on VIs alone by using single and multivariable linear regression analyses; and based on both VIs and PH by using multivariable linear regression analysis; (3) comparatively analyze the performances of maize AGB estimating models and map the distribution of maize AGB by using the optimal estimating model.

Research Field
The experiment was carried out on a field in Zhaojun town located in the southwest of Inner Mongolia, China (40 • 26 0.29" N, 109 • 36 25.99" E, elevation 1010 m). The research field, approximately 1.13 ha, planted with maize, was divided into 5 large regions. Three 6 × 6 m 2 areas within each region were chosen as sampling plots (yellow square in Figure 1b) for data collection from three sampling sites (green square in Figure 2a, 1 × 1 m 2 ) within the sampling plot. The mixture of soil samples at depths of 30, 60, and 90 cm were used to estimate soil characteristics. The soil type was loamy sand (80.7% sand, 13.7% powder, and 5.6% clay), using the United States Department of Agriculture soil taxonomy. The soil pH, organic matter, and C content were 9.27, 47.17 g/Kg, and 27.35 g/Kg, respectively. Maize (2) establish maize AGB estimation models based on PH alone by using linear and exponential regression analyses; based on VIs alone by using single and multivariable linear regression analyses; and based on both VIs and PH by using multivariable linear regression analysis; (3) comparatively analyze the performances of maize AGB estimating models and map the distribution of maize AGB by using the optimal estimating model.

Research Field
The experiment was carried out on a field in Zhaojun town located in the southwest of Inner Mongolia, China (40°26′0.29″ N, 109°36′25.99″ E, elevation 1010 m). The research field, approximately 1.13 ha, planted with maize, was divided into 5 large regions. Three 6 × 6 m 2 areas within each region were chosen as sampling plots (yellow square in Figure 1b) for data collection from three sampling sites (green square in Figure 2a, 1 × 1 m 2 ) within the sampling plot. The mixture of soil samples at depths of 30, 60, and 90 cm were used to estimate soil characteristics. The soil type was loamy sand (80.7% sand, 13.7% powder, and 5.6% clay), using the United States Department of Agriculture soil taxonomy. The soil pH, organic matter, and C content were 9.27, 47.

Field Measurements
At the maize vegetative stage (V5-V7, DOY 171-185), plant height (PH) and above-ground biomass (AGB) were measured on DOY 171, 177, and 185, respectively. A total of 15 sets of samples (PH and AGB) were obtained each day. There were 45 total biomass observations scaled up to the unit of kg/m 2 . The heights of maize were measured manually within each sampling site (green square in Figure 2a) using a measuring tape. As shown in Figure 2b, the distance between the ground and the highest leaf of maize was measured as the PH within each sampling site during the vegetative stage [40]. Five PH measurements were taken within each sampling site and the average of total fifteen measurements from three sampling sites was used as reference plant height (PHref) for the sampling plot (yellow square in Figure 1b). For measuring AGB, maize samples were harvested from a 0.5 × 0.5 m 2 subplot adjacent to the sampling plot and then was immediately put into a sealed bag to avoid the loss of water. Electronic balance (DJ-2000A, Suzhou Jiangdong Precision Instrument CO., Ltd., Suzhou, China) with a measurement accuracy of 0.01 g was used to measure the weight of maize fresh and dry AGB. For dry AGB, maize sample was put into a DHG-9053A oven (HASUC, Shanghai, China) to de-enzyme under 105 °C for two hours and then was dried to a constant weight under 80 °C. The time required varied from 36 to 72 h depending on the mass of maize samples on each sampling day. Each subplot's measurement was referred to as the reference biomass of corresponding sampling plot. Figure 3 shows the main steps in the acquisition and pretreatment of UAV RGB imagery. A consumer-grade quadrotor UAV RGB remote-sensing system, DJI Phantom 4 Pro (Shenzhen Dajiang Baiwang Technology Co., Ltd., Shenzhen, China) was adopted to collect RGB imagery of maize. This UAV has an integrated camera, with a 1-inch complementary metal-oxide semiconductor sensor that captures red-green-blue spectral information. The camera has an 84° field of view lens with an f/2.8 aperture, and a resolution of 4864 × 3648 pixels. This lens is also specially designed to eliminate image distortion. Detailed information about the digital camera and the UAV system are given in Table 1. On DOY 171, 177, and 185 (sunny days) between 11:00-13:00 (Chinese standard time), RGB imagery of maize were captured by DJI Phantom 4 Pro. Parameters of ISO and shutter were set to 400 and 1/1250 according to specific weather condition, and white balance was set to sunny. Flight was controlled by Altizure software (Everest Innovation Technology Ltd., Hong Kong, China), which directed the UAV flying along a serpentine image acquisition plan at a height of 30 m and a speed of 1.3 m/s with the camera looking downwards. Overlap of imagery to the front and side was 90%. The ground sample distance was 0.8 cm. To improve the geo-location accuracy, the geo-referencing of the point clouds was done using a combination of direct geo-referencing and five ground control points (Figure 1b, green round). The coordinates of the ground control points were measured by a KOLIDA RTK differential GNSS device (KOLIDA Instrument Co., Ltd., Guangzhou, China).

