Creating a Field-Wide Forage Canopy Model Using UAVs and Photogrammetry Processing

: Alfalfa canopy structure reveals useful information for managing this forage crop, but manual measurements are impractical at ﬁeld-scale. Photogrammetry processing with images from Unmanned Aerial Vehicles (UAVs) can create a ﬁeld-wide three-dimensional model of the crop canopy. The goal of this study was to determine the appropriate ﬂight parameters for the UAV that would enable reliable generation of canopy models at all stages of alfalfa growth. Flights were conducted over two separate ﬁelds on four different dates using three different ﬂight parameters. This provided a total of 24 ﬂights. The ﬂight parameters considered were the following: 30 m altitude with 90 ◦ camera gimbal angle, 50 m altitude with 90 ◦ camera gimbal angle, and 50 m altitude with 75 ◦ camera gimbal angle. A total of 32 three-dimensional canopy models were created using photogrammetry. Images from each of the 24 ﬂights were used to create 24 separate models and images from multiple ﬂights were combined to create an additional eight models. The models were analyzed based on Model Ground Sampling Distance (GSD), Model Root Mean Square Error (RMSE), and camera calibration difference. Of the 32 attempted models, 30 or 94% were judged acceptable. The models were then used to estimate alfalfa yield and the best yield estimates occurred with ﬂights at a 50 m altitude with a 75 ◦ camera gimbal angle; therefore, these ﬂight parameters are suggested for the most consistent results.


Introduction
Alfalfa was the third most valuable field crop in the United States in 2017 and 2018 [1,2]. However, unlike other top field crops, alfalfa is a forage primarily grown for its biomass. As a perennial, multiple harvests per year are produced over multiple years. Alfalfa's yield in any given harvest is determined by the height and density of the plant canopy [3]. Additionally, the nutritive value is a critical factor in the quality of the harvested biomass [4,5]. In alfalfa, there is a tradeoff between nutritive value and yield [6]. As the plant matures and grows taller, which increases yield, the concentration of less digestible components, such as lignin, increases and the nutritive value decreases [7,8]. The optimal relationship varies and is affected by location, enterprise management strategies, markets, anticipated end uses, weather, and climate factors unique to every farm and every harvest [9]. Furthermore, alfalfa stands weaken and thin over time and cannot be reseeded. Therefore, another unique decision in growing alfalfa is when to destroy the remaining stand and rotate into other crops. Alfalfa producers need to be aware of the state of the crop to select harvest times and crop rotation schedules that optimize the value to their enterprise [9]. All these considerations are required on top of monitoring for diseases, weeds, and other pest issues in order to deploy appropriate mitigation strategies that are just as important in alfalfa as in any field crop. The structure of the alfalfa canopy is directly related to many important production factors such as yield, quality, stand health, and pest control [10][11][12]. A three dimensional (3D) model of the physical structure of the alfalfa canopy can be used to altitudes. Alfalfa can be particularly challenging since once the canopy closes, the field appears as a consistent green surface that undulates in windy conditions. This type of surface and the lower flight altitudes limit the unique and consistent features that the photogrammetry model creation process requires in order to produce useful models. In these challenging conditions, the process can simply fail or produce unusable models with many odd artifacts and multiple ground surfaces (Figure 1).
Remote Sens. 2021, 13, x FOR PEER REVIEW 3 of 16 2.5 cm resolution [11] and achieving this level of detail in the canopy model calls for lower flight altitudes. Alfalfa can be particularly challenging since once the canopy closes, the field appears as a consistent green surface that undulates in windy conditions. This type of surface and the lower flight altitudes limit the unique and consistent features that the photogrammetry model creation process requires in order to produce useful models. In these challenging conditions, the process can simply fail or produce unusable models with many odd artifacts and multiple ground surfaces ( Figure 1). The challenges in creating alfalfa canopy models and the questions from others motivated the goal of this study, which was to develop and compare alfalfa canopy models derived from UAV imagery with different flight parameters (altitude and camera angle). Our initial motivation was to identify an efficient set of flight parameters that would reliably create successful models to enable the determination of canopy heights within the 2.5 cm resolution used in current manual field measurements [11]. However, other applications and uses of the 3D forage canopy models may call for different accuracy levels. One application of these models is to use them to estimate crop parameters such as yield. Therefore, to provide an illustration of the effects of these differences in 3D canopy creation, these canopies were used with previously developed models to estimate yield based on 3D canopy point clouds. In this study, our objective was to conduct an assessment on the ability to successfully create 3D canopy models using various flight parameters. Additionally, we wanted to ensure that the suggested methods work throughout the entire alfalfa regrowth process-from the initial sparse plant crowns to the fully closed canopies before harvest. The results of this investigation will help engineers and researchers utilizing 3D plant canopy models in the identification of the most useful and efficient methods to successfully create these models.

