Evaluation of Seed Emergence Uniformity of Mechanically Sown Wheat with UAV RGB Imagery

The uniformity of wheat seed emergence is an important characteristic used to evaluate cultivars, cultivation mode and field management. Currently, researchers typically investigated the uniformity of seed emergence by manual measurement, a time-consuming and laborious process. This study employed field RGB images from unmanned aerial vehicles (UAVs) to obtain information related to the uniformity of wheat seed emergence and missing seedlings. The calculation of the length of areas with missing seedlings in both drill and broadcast sowing can be achieved by using an area localization algorithm, which facilitated the comprehensive evaluation of uniformity of seed emergence. Through a comparison between UAV images and the results of manual surveys used to gather data on the uniformity of seed emergence, the root-mean-square error (RMSE) was 0.44 for broadcast sowing and 0.64 for drill sowing. The RMSEs of the numbers of missing seedling regions for broadcast and drill sowing were 1.39 and 3.99, respectively. The RMSEs of the lengths of the missing seedling regions were 12.39 cm for drill sowing and 0.20 cm2 for broadcast sowing. The UAV image-based method provided a new and greatly improved method for efficiently measuring the uniformity of wheat seed emergence. The proposed method could provide a guideline for the intelligent evaluation of the uniformity of wheat seed emergence.


Introduction
Wheat seedlings grown under non-uniform conditions will result in yield loss. Excessive planting density will reduce the availability of nutrients and intensify intraspecific competition between plants, which will further affect growth and development as well as reduce productivity [1]. However, while a sparse planting density can ensure full development of individual plants, poor soil fertility and a lack of luminous energy may make it difficult to achieve ideal output, and result in different levels of reduced production. Many factors affect the uniformity of wheat seed emergence. Some problems with seed emergence are also research hotspots, such as the effects of straw mulching [2,3], as well as the effects of temperature, moisture and sowing depth [4,5]. However, most current research on the uniformity of wheat seed emergence involves manually selecting several regions, investigating the amount of plants in those regions, and calculating the coefficients of variation of the plants in those regions to reflect the uniformity of seed emergence. Such method for measuring the uniformity of seed emergence are time-consuming and laborious as well as inadequate because it cannot directly reflect the seed emergence conditions in a field scale.
Currently, UAV remote sensing is playing an important role in crop monitoring and features high efficiency, high resolution, low cost, low risk, flexibility and other characteristics [6][7][8]. Previous research studies have found that the application of UAV remote sensing is mainly used for crop classification, disaster monitoring, wheat plants density, estimation of agronomic parameters and growth analysis in agricultural applications [9,10]. In crop classification research, UAVs mainly acquire data on different types of planting areas to facilitate crop management and related decision-making [11]. Disaster monitoring has been an important topic in recent years, including both the monitoring of field pests and the abnormal growth of crops caused by external factors, such as the monitoring of lodging [12,13]. The plants density of wheat was estimated using image processing algorithm and RGB imagery [14].
The estimation of the agronomic parameters mainly includes growth traits including leaf area index, biomass, nitrogen content and plant height through the establishment of models. Then, researchers looked for UAV image parameters that could reflect the quality of crop population to evaluate the status of crop growth [15,16]. Although many applications of UAVs are available related to crop growth and management, few reports were about the application of UAV remote sensing to the evaluation of the uniformity of wheat seed emergence.
This study used UAV images to evaluate the uniformity of wheat seed emergence under the conditions of mechanical broadcast and mechanical drill sowing, to explore the evaluation methods and implementation theories of the study of the uniformity of seed emergence under different sowing patterns. The goal was to provide technical support and a theoretical basis for the comprehensive evaluation of the uniformity of wheat seed emergence.