Field Measurements
At the maize vegetative stage (V5-V7, DOY 171-185), plant height (PH) and above-ground biomass (AGB) were measured on DOY 171, 177, and 185, respectively. A total of 15 sets of samples (PH and AGB) were obtained each day. There were 45 total biomass observations scaled up to the unit of kg/m 2 . The heights of maize were measured manually within each sampling site (green square in Figure 2a) using a measuring tape. As shown in Figure 2b, the distance between the ground and the highest leaf of maize was measured as the PH within each sampling site during the vegetative stage [40]. Five PH measurements were taken within each sampling site and the average of total fifteen measurements from three sampling sites was used as reference plant height (PH ref ) for the sampling plot (yellow square in Figure 1b).
For measuring AGB, maize samples were harvested from a 0.5 × 0.5 m 2 subplot adjacent to the sampling plot and then was immediately put into a sealed bag to avoid the loss of water. Electronic balance (DJ-2000A, Suzhou Jiangdong Precision Instrument CO., Ltd., Suzhou, China) with a measurement accuracy of 0.01 g was used to measure the weight of maize fresh and dry AGB. For dry AGB, maize sample was put into a DHG-9053A oven (HASUC, Shanghai, China) to de-enzyme under 105 • C for two hours and then was dried to a constant weight under 80 • C. The time required varied from 36 to 72 h depending on the mass of maize samples on each sampling day. Each subplot's measurement was referred to as the reference biomass of corresponding sampling plot. Figure 3 shows the main steps in the acquisition and pretreatment of UAV RGB imagery. A consumer-grade quadrotor UAV RGB remote-sensing system, DJI Phantom 4 Pro (Shenzhen Dajiang Baiwang Technology Co., Ltd., Shenzhen, China) was adopted to collect RGB imagery of maize. This UAV has an integrated camera, with a 1-inch complementary metal-oxide semiconductor sensor that captures red-green-blue spectral information. The camera has an 84 • field of view lens with an f/2.8 aperture, and a resolution of 4864 × 3648 pixels. This lens is also specially designed to eliminate image distortion. Detailed information about the digital camera and the UAV system are given in Table 1. On DOY 171, 177, and 185 (sunny days) between 11:00-13:00 (Chinese standard time), RGB imagery of maize were captured by DJI Phantom 4 Pro. Parameters of ISO and shutter were set to 400 and 1/1250 according to specific weather condition, and white balance was set to sunny. Flight was controlled by Altizure software (Everest Innovation Technology Ltd., Hong Kong, China), which directed the UAV flying along a serpentine image acquisition plan at a height of 30 m and a speed of 1.3 m/s with the camera looking downwards. Overlap of imagery to the front and side was 90%. The ground sample distance was 0.8 cm. To improve the geo-location accuracy, the geo-referencing of the point clouds was done using a combination of direct geo-referencing and five ground control points (Figure 1b, green round). The coordinates of the ground control points were measured by a KOLIDA RTK differential GNSS device (KOLIDA Instrument Co., Ltd., Guangzhou, China). After acquisition, the image mosaic processing was performed by using Pix4DMapper software (Pix4DInc., Switzerland, https://www.pix4d.com/) which was specifically designed to process UAV imagery using techniques rooted in both computer vision and photogrammetry [41]. The whole processing workflow was: firstly, initializing all geo-located imagery captured in each flight by automatically searching for feature tie-points in match pairs of imagery and correcting them based on the camera model. After that, five ground control points were incorporated to correct the geographic coordinates of imagery. Then, the densified point clouds were generated with a 7 × 7 pixels match window. Finally, the digital surface model and orthophoto with a spatial resolution of 0.8 cm were generated by using the inverse distance weighting method. More detailed information about the image mosaic processing based on Pix4DMapper software could be seen in Duan et al. [42].    After acquisition, the image mosaic processing was performed by using Pix4DMapper software (Pix4DInc., Switzerland, https://www.pix4d.com/) which was specifically designed to process UAV imagery using techniques rooted in both computer vision and photogrammetry [41]. The whole processing workflow was: firstly, initializing all geo-located imagery captured in each flight by automatically searching for feature tie-points in match pairs of imagery and correcting them based on Remote Sens. 2019, 11, 1261 6 of 21 the camera model. After that, five ground control points were incorporated to correct the geographic coordinates of imagery. Then, the densified point clouds were generated with a 7 × 7 pixels match window. Finally, the digital surface model and orthophoto with a spatial resolution of 0.8 cm were generated by using the inverse distance weighting method. More detailed information about the image mosaic processing based on Pix4DMapper software could be seen in Duan et al. [42].

The Extraction of Maize Plant Height
At the vegetative stage (V5-V7, DOY 171-185), maize canopy did not reach full coverage. Therefore, point clouds include both soil and maize plants (Figure 4b). PH could be calculated by subtracting the ground level from estimated plant level derived from point clouds. As shown in Figure 4a, the 1st percentile of the cumulative probability distribution of elevation derived from the mixed point clouds was chosen to represent the ground level. To search for an optimal PH feature, the 90th, 95th, 97th, 98th, and 99th percentiles, and the maximum value of point clouds altitude within each sampling plot were selected to represent the plant level.

The Extraction of Maize Plant Height
At the vegetative stage (V5-V7, DOY 171-185), maize canopy did not reach full coverage. Therefore, point clouds include both soil and maize plants (Figure 4b). PH could be calculated by subtracting the ground level from estimated plant level derived from point clouds. As shown in Figure 4a, the 1st percentile of the cumulative probability distribution of elevation derived from the mixed point clouds was chosen to represent the ground level. To search for an optimal PH feature, the 90th, 95th, 97th, 98th, and 99th percentiles, and the maximum value of point clouds altitude within each sampling plot were selected to represent the plant level. Then, six PH features (PH90, PH95, PH97, PH98, PH99, and PHmax) for each sampling plot were obtained. Regression analyses between six PH features and PHref were adopted to search for the optimal PH feature. There were 45 total PHref values averaged from observations, 15 PHref values for each day. The estimation precisions of six PH features were evaluated by their coefficient of determination (R 2 ) and root mean square error (RMSE).