Experiment Fields
In this study, we used two alfalfa fields, Field 1 and Field 2 ( Figure 2), which are located at the University of Kentucky's North Farm with the approximate centroid of each field located at (38.128688, −84.509497) and (38.118587, −84.509612), respectively. The soil types in Field 1 were Armour silt loam, Egam silt loam, and Huntington silt loam. Field 2 consisted of Armour silt loam, Huntington silt loam, and Bluegrass-Maury silt loam. Data collection occurred from 17 May 2019 to 4 June 2019 with four separate data collection days. The first alfalfa cutting in the field for the 2019 growing season was cut on 7 May 2019 with a John Deere 630 Discbine mower (Moline, Illinois) and collected as haylage on 9 May 2019. This data collection period covered the alfalfa growth period between the first and second cuttings of the year. The second cutting occurred several days after data collection ended. Maturity levels ranged from 0 to 5 using the maturity stage evaluation scale from Kalu and Fick [33]. The alfalfa in Field 1 was a reduced lignin and glyphosate resistant variety, Ameristand 400 HVXRR, and the alfalfa in Field 2 was a glyphosate The challenges in creating alfalfa canopy models and the questions from others motivated the goal of this study, which was to develop and compare alfalfa canopy models derived from UAV imagery with different flight parameters (altitude and camera angle). Our initial motivation was to identify an efficient set of flight parameters that would reliably create successful models to enable the determination of canopy heights within the 2.5 cm resolution used in current manual field measurements [11]. However, other applications and uses of the 3D forage canopy models may call for different accuracy levels. One application of these models is to use them to estimate crop parameters such as yield. Therefore, to provide an illustration of the effects of these differences in 3D canopy creation, these canopies were used with previously developed models to estimate yield based on 3D canopy point clouds. In this study, our objective was to conduct an assessment on the ability to successfully create 3D canopy models using various flight parameters. Additionally, we wanted to ensure that the suggested methods work throughout the entire alfalfa regrowth process-from the initial sparse plant crowns to the fully closed canopies before harvest. The results of this investigation will help engineers and researchers utilizing 3D plant canopy models in the identification of the most useful and efficient methods to successfully create these models.

Experiment Fields
In this study, we used two alfalfa fields, Field 1 and Field 2 ( This data collection period covered the alfalfa growth period between the first and second cuttings of the year. The second cutting occurred several days after data collection ended. Maturity levels ranged from 0 to 5 using the maturity stage evaluation scale from Kalu and Fick [33]. The alfalfa in Field 1 was a reduced lignin and glyphosate resistant variety, Ameristand 400 HVXRR, and the alfalfa in Field 2 was a glyphosate resistant variety, Allied 428RR. One herbicide spray application was performed during the data collection period, Remote Sens. 2021, 13, 2487 4 of 15 on 21 May 2019. Weed, insect, and disease pressure were manually evaluated at 10 points (1 m 2 sampling quadrats) in each field on each of the four separate data collection days. After performing the drone flights on each day, the alfalfa in each sampling quadrat was hand harvested, weighed, and dried to determine the dry matter yield for each quadrat.
ually evaluated at 10 points (1 m sampling quadrats) in each field on each of the four separate data collection days. After performing the drone flights on each day, the alfalfa in each sampling quadrat was hand harvested, weighed, and dried to determine the dry matter yield for each quadrat.

Field Conditions
Weed pressure was low throughout the sampling period in both fields, with nearly all measured locations having less than 5% weed presence and only a minority in the 5% to 20% range. Insect and disease pressure were initially similar to weed pressure, with most locations possessing less than 5% of plants exhibiting damage from insects or disease. However, insect and disease pressure increased during the data collection period and by 4 June 2019, more locations had between 5% and 20% of plants showing damage from these sources.