Field Trials
This study employed an experimental wheat variety, Yangfumai 4, with a seed emergence rate of 80% and a field emergence rate of 60% and used both mechanical drill and mechanical broadcast sowing patterns. Drill sowing machine put seeds in different rows, and the space between two lines was 25 cm. The normal distance between seeds was about 1 cm to 2.5 cm in one line. Broadcast sowing machine spread seeds on the field symmetrically, and the normal distance among seeds was about 6 cm to 9 cm. Three planting densities were selected: 180 × 10 4 ha −1 , 240 × 10 4 ha −1 , and 300 × 10 4 ha −1 (Figure 1). The size of each experiment plot is 2.7 m wide and 33 m long. There were 6 plots in the experiment: one replicate for each sowing method and density. Trials were made in both 2014 and 2015. The basic information of the experiments is shown in Table 1. Currently, UAV remote sensing is playing an important role in crop monitoring and features high efficiency, high resolution, low cost, low risk, flexibility and other characteristics [6][7][8]. Previous research studies have found that the application of UAV remote sensing is mainly used for crop classification, disaster monitoring, wheat plants density, estimation of agronomic parameters and growth analysis in agricultural applications [9,10] In crop classification research, UAVs mainly acquire data on different types of planting areas to facilitate crop management and related decisionmaking [11]. Disaster monitoring has been an important topic in recent years, including both the monitoring of field pests and the abnormal growth of crops caused by external factors, such as the monitoring of lodging [12,13]. The plants density of wheat was estimated using image processing algorithm and RGB imagery [14].
The estimation of the agronomic parameters mainly includes growth traits including leaf area index, biomass, nitrogen content and plant height through the establishment of models. Then, researchers looked for UAV image parameters that could reflect the quality of crop population to evaluate the status of crop growth [15,16]. Although many applications of UAVs are available related to crop growth and management, few reports were about the application of UAV remote sensing to the evaluation of the uniformity of wheat seed emergence.
This study used UAV images to evaluate the uniformity of wheat seed emergence under the conditions of mechanical broadcast and mechanical drill sowing, to explore the evaluation methods and implementation theories of the study of the uniformity of seed emergence under different sowing patterns. The goal was to provide technical support and a theoretical basis for the comprehensive evaluation of the uniformity of wheat seed emergence.

Field Trials
This study employed an experimental wheat variety, Yangfumai 4, with a seed emergence rate of 80% and a field emergence rate of 60% and used both mechanical drill and mechanical broadcast sowing patterns. Drill sowing machine put seeds in different rows, and the space between two lines was 25 cm. The normal distance between seeds was about 1 cm to 2.5 cm in one line. Broadcast sowing machine spread seeds on the field symmetrically, and the normal distance among seeds was about 6 cm to 9 cm. Three planting densities were selected: 180 × 10 4 ha −1 , 240 × 10 4 ha −1 , and 300 × 10 4 ha −1 (Figure 1). The size of each experiment plot is 2.7 m wide and 33 m long. There were 6 plots in the experiment: one replicate for each sowing method and density. Trials were made in both 2014 and 2015. The basic information of the experiments is shown in Table 1.

Image Acquisition
A DJI 1 Inspire RAW UAV was selected to obtain images. The camera calibration function of DJI Go APP (http://www.dji.com/) was used for camera calibration. It was equipped with a RGB camera with 16 megapixels, which is capable of acquiring images vertically with a flight height of 7 m (the spatial resolution is about 0.6 mm). The coordinates of the subsample over each microplot were manually defined using the orthomosaic image generated by the Agisoft Photoscan software (Version 1.2.2, Agisoft LLC., St. Petersburg, Russia). Matlab (V2016a, MathWorks, Natick, MA, USA) was used to process the images. The flowchart of UAV image preprocessing is shown in Figure 2. In this study, UVA data wree acquired on 17

Image Acquisition
A DJI 1 Inspire RAW UAV was selected to obtain images. The camera calibration function of DJI Go APP (http://www.dji.com/) was used for camera calibration. It was equipped with a RGB camera with 16 megapixels, which is capable of acquiring images vertically with a flight height of 7 m (the spatial resolution is about 0.6 mm). The coordinates of the subsample over each microplot were manually defined using the orthomosaic image generated by the Agisoft Photoscan software (Version 1.