The Calculation of UAV RGB Vegetation Index
Many previous studies have used different vegetation indices (VIs) derived from RGB imagery to estimate the AGB of crops [40,[43][44][45][46][47]. In this study, five VIs and one VI combination were calculated using visible bands, including the normalized green-red difference index (NGRDI) [44], excess green index (ExG) [48], excess green minus excess red (ExGR), color index of vegetation (CIVE) [49], vegetation index (VEG) [50], and the combination of ExG, ExGR, CIVE and VEG (COM) [51]. Their calculation formulas are as follows: where RR, RG, and RB represent the digital numbers of red, green, and blue bands which were normalized by dividing 255 (8-bit image in this study) [52,53]. Figure 5 shows the main procedures to obtain the optimal estimation model of maize AGB based

The Calculation of UAV RGB Vegetation Index
Many previous studies have used different vegetation indices (VIs) derived from RGB imagery to estimate the AGB of crops [40,[43][44][45][46][47]. In this study, five VIs and one VI combination were calculated using visible bands, including the normalized green-red difference index (NGRDI) [44], excess green index (ExG) [48], excess green minus excess red (ExGR), color index of vegetation (CIVE) [49], vegetation index (VEG) [50], and the combination of ExG, ExGR, CIVE and VEG (COM) [51]. Their calculation formulas are as follows: where R R , R G , and R B represent the digital numbers of red, green, and blue bands which were normalized by dividing 255 (8-bit image in this study) [52,53]. Figure 5 shows the main procedures to obtain the optimal estimation model of maize AGB based on maize features derived from UAV RGB imagery. Three estimation models of maize AGB were adopted in this study, namely, estimation models based on (1) PH; (2) Vis; and (3) both PH and VIs. The PH feature derived from point clouds with the best performance and VIs with the better correlations were adopted. Linear regression analysis was used for VIs, and both linear and exponential regression analyses were used for PH. In the establishment of the third kind estimation model, multivariable linear regression analysis was adopted.

Estimation Models of Plant Height and Biomass
Remote Sens. 2019, 11, x FOR PEER REVIEW 7 of 21 adopted in this study, namely, estimation models based on (1) PH; (2) Vis; and (3) both PH and VIs. The PH feature derived from point clouds with the best performance and VIs with the better correlations were adopted. Linear regression analysis was used for VIs, and both linear and exponential regression analyses were used for PH. In the establishment of the third kind estimation model, multivariable linear regression analysis was adopted.

Statistical Analysis
For statistical analysis, the R programming language (R-3.4.3, https://www.r-project.org/) was adopted. More specifically, the lm () function was used to perform regression analyses and AIC ()function was used to calculate the Akaike Information Criterion (AIC). The coefficient of determination (R 2 ), the root mean square error (RMSE), and the AIC [54,55] were calculated for the comparison. R 2 is not suitable to evaluate the predictive and fitting performance among models with different forms and different numbers of parameters. Therefore, the AIC was further used for the comparison of models with the lower value representing the better estimation performance.

Statistical Analysis
For statistical analysis, the R programming language (R-3.4.3, https://www.r-project.org/) was adopted. More specifically, the lm () function was used to perform regression analyses and AIC ()function was used to calculate the Akaike Information Criterion (AIC). The coefficient of determination (R 2 ), the root mean square error (RMSE), and the AIC [54,55] were calculated for the comparison. R 2 is not suitable to evaluate the predictive and fitting performance among models with different forms and Remote Sens. 2019, 11, 1261 8 of 21 different numbers of parameters. Therefore, the AIC was further used for the comparison of models with the lower value representing the better estimation performance.

Comparison between Maize Plant Height Derived from UAV Point Clouds and Ground-Truth Values
The correlations between six PH features (PH 90 , PH 95 , PH 97 , PH 98 , PH 99 , and PH max ) and PH ref were analyzed ( Figure 6). The PH 90  also had the lowest extraction error with the RMSE value of 0.12 m. At the same time, it could be observed that when the 90th, 95th, 97th, 98th, and 99th percentiles of point clouds altitude were used to represent the maize plant level, a relatively lower estimation was obtained, compared to ground-truth PH with the fitting line below the 1:1 line ( Figure 6). With the increase of percentile of UAV RGB point clouds altitude, there was a decreasing trend of the difference between PH derived from point clouds and PH ref . When the maximum value of point clouds altitude was used to represent the maize plant level, a relatively higher estimation was obtained with the fitting line above the 1:1 line.