Ground Control Points
On each data collection day, the first step was to record Ground Control Points (GCPs) and we also conducted our maturity, weed, insect, and disease evaluations throughout the field. The GCPs were the brightly painted corners of sampling quadrats (10 per field and 20 on each date), which were distributed in the field. These were rigid, located above the alfalfa canopy, and easily identifiable, which rendered them good GCPs. The GCPs were located using a Trimble 5800 receiver (Trimble Inc., Sunnyvale, CA, USA) and the Lefebure NTRIP Client (Cedar Rapids, Iowa), which provided RTK corrections from the Kentucky Continuous Operated Reference Stations (KYCORS) network.

Flights
After the GCPs were collected, three flight missions using different flight parameters were flown on each field using a DJI Phantom 4 Pro UAV (Shenzhen, China) for a total of six flights per day and 24 total unique flights. The Phantom 4 Pro has an integrated 20megapixel 2.54 cm CMOS visible camera mounted on a gimbal. The camera lens has an

Field Conditions
Weed pressure was low throughout the sampling period in both fields, with nearly all measured locations having less than 5% weed presence and only a minority in the 5% to 20% range. Insect and disease pressure were initially similar to weed pressure, with most locations possessing less than 5% of plants exhibiting damage from insects or disease. However, insect and disease pressure increased during the data collection period and by 4 June 2019, more locations had between 5% and 20% of plants showing damage from these sources.

Ground Control Points
On each data collection day, the first step was to record Ground Control Points (GCPs) and we also conducted our maturity, weed, insect, and disease evaluations throughout the field. The GCPs were the brightly painted corners of sampling quadrats (10 per field and 20 on each date), which were distributed in the field. These were rigid, located above the alfalfa canopy, and easily identifiable, which rendered them good GCPs. The GCPs were located using a Trimble 5800 receiver (Trimble Inc., Sunnyvale, CA, USA) and the Lefebure NTRIP Client (Cedar Rapids, Iowa), which provided RTK corrections from the Kentucky Continuous Operated Reference Stations (KYCORS) network.

Flights
After the GCPs were collected, three flight missions using different flight parameters were flown on each field using a DJI Phantom 4 Pro UAV (Shenzhen, China) for a total of six flights per day and 24 total unique flights. The Phantom 4 Pro has an integrated 20-megapixel 2.54 cm CMOS visible camera mounted on a gimbal. The camera lens has an 84 • field of view and an 8.8 mm focal length. The missions were developed using DJI Go and DJI GS Pro Mobile applications. All missions were set to have the images captured at equal intervals perpendicular to the main path without stopping. The images had a front Remote Sens. 2021, 13, 2487 5 of 15 overlap of 85% and a side overlap of 75%. The flight parameters varied based on elevation and gimbal angle (Table 1). Altering elevation and gimbal angle changed the resolution and the mission planning app automatically adjusted flight speed to ensure image quality (also shown in Table 1). In order to distinguish between these flight parameters (FP), they will hereafter be referred to by the elevation and gimbal angle as the 50-90 • , 50-75 • , and 30-90 • flight parameters. Missions providing complete field coverage took over 21 min for Field 1 (7.02 ha) and over 10 min for Field 2 (3.09 ha), even when using the highest speed 50-90 • flight parameters (Table 2). However, the GCPs (based on equipment placement for another experiment) were all located in a smaller region of each field. In order to reduce the flight time, images captured, and the resulting processing time, the mission areas were reduced from 7.02 ha to 1.3 ha for Field 1 and from 3.09 ha to 2.02 ha for Field 2 ( Figure 3). Some of the missions flown at 50 m still covered the entire field area to provide full field models, but most flights were of the smaller partial field region. Whether or not a particular flight (day/field/flight parameters) covered the entire field or part of the field is indicated in the results. For this analysis, flight routes and mission durations are dependent on both flight elevation and gimbal angle. The mission duration over the partial subsection of Field 1 at 50 m did increase by 45%, from 2 min 52 s to 4 min 9 s when the gimbal angle changed from 75 • to 90 • , even though the covered areas and flight elevation remained the same. Changes to the flight route enabled by the difference in gimbal angle provided the time reduction.  After flying the three missions that provided scans of the field, a fourth mission was conducted that flew around each of the sampling quadrats individually. The UAV was flown in a circle with a diameter of between 10 m and 20 m around the quadrat at a height of approximately 10 m. The UAV flew this circular route three times around the quadrat with the camera locked onto the quadrat while recording video (30 frames per second, 4K UHD: 3840 × 2160 pixels). This provided imagery of the alfalfa within the quadrat from many different angles at a close distance. During photogrammetry processing, this video was subsampled and every 30th frame was extracted. This provided between 40 and 80 images of each sampling quadrat. For this mission, photogrammetry processing was performed individually for each quadrat using the images extracted from the videos. Ground control points were not used since each quadrat was isolated and its geometry kept it at known distance from the ground. All other things being equal, the photogrammetry processing steps were the same.