Image Segmentation
Based on the row spacing of drill sowing, the entire image in each date was divided into 30 cm × 30 cm sub-regions (Figure 3), and then the percentage of wheat cover in each sub-region was calculated. Each replicate was divided in to 990 sub-regions. Coverage refers to the percentage of wheat seedling pixels in each sub-region. The image segmentation methods for drill and broadcast sowing were the same. The vegetation index (Equation (1)) and Otsu's algorithm were used to extract wheat seedlings in the image [17]. The reference threshold value obtained by Otsu's algorithm was

Image Segmentation
Based on the row spacing of drill sowing, the entire image in each date was divided into 30 cm × 30 cm sub-regions (Figure 3), and then the percentage of wheat cover in each sub-region was calculated. Each replicate was divided in to 990 sub-regions. Coverage refers to the percentage of wheat seedling pixels in each sub-region. The image segmentation methods for drill and broadcast sowing were the same. The vegetation index (Equation (1)) and Otsu's algorithm were used to extract wheat seedlings in the image [17]. The reference threshold value obtained by Otsu's algorithm was about 40 in both dates. The rows of drill sowing should be detected before getting the missing seedling region. The missing seedling region of broadcast sowing could be obtained without detecting rows. Thus, different analytical methods were designed in this study.
The values of the three color components (red, green, and blue) in RGB color images (24 bits) are, respectively, represented as R, G, and B.
Remote Sens. 2017, 9,1241 4 of 15 about 40 in both dates. The rows of drill sowing should be detected before getting the missing seedling region. The missing seedling region of broadcast sowing could be obtained without detecting rows. Thus, different analytical methods were designed in this study.
The values of the three color components (red, green, and blue) in RGB color images (24 bits) are, respectively, represented as R, G, and B.

Calculation of Seed Emergence Uniformity
The coefficient of variation (CVs) between sub-regions were calculated by Equations (2)-(4) and the coverage of each sub-region. "x", "i", and "n" in Equations (2) and (3) are the coverage of each sub-region, the serial number of each sub-region, and the total number of each sub-region, respectively. The coefficient of variation method is used to represent a uniformity index by dividing the standard deviation (δx) by the average value (Ex). A smaller coefficient of variation indicates higher uniformity [18]. Equations (2)-(4) provide the calculations [19]:

Localization of Seedlingless Ridges in Drill Sowing
Localization of seedlingless ridges refers to seeking the condition of missing seedlings with drill sowing of wheat in the row direction, identifying the regions with missing seedlings, and calculating the lengths of these regions. The localization steps were as follows. (1) Traverse the image based on a 3 × 3 pixels (about 2 × 2 cm) template, extract the coverage of the template by using Equation (1) and Otsu's algorithm, and construct a three-dimensional image of the area covered by the image. The three-dimensional image was constructed by coverage values (z-axis) and the coordinate (x-axis, y-

Calculation of Seed Emergence Uniformity
The coefficient of variation (CVs) between sub-regions were calculated by Equations (2)-(4) and the coverage of each sub-region. "x", "i", and "n" in Equations (2) and (3) are the coverage of each sub-region, the serial number of each sub-region, and the total number of each sub-region, respectively. The coefficient of variation method is used to represent a uniformity index by dividing the standard deviation (δ x ) by the average value (E x ). A smaller coefficient of variation indicates higher uniformity [18]. Equations (2)-(4) provide the calculations [19]:

Localization of Seedlingless Ridges in Drill Sowing
Localization of seedlingless ridges refers to seeking the condition of missing seedlings with drill sowing of wheat in the row direction, identifying the regions with missing seedlings, and calculating the lengths of these regions. The localization steps were as follows. (1) Traverse the image based on a 3 × 3 pixels (about 2 × 2 cm) template, extract the coverage of the template by using Equation (1) and Otsu's algorithm, and construct a three-dimensional image of the area covered by the image.
The three-dimensional image was constructed by coverage values (z-axis) and the coordinate (x-axis, y-axis) of the original image. In this type of image, the "hills" in a three-dimensional image represent the regions with wheat seedlings and the "cliff" areas in the row direction represent regions with missing seedlings. The regions of missing seedlings can be more intuitively seen through three-dimensional images. (2) Localize the wheat rows (the ridge in the three-dimensional image) using Equation (5). Each three-dimensional image is defined as I (m, k) , where m is the height of the image and k is the width of the image. (x i , y j ) is the coordinate of each pixel in the image, where Y j is the total value of each row's coverage in the image. Rg is the center line of the wheat row which is obtained by Equation (6).
(3) With row as the unit, calculate the change in horizontal coverage. (4) The portion where the mean value of the coverage change curve is equal to 0 is the region of missing seedlings. The length of the region of missing seedlings was calculated by adding all the length of 0 value points on the coverage change curve

Localization of Missing Seedling in Broadcast Sowing
The extraction steps for locating regions of missing wheat seedlings with broadcast sowing included: (1) confirming the coordinates of the original image (this step was used to confirm the length and width); (2) traversing the image based on a 2 cm × 2 cm template and extracting the coverage of wheat in the template by using Equation (1) and Otsu's algorithm; (3) extracting the regions whose value is 0 in the coverage distribution diagram and also with an area larger than 0.02 m 2 ; and (4) counting the number and measuring the area of the regions of missing seedlings as well as collecting any relevant information from between the regions of missing seedlings.

Uniformity Evaluation
The standard of evaluating uniformity was formulated by exploring the change of yield under different uniformity conditions. The field was divided into 2.5 m × 2.5 m regions. The grain yield and uniformity of each region was measured correspondingly. The grains were dried and weighed after manual harvesting, and then the grain yield of each region was obtained. The manual measurement steps of seed emergence uniformity are: (1) selecting 50 points randomly in each treatment; (2) marking each point by a 30 cm × 30 cm square and counting the seedling number manually; and (3) calculating the uniformity by Equations (2)-(4). The reference number of missing seedling region was counted manually. The length of missing seedling region for drill sowing field were measured by "Ruler" tool in the Agisoft Photoscan software. The area of missing seeding for broadcast sowing field were measured by "Polygon" tool in the Agisoft. The miss seedling rates of broadcast and drill sowing are calculated by Equation (7).
MS is the length of missing seedlings for drill sowing, and is the area of missing seedlings for broadcast sowing. TS is the total length of wheat seedlings for drill sowing, and is the total area of wheat seedlings for broadcast sowing. Figure 4 shows the three-dimensional effect of drill sowing on wheat coverage. In the threedimensional diagram, wheat was clearly divided into different "hills." The ridges can be easily positioned based on Equations (5) and (6). Then, through the extraction value of the hill line, the method can clearly delineate the position of the missing seedlings and seedlingless ridges regions. Figure 5 shows the ranges values for extraction of the first and fifth hill results in Figure 4. The result shows that the locations whose coverage values are 0 are all seedlingless ridge regions.   Then, through the extraction value of the hill line, the method can clearly delineate the position of the missing seedlings and seedlingless ridges regions. Figure 5 shows the ranges values for extraction of the first and fifth hill results in Figure 4. The result shows that the locations whose coverage values are 0 are all seedlingless ridge regions. Then, through the extraction value of the hill line, the method can clearly delineate the position of the missing seedlings and seedlingless ridges regions. Figure 5 shows the ranges values for extraction of the first and fifth hill results in Figure 4. The result shows that the locations whose coverage values are 0 are all seedlingless ridge regions.

Test of Wheat Seed Emergence Uniformity
The summary of the measured uniformity by UAV is shown in Table 2. The test results of UAV image-based uniformity delineation of wheat seed emergence is shown in Figure 7. From the 1:1 line, it can be seen that the uniformity measurement value of different sowing methods only slightly varied when compared with the manual measurement value. The R 2 values of drill and broadcast sown wheat were 0.83** and 0.89**, respectively. The RMSE values were 0.44 and 0.64, respectively. The maximum deviation of broadcast and drill sown wheat were 1.22 and 2.24, respectively.