Comparison between Maize Plant Height Derived from UAV Point Clouds and Ground-Truth Values
The correlations between six PH features (PH90, PH95, PH97, PH98, PH99, and PHmax) and PHref were analyzed ( Figure 6). The PH90 and PH95 had the lowest correlations with PHref, with R 2 values of 0.20 and 0.51, respectively. The other four PH features (PH97, PH98, PH99 and PHmax) all had good correlations with PHref with R 2 values no less than 0.70. The PH99 had the highest correlation with PHref with an R 2 value of 0.90. The PH99 also had the lowest extraction error with the RMSE value of 0.12 m. At the same time, it could be observed that when the 90th, 95th, 97th, 98th, and 99th percentiles of point clouds altitude were used to represent the maize plant level, a relatively lower estimation was obtained, compared to ground-truth PH with the fitting line below the 1:1 line ( Figure  6). With the increase of percentile of UAV RGB point clouds altitude, there was a decreasing trend of the difference between PH derived from point clouds and PHref. When the maximum value of point clouds altitude was used to represent the maize plant level, a relatively higher estimation was obtained with the fitting line above the 1:1 line. Figure 6. Correlations between maize plant heights derived from UAV RGB point clouds and groundtruth plant heights (PHref). Maize plant height estimations were obtained by using the 90th (a); 95th (b); 97th (c); 98th (d); and 99th (e) percentiles; and maximum value (f) of point clouds altitude to represent the plant level, respectively. R 2 represents the coefficient of determination and RMSE represents the root mean square error.

Estimation Models of Maize Biomass Based on Plant Height Derived from UAV RGB Point Clouds
Both linear and exponential regression analyses were used to estimate maize fresh and dry AGB based on maize PH99, which had the best estimation performance. As shown in Figure 7

Estimation Models of Maize Biomass Based on Plant Height Derived from UAV RGB Point Clouds
Both linear and exponential regression analyses were used to estimate maize fresh and dry AGB based on maize PH 99, which had the best estimation performance. As shown in Figure 7, there were high positive correlations between PH 99 and fresh or dry AGB, with R 2 values no less than 0.59 (p < 0.01). For both fresh and dry AGB, the exponential regression model had higher correlations with R 2 values of 0.77 and 0.76, respectively, compared to the linear regression model with R 2 values both of 0.59. However, when RMSE and AIC were considered, the linear regression model had relatively  (a,c) were the linear and exponential regression models for fresh above-ground biomass, respectively; (b,d) were the linear and exponential regression models for dry above-ground biomass, respectively.

Estimation Models of Maize Biomass Based on UAV RGB Vegetation Indices
In this study, five VIs, NGRDI, ExG, ExGR, CIVE, VEG, and one VI combination (COM), derived from UAV RGB imagery were used to establish linear regression models for both maize fresh and dry AGB. As shown in Figures A1 and A2, and Table 2, except for CIVE, the other five VIs all had significant positive correlations (p < 0.01) with both maize fresh and dry AGB. For both maize fresh and dry AGB, the ExGR had the highest correlations with R 2 values both of 0.73, followed by the NGRDI and VEG with R 2 values of 0.70 and 0.68, and 0.65 and 0.63, respectively. Lowest correlations were observed by using ExG to estimate maize fresh and dry AGB, with R 2 values of 0.34 and 0.33, respectively. The COM also had a high correlation with both maize fresh and dry AGB with R 2 values of 0.73 and 0.72, respectively, due to the fact that the COM was a combination of ExG, ExGR, CIVE, Figure 7. Correlations between maize plant heights derived from UAV RGB point clouds and ground-truth fresh or dry above-ground biomass (AGB). Maize plant height (PH 99 ) observations were obtained by using the 99th percentile of point clouds altitude to represent the plant level. AIC represents the Akaike Information Criterion; (a,c) were the linear and exponential regression models for fresh above-ground biomass, respectively; (b,d) were the linear and exponential regression models for dry above-ground biomass, respectively.

Estimation Models of Maize Biomass Based on UAV RGB Vegetation Indices
In this study, five VIs, NGRDI, ExG, ExGR, CIVE, VEG, and one VI combination (COM), derived from UAV RGB imagery were used to establish linear regression models for both maize fresh and dry AGB. As shown in Figures A1 and A2, and Table 2, except for CIVE, the other five VIs all had significant positive correlations (p < 0.01) with both maize fresh and dry AGB. For both maize fresh and dry AGB, the ExGR had the highest correlations with R 2 values both of 0.73, followed by the NGRDI and VEG with R 2 values of 0.70 and 0.68, and 0.65 and 0.63, respectively. Lowest correlations were observed by using ExG to estimate maize fresh and dry AGB, with R 2 values of 0.34 and 0.33, respectively. The COM also had a high correlation with both maize fresh and dry AGB with R 2 values of 0.73 and 0.72, respectively, due to the fact that the COM was a combination of ExG, ExGR, CIVE, and VEG. When it comes to the RMSE, a similar observation was obtained. For maize fresh AGB, the ExGR also had the lowest RMSE with a value of 0.32, following by the NGRDI and VEG with RMSE values of 0.34 and 0.37. The CIVE had the largest RMSE with a value of 0.62. For the maize dry AGB, similar observations were found. Table 2. Correlations between vegetation indices derived from UAV RGB imagery and ground-truth fresh or dry above-ground biomass. Five vegetation indices and one vegetation index combination, normalized green-red difference index (NGRDI), excess green index (ExG), excess green minus excess red (ExGR), color index of vegetation (CIVE), vegetation index (VEG) and the combination of ExG, ExGR, CIVE, and VEG (COM), were used to establish linear regression estimation models for both maize ground-truth fresh and dry above-ground biomass.