Photogrammetry Processing
For each mission and its set of images, we created a new photogrammetry project in Pix4Dmapper (Pix4D S.A., Prilly, Switzerland). Pix4D was selected as it was a photogrammetry platform that has been used in many of the previously cited research projects [13,14,[25][26][27]29,31]. It is based on the computer vision techniques of [34,35] and implements various algorithms described in [36][37][38] to carry out the photogrammetry process. The GNSS coordinates, coordinate system, and camera model were automatically detected from the metadata associated with each image file. We operated in the World Geodetic System 1984 (EGM 96 Geoid) with a camera model code of FC6310_8.8_5472x3648(RGB). The images used for a single model were loaded and the processing template "3D Model", which set several program options, was chosen ( Table  3). The images' locations and layout were verified in map view before adding the GCPs. After flying the three missions that provided scans of the field, a fourth mission was conducted that flew around each of the sampling quadrats individually. The UAV was flown in a circle with a diameter of between 10 m and 20 m around the quadrat at a height of approximately 10 m. The UAV flew this circular route three times around the quadrat with the camera locked onto the quadrat while recording video (30 frames per second, 4K UHD: 3840 × 2160 pixels). This provided imagery of the alfalfa within the quadrat from many different angles at a close distance. During photogrammetry processing, this video was subsampled and every 30th frame was extracted. This provided between 40 and 80 images of each sampling quadrat. For this mission, photogrammetry processing was performed individually for each quadrat using the images extracted from the videos. Ground control points were not used since each quadrat was isolated and its geometry kept it at known distance from the ground. All other things being equal, the photogrammetry processing steps were the same.

Photogrammetry Processing
For each mission and its set of images, we created a new photogrammetry project in Pix4Dmapper (Pix4D S.A., Prilly, Switzerland). Pix4D was selected as it was a photogrammetry platform that has been used in many of the previously cited research projects [13,14,[25][26][27]29,31]. It is based on the computer vision techniques of [34,35] and implements various algorithms described in [36][37][38] to carry out the photogrammetry process. The GNSS coordinates, coordinate system, and camera model were automatically detected from the metadata associated with each image file. We operated in the World Geodetic System 1984 (EGM 96 Geoid) with a camera model code of FC6310_8.8_5472x3648(RGB). The images used for a single model were loaded and the processing template "3D Model", which set several program options, was chosen ( Table 3). The images' locations and layout were verified in map view before adding the GCPs. Utilizing the GCP/MTP Manager, the GCPs for the given field were imported and the Basic Editor tool was used to tag each GCP on 3-10 separate images. Once complete, the rough locations of all the GCPs were verified in the map view and compared with field notes to ensure the point had been marked correctly. Finally, we ran only the initial processing step. The initial processing step could be completed rapidly while the final two steps of model generation would take approximately one hour per 100 images. Verifying the model after the initial processing enabled quick identification and resolution of any errors detected at this stage. If failures were noted or the model did not pass the initial assessment detailed in the next section, the initial processing step was repeated. After processing each flight to create individual models, all the images from all three flights over a field on a single day were processed together to create one final model (the Combined model).