Test of Wheat Seed Emergence Uniformity
The summary of the measured uniformity by UAV is shown in Table 2. The test results of UAV image-based uniformity delineation of wheat seed emergence is shown in Figure 7. From the 1:1 line, it can be seen that the uniformity measurement value of different sowing methods only slightly varied when compared with the manual measurement value. The R 2 values of drill and broadcast sown wheat were 0.83** and 0.89**, respectively. The RMSE values were 0.44 and 0.64, respectively. The maximum deviation of broadcast and drill sown wheat were 1.22 and 2.24, respectively.

Test of Wheat Seed Emergence Uniformity
The summary of the measured uniformity by UAV is shown in Table 2. The test results of UAV image-based uniformity delineation of wheat seed emergence is shown in Figure 7. From the 1:1 line, it can be seen that the uniformity measurement value of different sowing methods only slightly varied when compared with the manual measurement value. The R 2 values of drill and broadcast sown wheat were 0.83** and 0.89**, respectively. The RMSE values were 0.44 and 0.64, respectively. The maximum deviation of broadcast and drill sown wheat were 1.22 and 2.24, respectively.

Test of Drill Sowing Wheat Seedling-Missing Region
The summary of the measured information of missing seedlings is shown in Table 3. The 1:1 line between the UAV and manual measurement results of the number and length of drill sowing region of missing wheat seedlings is shown in Figure 8. The R 2 value of the number of this region was 0.84**, and RMSE was 3.99; the R 2 value of the length of this region was 0.61**, and RMSE was 12.39. The average numbers of regions of missing seedlings obtained from UAV images and manual measurement were 22.3 and 23.2, respectively. The average length of these regions obtained from UAV and manual measurement was 38.7 cm and 43.9 cm, respectively, which shows only a small difference.

Test of Drill Sowing Wheat Seedling-Missing Region
The summary of the measured information of missing seedlings is shown in Table 3. The 1:1 line between the UAV and manual measurement results of the number and length of drill sowing region of missing wheat seedlings is shown in Figure 8. The R 2 value of the number of this region was 0.84**, and RMSE was 3.99; the R 2 value of the length of this region was 0.61**, and RMSE was 12.39. The average numbers of regions of missing seedlings obtained from UAV images and manual measurement were 22.3 and 23.2, respectively. The average length of these regions obtained from UAV and manual measurement was 38.7 cm and 43.9 cm, respectively, which shows only a small difference.

Test of Broadcast Sowing Wheat Seedling-Missing Region
The summary of the measured information of missing seedlings is shown in Table 4. The 1:1 line between the UAV measurement results and manual measurement results of the number and area of broadcast sowing region of missing wheat seedlings is shown in Figure 9. The R 2 value of the number of region of missing seedlings was 0.93**, and RMSE was 1.39; the R 2 value of the area of region of missing seedlings was 0.78**, and RMSE was 0.20. The average number of these regions obtained from UAV images and manual measurement was 4.5 and 5.2, respectively. The average area of these regions obtained from UAV and manual measurement was 0.57 cm 2 and 0.51 cm 2 , respectively, which shows little difference.

Test of Broadcast Sowing Wheat Seedling-Missing Region
The summary of the measured information of missing seedlings is shown in Table 4. The 1:1 line between the UAV measurement results and manual measurement results of the number and area of broadcast sowing region of missing wheat seedlings is shown in Figure 9. The R 2 value of the number of region of missing seedlings was 0.93**, and RMSE was 1.39; the R 2 value of the area of region of missing seedlings was 0.78**, and RMSE was 0.20. The average number of these regions obtained from UAV images and manual measurement was 4.5 and 5.2, respectively. The average area of these regions obtained from UAV and manual measurement was 0.57 cm 2 and 0.51 cm 2 , respectively, which shows little difference.