Vegetation Index and Combination
Fresh After estimating maize fresh and dry AGB by using single vegetation index, three VIs, NGRDI, ExGR, and VEG, which had the highest correlations were chosen to estimate maize fresh and dry AGB by using multivariable linear regression analysis. When multivariable linear regression was adopted, the estimation performance of maize AGB was improved (Figure 8). For both fresh and dry AGB, the R 2 values were 0.82, with an increase of 0.09 compared to the highest R 2 values of 0.73 (ExGR) for the single vegetation index. The RMSE of fresh and dry AGB decreased with values of 0.06 and 0.01 kg/m 2 , compared to the lowest RMSE values of 0.32 and 0.04 kg/m 2 (ExGR) for single VI, respectively. The maize fresh and dry AGB could be calculated based on NGRDI, ExGR, and VEG by using Equations (7) and (8)  After estimating maize fresh and dry AGB by using single vegetation index, three VIs, NGRDI, ExGR, and VEG, which had the highest correlations were chosen to estimate maize fresh and dry AGB by using multivariable linear regression analysis. When multivariable linear regression was adopted, the estimation performance of maize AGB was improved (Figure 8). For both fresh and dry AGB, the R 2 values were 0.82, with an increase of 0.09 compared to the highest R 2 values of 0.73 (ExGR) for the single vegetation index. The RMSE of fresh and dry AGB decreased with values of 0.06 and 0.01 kg/m 2 , compared to the lowest RMSE values of 0.32 and 0.04 kg/m 2 (ExGR) for single VI, respectively. The maize fresh and dry AGB could be calculated based on NGRDI, ExGR, and VEG by using Equations (7) and (8) Figure 8. Correlations between maize above-ground biomass estimated based on NGRDI, ExGR, and VEG by using multivariable linear regression analysis and ground-truth above-ground biomass; (a) and (b) were maize fresh and dry above-ground biomass, respectively.

Figure 8.
Correlations between maize above-ground biomass estimated based on NGRDI, ExGR, and VEG by using multivariable linear regression analysis and ground-truth above-ground biomass; (a) and (b) were maize fresh and dry above-ground biomass, respectively.

Estimation Models of Maize Biomass Based on Both Plant Height and Vegetation Indices
After estimating maize fresh and dry AGB by using PH or VIs alone, multivariable linear regression analyses were adopted based on both PH and three VIs (NGRDI, ExGR, and VEG). As shown in Figure 9, when PH was added, the correlations with both maize fresh and dry AGB, and the estimation accuracy were improved slightly with an increase of 0.03 of R 2 for both maize fresh and dry AGB, and a decrease of 0.02 kg/m 2 of RMSE only for maize fresh AGB. At the same time, there were decreases of 6.23 and 7.76 of AIC for maize fresh and dry AGB, respectively.

Estimation Models of Maize Biomass Based on Both Plant Height and Vegetation Indices
After estimating maize fresh and dry AGB by using PH or VIs alone, multivariable linear regression analyses were adopted based on both PH and three VIs (NGRDI, ExGR, and VEG). As shown in Figure 9, when PH was added, the correlations with both maize fresh and dry AGB, and the estimation accuracy were improved slightly with an increase of 0.03 of R 2 for both maize fresh and dry AGB, and a decrease of 0.02 kg/m 2 of RMSE only for maize fresh AGB. At the same time, there were decreases of 6.23 and 7.76 of AIC for maize fresh and dry AGB, respectively. Figure 9. Correlations between maize above-ground biomass estimated based on NGRDI, ExGR, VEG, and plant height by using multivariable linear regression analysis and ground-truth aboveground biomass; (a,b) were maize fresh and dry above-ground biomass, respectively.

Mapping Maize Above-Ground Biomass Based on UAV RGB Remote-Sensing
Due to the insignificant improvement of the correlations and estimation accuracy when PH was added into the multivariable linear regression models, the distributions of both maize fresh and dry AGB at a farm scale were plotted by using VIs alone (Equations (7) and (8)). It could be observed that there were obvious increases of both maize fresh and dry AGB with the growth of maize ( Figure 10). For maize fresh AGB, the mean values for DOY 171, 177, and 185 were 0.85, 2.00, and 2.25 kg/m 2 , respectively. For maize dry AGB, the mean values for DOY 171, 177, and 185 were 0.11, 0.25, and 0.29 kg/m 2 , respectively. Figure 9. Correlations between maize above-ground biomass estimated based on NGRDI, ExGR, VEG, and plant height by using multivariable linear regression analysis and ground-truth above-ground biomass; (a,b) were maize fresh and dry above-ground biomass, respectively.

Mapping Maize Above-Ground Biomass Based on UAV RGB Remote-Sensing
Due to the insignificant improvement of the correlations and estimation accuracy when PH was added into the multivariable linear regression models, the distributions of both maize fresh and dry AGB at a farm scale were plotted by using VIs alone (Equations (7) and (8)). It could be observed that there were obvious increases of both maize fresh and dry AGB with the growth of maize ( Figure 10). For maize fresh AGB, the mean values for DOY 171, 177, and 185 were 0.85, 2.00, and 2.25 kg/m 2 , respectively. For maize dry AGB, the mean values for DOY 171, 177, and 185 were 0.11, 0.25, and 0.29 kg/m 2 , respectively.