Model Assessment
The Root Mean Square Error (RMSE) of the GCPs was a local indicator of how well Pix4Dmapper fits the model to the GCPs. It represented the consistency between the geolocation measured in the field and the geolocation Pix4Dmapper estimates when processing the model. The goal was to minimize this value. In order to calculate RMSE, the error between the actual location of each GCP and its location in the model was divided into directional errors (X, Y, and Z components). All of the directional errors for all the GCPs in a model were combined through the root-mean-square calculation to provide a RMSE for each direction [39]. In order to provide a single value for a model, the three directional RMSE values were averaged to create the Model RMSE. A Model RMSE of greater than 20 cm after the initial processing step indicated that this step had failed and that the project should be recreated from the beginning. These failed models had generated a point cloud, but it was corrupt and contained strange artifacts (e.g. Figure 1). Model RMSE varied for the final 3D models but in all cases successful models had values much lower than 20 cm.
The ground sampling distance (GSD) was a measure of the spatial resolution of the 3D model and was defined as the distance between two neighboring pixels. A larger value indicates a model with less detail. Higher flights at higher altitudes generally have higher GSD values. The GSD is calculated separately for every pair of neighboring pixels in the model. In order to provide a GSD value for the entire model, all of the individual GSDs in the model were averaged together to produce the Model GSD. The final variable, which is the camera optimization, is the percentage difference between the initial and optimized parameters of the camera.
After processing each flight with GCPs and creating a final 3D canopy model, we checked the model output. In the case of a completely failed model, Pix4DMapper would provide an error stating that the process failed to produce any models. The model creation process is a stochastic matching process and so it is affected by initial conditions. It can potentially be trapped in the local minima of the optimization function. Simply repeating the processing steps provided new initial conditions and often corrected issues within this step. Once a model had completed processing without errors, the final Model RMSE, Model GSD, and camera calibration difference were recorded.

Yield Estimation Using the 3D Canopy Models
For all the photogrammetry models, the 3D canopy inside the quadrats was isolated. The point cloud from this canopy was processed following the procedure in [13] and the yield estimation models developed in [13] were applied to produce yield estimates in each quadrat. In [13], yield models were developed using 2-variables, which were the mean and standard deviation of the canopy heights, and 3-variables, which also added the average maturity level of the field on that day to the mean and standard deviation of the canopy heights that were used in the 2-variable model. Both of these models are Gaussian Random Process models with rational quadratic kernel functions. The 2-variable and 3-variable models were applied to each of the quadrat canopies. The estimates from these models were then compared with the actual yield obtained by hand harvesting the quadrat. For each set of flight parameters, the RMSE and R 2 value for the relationship between the estimated and actual yield was calculated.

3D Model Creation
The image sets from the 24 flights (two fields on four dates while using three different flight parameters) were processed individually and combined as described in the methods section. This resulted in 32 unique models of the alfalfa canopy, which are summarized in Table 4. Not all models completed the processing stage on the initial attempt. Three models were processed twice and one model was processed three times before successfully generating a 3D canopy as is noted in Table 4. In the end, Pix4D was able to generate 3D canopy models for each one by following the procedures outlined above and none of them possessed obvious errors such as multiple ground planes, strange artifacts, or discontinuities. Table 4. The properties of the 32 models that were processed in this study. Based on the extreme outliers in Model RMSE, camera calibration difference, and Model GSD, two of the 32 models were judged not acceptable. One model had an unusually large camera calibration difference of 85.4%, which was 38 times the median value of 2.24% and the other had an unusually large Model RMSE of 13.1 cm, which was seven times greater than the median value of 1.91 cm. These two models were considered failed attempts. They were removed from further processing and not considered in any further results. Note that these models were two of the four that had required additional processing attempts to even create a 3D model in the first place. While the other two models that had issues during initial processing did eventually result in acceptable models, the final outputs of these two attempts were not acceptable. Thus, of the 32 models that were attempted, 30 of them produced acceptable models.

Flight Parameters (Height (m)-Gimbal
The values in Table 4 precisely describe the detail of the models, but images of the models are also useful in understanding the model detail. Examples of a section of the model created using the images captured with the 30-90 • flight parameters on 4 June 2019 in Field 1 are shown in Figures 4-6. The images show that the models captured gaps in canopy coverage and details such as the wheeled traffic effects from a spray application two weeks prior.        Although the UAV did not take pictures from this angle and this close to the ground, the photogrammetry process captured the transition from the short grass bordering the field to the taller alfalfa plants.  The results from the individual models were compiled by flight parameters used in the flight to capture the model data ( Table 5). The largest effect of the flight parameters was on the GSD. The flights at an altitude of 30 m had a much lower mean Model GSD compared to the flights at 50 m, as would be expected with models based on images taken at different altitudes. This corresponds to finer detail in the models from the flights at 30 m. There was also little variation (low standard deviation) in the GSD between models created using a given flight parameter. For all models created using the 30-90 • flight parameters, the GSD was always less than 1 cm. For models created using 50-75 • and 50-90 • flight parameters, the GSD was always below 1.5 cm. Table 5. Model GSD by flight parameters used in flights (including both fields on all four sampling days).