Uniformity Evaluation under Different Plant Density
The present study found that, for both drill and broadcast sown wheat, changes in the uniformity of seedling emergence always had certain effects on yield (Table 5). At three different planting densities, when the measure of uniformity was larger than four, the population yield remained relatively high and stable. When the measure of uniformity fell between two and four, the yield fluctuated more widely, and, when uniformity was less than two, yield was relatively low. Table 5. Relationship between uniformity and yield (kg ha −1 ).
The length of a drill sowing region with missing seedlings had little effect on the yield in a certain range. However, when the missing seedling rate is more than 20%, yield will reduced significantly ( Table 6). When the missing seedling rate is less than 10%, populations create a relatively uniform distribution and this will usually result in high yield. Table 6. Relationship between the missing seedling rates for drill sowing and yield (kg ha −1 ).

Uniformity Evaluation under Different Plant Density
The present study found that, for both drill and broadcast sown wheat, changes in the uniformity of seedling emergence always had certain effects on yield (Table 5). At three different planting densities, when the measure of uniformity was larger than four, the population yield remained relatively high and stable. When the measure of uniformity fell between two and four, the yield fluctuated more widely, and, when uniformity was less than two, yield was relatively low. The length of a drill sowing region with missing seedlings had little effect on the yield in a certain range. However, when the missing seedling rate is more than 20%, yield will reduced significantly ( Table 6). When the missing seedling rate is less than 10%, populations create a relatively uniform distribution and this will usually result in high yield. Table 6. Relationship between the missing seedling rates for drill sowing and yield (kg ha −1 ).
The effects of the missing seedling rate with both broadcast and drill sowing were similar (Table 7). When the missing seedling rate was more than 10%, the yield was affected to different degrees. In contrast, when the missing seedling rate is less than 5%, the plants formed a relatively uniform population with high yield. Table 7. Relationship between the missing seedling rate for broadcast sowing and yield (kg ha −1 ).
The uniformity of wheat seed emergence and the missing seedling rate certainly affect yield, but the effects are difficult to describe using a mathematical expression. The present study formulated a grading scale based on the effects of the uniformity of seed emergence on the population and the missing seedling rate on yield ( Table 8). The uniformity of wheat seed emergence was classified as high (U1), moderate (U2) and low (U3). Table 8 refers to the different uniformity levels listed in Table 5. L1 (low), L2 (moderate) and L3 (high), giving the length of missing wheat seedlings with drill sowing (Table 6). A1 (low), L2 (moderate) and L3 (high) give the area of missing wheat seedlings with broadcast sowing ( Table 7). The grading system developed here (Table 8) was formulated according to the various effects of missing seedling conditions on yield (Tables 5-7). The population seed emergence of levels I, II and II range from most to least uniform, laying the best to worst foundation for obtaining high yield. Based on the grading system developed in the present study, when the population seed emergence reaches levels VII, VIII and IX, it usually indicates serious conditions caused by missing seedlings that will significantly affect yield. In these situations, the gaps need to be filled with seedlings. Table 8. Uniformity grading scale used in the present study.

Discussion
The use of UAV image-based data related to the uniformity of wheat seed emergence proposed in this study, to some extent, solved some deficiency problems in terms of manual measurement, but some problems remain and need to be discussed.