Discussion
The unmanned aerial vehicle (UAV) remote-sensing platform as one of the most important farmscale data acquisition tools has a great application potential in the rapid, accurate, and economical estimation of crop above-ground biomass (AGB). It has been successfully used to estimate AGB of many crops, such as barely [32], sunflower [40], maize [38,56], wheat [36,37,57], etc. As the two main data sources to describe the growing status of crop, plant height (PH), and vegetation indices (VIs) have been widely used to estimate AGB of crops [23,35,58]. The UAV-based light detection and

Discussion
The unmanned aerial vehicle (UAV) remote-sensing platform as one of the most important farm-scale data acquisition tools has a great application potential in the rapid, accurate, and economical estimation of crop above-ground biomass (AGB). It has been successfully used to estimate AGB of many crops, such as barely [32], sunflower [40], maize [38,56], wheat [36,37,57], etc. As the two main data sources to describe the growing status of crop, plant height (PH), and vegetation indices (VIs) have been widely used to estimate AGB of crops [23,35,58]. The UAV-based light detection and ranging (LiDAR) system provides a way to estimate plant height accurately [59][60][61][62]. For example, Liu et al. [63] assessed the effectiveness of UAV-based LiDAR obtained point clouds to estimate the mean height of trees in the northern plains of Jiangsu Province, China, and showed a high prediction accuracy for Lorey's mean height with cross-validation coefficient of determination and relative root mean squared error of 0.97 and 2.83% by using partial least squares model. However, the high cost of UAV-based LiDAR system is the main limitation for its further application. To reduce the cost of estimating crop PH, researchers start to use low-cost UAV RGB remote-sensing system [64][65][66][67]. For example, Ziliani et al. [68] compared the PH derived from high-resolution UAV RGB imagery against PH derived from LiDAR, and showed a strong correlation between structure-from-motion derived heights and LiDAR scan data with the coefficient of determination (R 2 ) ranged from 0.77 to 0.99. Grüner et al. [69] estimated the PH of grass in northern Hesse, Germany from UAV RGB imagery by using photogrammetric structure from motion processing, and showed a high correlation between PH derived from UAV RGB imagery and ground-truth data with a R 2 value of 0.56 and a root mean squared error (RMSE) value of 0.13 m. In many of these studies, crop surface model derived from point clouds was adopted to estimate PH. However, relatively low estimations of PH were often obtained due to the fact that information below the crop canopy is also included in the crop surface model [70]. Compared to the crop surface model, choosing the appropriate PH feature directly from the point clouds to represent plant level could avoid estimation error caused by the information below the crop canopy. Therefore, in this study, point clouds derived from a consumer-grade UAV RGB imagery was directly adopted to estimate PH. More specifically, the 1st percentile of point clouds altitude was defined as the ground level, and six PH features, the 90th, 95th, 97th, 98th, and 99th percentiles, and the maximum value of point clouds altitude were used to search for an optimal PH feature. There was a highest correlation between PH directly derived from point clouds with the 99th percentile defined as the plant level and ground-truth PH (0.36-0.92 m) with an R 2 value of 0.90 and an RMSE value of 0.12 m. Compared to ground-truth PH, a relatively lower estimation was derived from the point clouds with the fitting line below the 1:1 line ( Figure 6). The reason for this relatively lower estimation may be that, during the study period, the highest position of maize appeared at the tip of the leaf (Figure 2b). Due to the limitation of accuracy of point clouds, the tip of the leaf may be difficult to be obtained.
Both linear and exponential regression analyses were adopted to estimate maize fresh and dry AGB based on PH 99 which had the best estimation performance of maize plant height. Compared to the linear regression model (both R 2 values were 0.59), the exponential regression models had higher correlations for both fresh and dry AGB with R 2 values of 0.77 and 0.76, respectively. However, with respect to RMSE, the exponential regression models had higher values of 0.49 and 0.07 kg/m 2 for maize fresh and dry AGB compared to the linear models with values of 0.41 and 0.07 kg/m 2 , respectively. Similar results were found for maize [71], barely [23,72]. For example, when LiDAR-derived metrics, mean height, and the coefficient of variation of normalized digital terrain model points' elevations, were used to estimate maize AGB in Heihe River Basin, Gansu province of China, the exponential regression models had higher R 2 values of 0.73 and 0.80 compared to R 2 values of 0.68 and 0.67 for the linear regression models. There were similar RMSE values for the linear and exponential regression models when the mean height was used; there was a higher RMSE value of 0.66 kg/m 2 for the exponential regression model compared to RMSE value of 0.60 kg/m 2 for the linear regression model when the coefficient of variation of normalized digital terrain model points' elevations was used. However, Tilly et al. [73] found an opposite result when PH derived from terrestrial laser scanning was used to estimate AGB of rice in Jiansanjiang, Heilongjiang Province, China. More specifically, for rice species of village 69 and village 36, there were lower R 2 values of 0.85 and 0.56 for the exponential regression model, compared to R 2 values of 0.90 and 0.60 for the linear regression model. The range of PH used to establish the AGB estimation model may be the reason for the opposite observations of estimation performance. The ranges of mean PH of each data collection campaign for this study, Bendig et al. [23], Tilly et al. [72], and Tilly et al. [73] were 0.36 to 0.92 m, 0.14 to 1.00 m, 0.17 to 0.81 m, and 0.25 to 0.57 m, respectively. When there was a relatively small range of PH, the linear regression model had better performance. However, when there was a relatively large range of PH, the exponential regression model had better performance. As shown in Figure 11, there was a rapidly increasing trend in slopes of the linear regression models between manually measured maize PH and biomass (both fresh and dry maize biomass), indicating that the exponential regression model was more suitable to describe the relationship between maize PH and biomass when all data (a relatively large range of PH) was adopted.
Remote Sens. 2019, 11, x FOR PEER REVIEW 14 of 21 relatively large range of PH, the exponential regression model had better performance. As shown in Figure 11, there was a rapidly increasing trend in slopes of the linear regression models between manually measured maize PH and biomass (both fresh and dry maize biomass), indicating that the exponential regression model was more suitable to describe the relationship between maize PH and biomass when all data (a relatively large range of PH) was adopted. When it comes to estimating crop AGB by using VIs derived from UAV remote-sensing system, there were two traditionally data sources, namely, multispectral [38,40] and hyperspectral [32,74]. For example, Meij et al. [75] estimated fresh AGB of oat based on simple difference vegetation index derived from UAV-based hyperspectral remote-sensing systems and showed an R 2 value of 0.56. Vega et al. [40] estimated AGB of sunflower by using the normalized difference vegetation index derived from UAV multispectral remote-sensing system and showed R 2 values ranged from 0.09 to 0.98. Recently, due to the high cost and low spatial resolution of UAV hyperspectral and multispectral remote-sensing systems, researchers started to estimate crop AGB based on vegetation index derived from UAV RGB remote-sensing system with a low cost and high spatial resolution [48,76], and showed a similar estimation performance. For example, Yue et al. [36] reported that there was a same R 2 value of 0.93 when random forest regression was adopted to estimate AGB of wheat in Changping District of Beijing, China based on vegetation indices derived from UAV hyperspectral and RGB remote-sensing systems, respectively. Therefore, in this study, five VIs and one VI combination, namely, normalized green-red difference index (NGRDI), excess green index (ExG), excess green minus excess red (ExGR), color index of vegetation (CIVE), vegetation index (VEG), and the combination of ExG, ExGR, CIVE, and VEG (COM), derived from consumer-grade UAV RGB imagery were adopted