Min.
Max There were fewer significant differences in the Model RMSE values (Table 6). While the models created using the images from the flights at 30 m did have lower mean Model RMSE, there were no significant differences (using α = 0.05) between the Model RMSE values based on flight parameters. Maximum RMSE values (from acceptable models) were as high as 5.7 cm, but such high values were rare. Of the 30 acceptable models, 28 models had Model RMSE values less than 2.5 cm. Interestingly, the combination models performed worse than models based on images taken only at 30 m. The additional images from the two flights using the 50 m flight parameters did not improve the model creation process. They only appeared to increase additional processing time.

Yield Estimation
There were 10 sampling quadrats per field on each of the four sampling days. This corresponds to a total of 80 samples for each flight parameter. However, not all samples created 3D canopy point clouds that could be used in the yield estimation model. The most common issue was that the highly visible geometric structure of the sampling quadrat would create canopy distortions inside the sampling region and the point cloud processing steps used to estimate yield would fail. Of the 80 possible quadrat samples, there were 75, 34, 34, 28, and 72 usable samples for the yield estimation model from the Quadrat Only, 30-90 • , 50-90 • , 50-75 • , and the Combination model, respectively. The Quadrat Only and Combination models had more views of the quadrats from different angles, which appeared to assist those models by reducing distortions from the quadrats. In actual field usage, the sampling quadrats should not be an issue since they are not normally present during alfalfa production. Future work with complete field imagery at 30 m and 50 m elevations will need to use different sampling quadrats to eliminate the introduction of artificial structures to the field canopy.
When the 3D canopy models were combined with the yield estimation models presented in [13], the yield estimation accuracy mostly aligned (Table 7) with the previously discussed canopy model accuracy represented by the Model RMS Error. The 50-75 • and the 30-90 • FP did better than the 50-90 • and the combined models. However, for yield estimation, the 50-75 • FP had the lowest error and the highest R 2 values when used to predict the yield within the test quadrats. The next best set of flight parameters that covered the entire field was the 30-90 • FP. The combined FP created the worst yield estimates. This set of data also included the Quadrat Only FP which consisted of individual canopy models created for each yield sampling quadrat at a height of approximately 10 meters with varying non-nadir gimbal angles. Although the Quadrat Only FP had the lowest flight altitude and had independent 3D canopies for each quadrat, it was only about as accurate as the 30-90 • FP. The error rates and R 2 values are different when the canopies are used in the 2-variable and 3-variable models, but the trends remain similar with both approaches for estimating yield: 50-75 • FP produced the best estimates and the Combined FP produced the worst. Table 7. Yield estimation accuracy using the 3D canopy models from different flight parameters.

Flight Parameters (Height (m)-Gimbal
Angle) The 2-variable and 3-variable models were created using individual flights and 3D models for each sampling quadrat in a manner that is most similar to the Quadrat Only FP presented in Table 7. In [13], the 2-variable and the 3-varible model were found to predict yields with RSME/R 2 of 799 kg ha −1 /0.63 and 734 kg ha −1 /0.69, respectively. These are similar to the values obtained with the Quadrat Only FP parameters. This set of experiments was conducted in different fields in a different growing year and so both yield estimation models are consistent even with the changes in varieties, fields, and weather conditions. As in [13], the 3-variable model performed better than the 2-variable model with lower RMS Error and higher R 2 values when considering the average of all flight methods. However, for the 30-90 • and 50-90 • flight parameters with the 2-variable model, the R 2 values fell below 0.4 and so the 2-variable model does not work well with those flight parameters. Interestingly, using the 50-75 • FP produced better results than the high-resolution models from the Quadrat Only canopies. The RMSE and R 2 values for yield estimates with the 50-75 • FP were even better than those obtained when initially creating the models.