Acquisition Time
The UAV image acquisition time in the wheat growing period may affect the results. Wheat plants at the first and third leaf stage have relatively small leaves (Figure 10a) so that the UAV images do not provide a clear image of the plants. In addition, some wheat seeds may still have not germinated, so that the use of UAV images at that time provides only relatively partial data related to the uniformity of seedling coverage (Figure 10b). When wheat reaches the fourth and fifth leaf stage (Figure 10c), tillering begins to occur, and the region of missing seedlings is more obvious. Therefore, using the coverage of seedlings to evaluate the uniformity of seed emergence in a population is more appropriate (Figure 10d). By the time the wheat reaches the jointing stage, in some fields the rows will be fully covered by wheat leaves, and the regions with missing seedlings will be hidden by the leaves of other plants (Figure 10e). The jointing stage loses its usefulness in evaluating the uniformity of seed emergence (Figure 10f). In summary, the best time to use UAVs for evaluating the uniformity of wheat seed emergence should be around the fifth leaf stage.
to the uniformity of seedling coverage (Figure 10b). When wheat reaches the fourth and fifth leaf stage (Figure 10c), tillering begins to occur, and the region of missing seedlings is more obvious. Therefore, using the coverage of seedlings to evaluate the uniformity of seed emergence in a population is more appropriate (Figure 10d). By the time the wheat reaches the jointing stage, in some fields the rows will be fully covered by wheat leaves, and the regions with missing seedlings will be hidden by the leaves of other plants (Figure 10e). The jointing stage loses its usefulness in evaluating the uniformity of seed emergence (Figure 10f). In summary, the best time to use UAVs for evaluating the uniformity of wheat seed emergence should be around the fifth leaf stage.

Flight Attitude
A reasonable range of UAV flight height falls between 5 m and 15 m when acquiring images. In the present study, the relationship between UAV image acquisition altitude and actual area of each pixel was demonstrated through standard editing methods ( Figure 11). A low flight height makes it difficult to acquire stable images because the wheat seedlings may sway in the wind, resulting in increased labor to deal with image processing. At the other extreme, a high flight height will lead to the collection of obscure images.

Flight Attitude
A reasonable range of UAV flight height falls between 5 m and 15 m when acquiring images. In the present study, the relationship between UAV image acquisition altitude and actual area of each pixel was demonstrated through standard editing methods ( Figure 11). A low flight height makes it difficult to acquire stable images because the wheat seedlings may sway in the wind, resulting in increased labor to deal with image processing. At the other extreme, a high flight height will lead to the collection of obscure images.

Illumination Condition
Weather only slightly affects UAV image acquisition. After collecting the field data, images acquired in the morning or evening with relatively soft light or cloudy weather were available for review. However, particularly strong illumination is unsuitable for data collection because strong light may cause the wheat seedlings to appear white (Figure 13), making it impossible to completely extract the plants using Equation (1).  When the UAV flies at an altitude of 5 m, the image is relatively clear and some of the details of the wheat are visible. At 15 m, sporadic regions are fuzzy, but this has smaller impact on the height extraction ( Figure 12). When the flight height reaches 30 m, the image becomes blurred, and the uniformity extraction error is relatively large. Meanwhile, at 60 m, poor image quality makes it impossible to carry out uniformity evaluation.

Illumination Condition
Weather only slightly affects UAV image acquisition. After collecting the field data, images acquired in the morning or evening with relatively soft light or cloudy weather were available for review. However, particularly strong illumination is unsuitable for data collection because strong light may cause the wheat seedlings to appear white (Figure 13), making it impossible to completely extract the plants using Equation (1).

Illumination Condition
Weather only slightly affects UAV image acquisition. After collecting the field data, images acquired in the morning or evening with relatively soft light or cloudy weather were available for review. However, particularly strong illumination is unsuitable for data collection because strong light may cause the wheat seedlings to appear white (Figure 13), making it impossible to completely extract the plants using Equation (1).

The Component of Field
In terms of field image composition, the fields used in the present study were sprayed with herbicides before the experiment. Therefore, relatively few weeds grew in the field when acquiring the images, and the influence of weeds was ignored during the evaluation of uniformity. Fields with many weeds may be difficult to evaluate using the method proposed in this study. In terms of image processing, the great differences in wheat posture between the broadcast and drill sown patterns presented in the images make it impossible to design a universal method for the extraction of localized regions of missing wheat seedlings and related data. Thus, the methods designed here were unique to the present study.