to estimate maize fresh and dry AGB by using simple and multivariable linear regression analyses. The results showed that there were good correlations when ExGR, NGRDI, and VEG were used to estimate maize AGB with R 2 values ranging from 0.63 to 0.73. It could be also observed that when ExGR, NGRDI, and VEG were used together in the multivariable linear regression model, a higher R 2 value of 0.82 was found for both of maize fresh and dry AGB. Similar results have been found for wheat [36,37]. For example, Lu et al. [37] reported that when VIs derived from UAV RGB remote-sensing system with R 2 values ranging from 0.04 to 0.62 were put into a random forest regression model, an ensemble learning method combining a large number of decision trees to improve the accuracy of classification and regression trees, a higher correlation with wheat AGB was obtained with an R 2 value of 0.70. The different recognition abilities of VIs for different When it comes to estimating crop AGB by using VIs derived from UAV remote-sensing system, there were two traditionally data sources, namely, multispectral [38,40] and hyperspectral [32,74]. For example, Meij et al. [75] estimated fresh AGB of oat based on simple difference vegetation index derived from UAV-based hyperspectral remote-sensing systems and showed an R 2 value of 0.56. Vega et al. [40] estimated AGB of sunflower by using the normalized difference vegetation index derived from UAV multispectral remote-sensing system and showed R 2 values ranged from 0.09 to 0.98. Recently, due to the high cost and low spatial resolution of UAV hyperspectral and multispectral remote-sensing systems, researchers started to estimate crop AGB based on vegetation index derived from UAV RGB remote-sensing system with a low cost and high spatial resolution [48,76], and showed a similar estimation performance. For example, Yue et al. [36] reported that there was a same R 2 value of 0.93 when random forest regression was adopted to estimate AGB of wheat in Changping District of Beijing, China based on vegetation indices derived from UAV hyperspectral and RGB remote-sensing systems, respectively. Therefore, in this study, five VIs and one VI combination, namely, normalized green-red difference index (NGRDI), excess green index (ExG), excess green minus excess red (ExGR), color index of vegetation (CIVE), vegetation index (VEG), and the combination of ExG, ExGR, CIVE, and VEG (COM), derived from consumer-grade UAV RGB imagery were adopted to estimate maize fresh and dry AGB by using simple and multivariable linear regression analyses. The results showed that there were good correlations when ExGR, NGRDI, and VEG were used to estimate maize AGB with R 2 values ranging from 0.63 to 0.73. It could be also observed that when ExGR, NGRDI, and VEG were used together in the multivariable linear regression model, a higher R 2 value of 0.82 was found for both of maize fresh and dry AGB. Similar results have been found for wheat [36,37]. For example, Lu et al. [37] reported that when VIs derived from UAV RGB remote-sensing system with R 2 values ranging from 0.04 to 0.62 were put into a random forest regression model, an ensemble learning method combining a large number of decision trees to improve the accuracy of classification and regression trees, a higher correlation with wheat AGB was obtained with an R 2 value of 0.70. The different recognition abilities of VIs for different aspects of crops may be the reason for the increase of estimation performance when multiple VIs were adopted. For example, NGRDI is sensitive to AGB before canopy closure [44] and VEG is invariant over the range of natural daylight illumination [50]. The addition of VEG into the AGB estimation model based on NGRDI may enhance the adaptability to different natural daylight illumination of this model, resulting in a better estimation performance. Further studies are still needed.
For the further improvement of estimation performance, researchers start to estimate crop AGB by using the crop VIs and PH simultaneously. However, in this study, the optimal PH observation of maize directly derived from UAV RGB point clouds was added into the AGB multivariable linear regression model based on ExGR, NGRDI, and VEG, resulting in a slight increase of 0.03 of the R 2 value for both fresh and dry AGB. For the RMSE, there was only a decrease of 0.01 kg/m 2 for fresh AGB. Similar results have been found for wheat [36,37] and barley [35]. For example, Lu et al. [37] reported an increase of 0.06 of R 2 value when wheat height was added into the random forest regression model based on UAV RGB VIs. Näsi reported a decrease of 0.01 of R 2 value when barley height information derived from UAV RGB imagery was added into the random forest regression model based on UAV RGB VIs. There may be two reasons for the slight improvement of performance when PH derived from UAV RGB imagery was added into the AGB estimation model based on VIs derived from UAV RGB imagery. On one hand, the reason may be the weaker ability of UAV RGB remote-sensing system to obtain an accurate ground level when estimating PH. Ota et al. [31] reported that the accuracy of ground level derived from the UAV RGB imagery was not accurate enough to estimate AGB of forest in the tropics, and that, for the accurate estimation of AGB, at least a single acquisition of airborne LiDAR was required to obtain ground level. On the other hand, the VIs have a significant relationship with PH. For example, Schirrmann et al. [48] reported a highly significant relationship between blue-green ratio index and winter wheat PH with an R 2 value of −0.93. As shown in Figure 12, in this study, high significant relationships were also found between VIs and maize PH with R 2 values of 0.77, 0.85, and 0.78 for NGRDI, ExGR, and VEG, respectively. Therefore, when estimating crop AGB based on UAV RGB remote-sensing system alone, exploring the optimized VIs and establishing estimation models with high performance based on advanced algorithms (e.g., machine learning technology) may be a better way [35,36]. aspects of crops may be the reason for the increase of estimation performance when multiple VIs were adopted. For example, NGRDI is sensitive to AGB before canopy closure [44] and VEG is invariant over the range of natural daylight illumination [50]. The addition of VEG into the AGB estimation model based on NGRDI may enhance the adaptability to different natural daylight illumination of this model, resulting in a better estimation performance. Further studies are still needed.
For the further improvement of estimation performance, researchers start to estimate crop AGB by using the crop VIs and PH simultaneously. However, in this study, the optimal PH observation of maize directly derived from UAV RGB point clouds was added into the AGB multivariable linear regression model based on ExGR, NGRDI, and VEG, resulting in a slight increase of 0.03 of the R 2 value for both fresh and dry AGB. For the RMSE, there was only a decrease of 0.01 kg/m 2 for fresh AGB. Similar results have been found for wheat [36,37] and barley [35]. For example, Lu et al. [37] reported an increase of 0.06 of R 2 value when wheat height was added into the random forest regression model based on UAV RGB VIs. Näsi reported a decrease of 0.01 of R 2 value when barley height information derived from UAV RGB imagery was added into the random forest regression model based on UAV RGB VIs. There may be two reasons for the slight improvement of performance when PH derived from UAV RGB imagery was added into the AGB estimation model based on VIs derived from UAV RGB imagery. On one hand, the reason may be the weaker ability of UAV RGB remote-sensing system to obtain an accurate ground level when estimating PH. Ota et al. [31] reported that the accuracy of ground level derived from the UAV RGB imagery was not accurate enough to estimate AGB of forest in the tropics, and that, for the accurate estimation of AGB, at least a single acquisition of airborne LiDAR was required to obtain ground level. On the other hand, the VIs have a significant relationship with PH. For example, Schirrmann et al. [48] reported a highly significant relationship between blue-green ratio index and winter wheat PH with an R 2 value of −0.93. As shown in Figure 12, in this study, high significant relationships were also found between VIs and maize PH with R 2 values of 0.77, 0.85, and 0.78 for NGRDI, ExGR, and VEG, respectively. Therefore, when estimating crop AGB based on UAV RGB remote-sensing system alone, exploring the optimized VIs and establishing estimation models with high performance based on advanced algorithms (e.g., machine learning technology) may be a better way [35,36].

Conclusions
Unmanned aerial vehicle (UAV) RGB remote-sensing system as one of the most important farmscale data acquisition tools has a great application potential in the rapid, accurate, and economical estimation of above-ground biomass at the farm scale. Our results confirmed that there was a high correlation between plant height directly derived from UAV RGB point clouds and ground-truth plant height with a coefficient of determination (R 2 ) value of 0.90 and an RMSE value of 0.12 m.

Conclusions
Unmanned aerial vehicle (UAV) RGB remote-sensing system as one of the most important farm-scale data acquisition tools has a great application potential in the rapid, accurate, and economical estimation of above-ground biomass at the farm scale. Our results confirmed that there was a high correlation between plant height directly derived from UAV RGB point clouds and ground-truth plant height with a coefficient of determination (R 2 ) value of 0.90 and an RMSE value of 0.12 m. Compared to the linear regression model (both R 2 values were 0.59), the exponential regression models based on plant height alone had higher correlations for both fresh and dry above-ground biomass with R 2 values of 0.77 and 0.76, respectively. The range of plant height used to establish the estimation model may affect the estimation performances of the linear and exponential regression models. The vegetation indices derived from UAV RGB imagery had great potential to estimate maize above-ground biomass with R 2 values ranging from 0.63 to 0.73. When multiple vegetation indices were adopted, a higher correlation was obtained with an R 2 value of 0.82. The different recognition abilities of vegetation indices to different aspects of the crop may be the reason for the increase of estimation performance. It was also confirmed that the addition of plant height derived from UAV RGB imagery into the above-ground biomass estimation model based on vegetation index could not significantly enhance the estimation performance. When estimating crop above-ground biomass based on UAV RGB remote-sensing system alone, exploring the optimized vegetation indices and establishing estimation models with high performance based on advanced algorithms (e.g., machine learning technology) may be a better way. color index of vegetation (CIVE); (e) vegetation index (VEG); and (f) the combination of ExG, ExGR, CIVE and VEG (COM). R 2 represents the coefficient of determination and RMSE represents the root mean square error.