The Effect of Emergence Uniformity
The uniformity of wheat seed emergence and the missing seedling rate had certain effects on yield, for the reason that plant density is a key factor influencing grain yield [14,20]. However, this effect proved difficult to quantify. Through studying the effects of uniformity and missing seedling rate on the yield, this paper designed an evaluation criterion (Table 8) for the wheat seed emergence rate, which was formulated based on the range of change of uniformity in the missing seedling rate. Only three stages were classified in this range, namely "good," "medium" and "bad." Discovering how to refine the evaluation criteria with the goal of further improving the precision of the evaluation will be the focus of future research studies.

Comparison to Others
Recent developments in UAV have provided research opportunities in assessing crop growth and physiological traits [21]. Imagery from UAV coming from different sensors [RGB cameras, multispectral imagery, hyperspectral imagery, and thermal camera] have been used to estimate LAI [22], biomass [23], height [24], nitrogen [25], and carotenoid [26]. Most of these studies estimated agronomic parameters by the spectral features of crop. However, some agronomic parameters could not be acquired only using spectral features, such as the information of missing seedling and emergence uniformity. In this case, the geometric and morphological analysis in the proposed method are available. Further, this work could be selected as a reference when the researchers want

The Component of Field
In terms of field image composition, the fields used in the present study were sprayed with herbicides before the experiment. Therefore, relatively few weeds grew in the field when acquiring the images, and the influence of weeds was ignored during the evaluation of uniformity. Fields with many weeds may be difficult to evaluate using the method proposed in this study. In terms of image processing, the great differences in wheat posture between the broadcast and drill sown patterns presented in the images make it impossible to design a universal method for the extraction of localized regions of missing wheat seedlings and related data. Thus, the methods designed here were unique to the present study.

The Effect of Emergence Uniformity
The uniformity of wheat seed emergence and the missing seedling rate had certain effects on yield, for the reason that plant density is a key factor influencing grain yield [14,20]. However, this effect proved difficult to quantify. Through studying the effects of uniformity and missing seedling rate on the yield, this paper designed an evaluation criterion (Table 8) for the wheat seed emergence rate, which was formulated based on the range of change of uniformity in the missing seedling rate. Only three stages were classified in this range, namely "good," "medium" and "bad." Discovering how to refine the evaluation criteria with the goal of further improving the precision of the evaluation will be the focus of future research studies.

Comparison to Others
Recent developments in UAV have provided research opportunities in assessing crop growth and physiological traits [21]. Imagery from UAV coming from different sensors [RGB cameras, multispectral imagery, hyperspectral imagery, and thermal camera] have been used to estimate LAI [22], biomass [23], height [24], nitrogen [25], and carotenoid [26]. Most of these studies estimated agronomic parameters by the spectral features of crop. However, some agronomic parameters could not be acquired only using spectral features, such as the information of missing seedling and emergence uniformity. In this case, the geometric and morphological analysis in the proposed method are available. Further, this work could be selected as a reference when the researchers want to use UAV imagery to acquire some agronomic parameters that are difficult to analyze only by spectral features.

Conclusions
This study established a comprehensive method to measure the uniformity of wheat seed emergence using UAV images. This method allows the measurement of the overall uniformity of wheat seed emergence (reciprocal value of the coefficient of variation), and determine the number of regions of missing seedlings with both drill sowing and broadcast sowing of wheat. Based on comparison with the conventional manual measurement method, the UAV image-based method of measuring the overall uniformity of seed emergence and the detection of regions of missing seedlings was achieved with high accuracy. Through a comparison between UAV images and the results of manual surveys used to gather data on the uniformity of seed emergence, the RMSEs was 0.44 for broadcast sowing and 0.64 for drill sowing. The RMSEs of the numbers of missing seedling regions for broadcast and drill sowing were 1.39 and 3.99, respectively. The RMSEs of the lengths of the missing seedling regions were 12.39 cm for drill sowing and 0.20 cm 2 for broadcast sowing. The UAV image-based method provided a new and greatly improved method for efficiently measuring the uniformity of wheat seed emergence.