Discussion
While 94% of the 32 attempted models were created successfully, the two failed models demonstrate that these flight parameters are not foolproof. If a model is being used for mission critical decision making and a usable model is crucial, multiple flights using different flight parameters may be warranted. In cases where a model failed in this study, the models based on a different set of flight parameters over the same field at the same time were always able to create a successful model. The canopies captured during this project include entire whole alfalfa forage growth cycle. They start at seven days after cutting when regrowth was just beginning. The final canopy was measured 28 days after cutting when the crop was, again, ready for harvest. Therefore, these parameters appear acceptable for utilization in modeling the canopy during the entire growth cycle of the alfalfa forage.
The models produced with the 30-90 • FP were more detailed as determined by the GSD, but they were not more accurate based on the lack of significant differences in Model RMSE between the different flight parameters. These models were also less accurate when used to predict yield. We also found little benefit from combining images from flights with multiple parameters. While the combined models did appear to have a denser point cloud than many of the other models, the additional processing time, minimal difference in model accuracy parameters, and the lower ability to predict alfalfa yield when used with the alfalfa yield prediction models suggests that this approach is actually detrimental to the usefulness of a model.
We suggest using the 50-75 • parameters for the creation of alfalfa canopies with the 3-variable model for predicting yield. Lower altitudes required much longer flight and processing times and the models were not always acceptable after processing. In comparing the gimbal angles at 50 m, the 75 • gimbal angle provided more consistent Model RMSEs that were always less than 2.2 cm, unlike the 90 • gimbal angle which had a maximum of 5.7 cm. The tests, when using the different flight parameters to predict alfalfa yield, further strengthen this suggestion. The 50-75 • parameters were clearly the best for this particular use. The 3-variable model was more consistent across flight parameters, which could be useful for real world applications where field elevation changes may make it difficult to maintain an exact flight height above the ground. While the 2-variable model with the 50-75 • parameters had slightly lower RMSE than the 3-variable model, its R 2 with all flight parameters was always lower than the 3-variable model. In certain situations, the R 2 could be as low as 0.30 (30-90 • ) or 0.35 (50-90 • ). In our previous study, the RMSE followed the trend of the R 2 values with the 3-variable model providing better results than the 2-variable model. For the sake of consistency, the 3-variable model appears to be the better option. Further testing would be necessary to see if flights at slightly higher altitudes may also provide acceptable accuracy levels, but based on the results of other studies we should see deteriorating accuracies and detail as we increase in height [27,28,30,31].

Comparisons with Other Studies
While other studies aimed at estimating biomass or crop yield did not report the effect of varying flight heights, they flew missions at similar heights between 30 m and 50 m above ground level. However, these studies all appeared to use cameras with a 90 • gimbal angle. The results of our study indicate that switching to a 70 • gimbal angle has the potential to improve crop height models and, thus, their biomass or yield estimation results.
A major difference between our study and the other photogrammetry studies is our focus on the ability to reliably create an acceptable model and consistently estimate alfalfa yield correctly. Our study performed multiple repetitions of flights using each flight parameter on different days and on multiple fields. In commercial alfalfa production, farmers must periodically assess the crop to determine appropriate harvest time as well as other production decisions such as when to spray various chemicals. The crop is constantly developing and timely information is critical in the production decision making process. In this project, we followed this periodic monitoring requirement and flew every seven days exactly. Other studies had different goals and many based their analysis on a single set of flights. Since they were not flying on a set schedule, they could have waited for perfect flight capture conditions (e.g. no wind) and repeated any flights that provided unacceptable results after photogrammetry processing. It is also possible that the fields and environments targeted in these studies have enough natural variation that model failure is not an issue. Given the inquiries into our methods and our own difficulty in initially identifying appropriate flight parameters, consistency and reliability are important additional factors, which this study has addressed.

Conclusions
Based on this study, we suggest the utilization of a set of flight parameters for creating 3D models of alfalfa canopies. Following the procedures provided here, producers and researchers can be confident that they will consistently be able to generate a 3D canopy model with a single coverage flight based on our acceptable model creation rate of 94%. Furthermore, this remains true for all of the alfalfa growth stages from initial regrowth from sparse crowns to a fully closed canopy. In the cases where one of our canopy models was not acceptable, our other trials on that day and in that field did provide acceptable models. If a successful 3D canopy model is absolutely required, it is suggested that a second flight with different image capture points be performed. Model detail (measured by GSD) varied by flight altitude and gimbal angle between the means of 0.767 cm for the 30-90 • flight parameters and 1.41 cm with the 50-75 • flight parameters. However, the geolocation error (measured by Model RMSE) was similar for the different flight parameters with an overall mean of 1.91 cm. When these canopies were used with alfalfa prediction models, the 50-75 • flight parameter clearly provided the best estimate of alfalfa yield. Given the lower flight durations and processing times as well as increased consistency in geolocation accuracy, most users should opt for flights with a 50 m altitude and 75 • gimbal angle. If the target application for the 3D canopy is alfalfa yield estimation, then the 50-75 • flight parameter is strongly suggested.  Data Availability Statement: Upon acceptance of this manuscript, we will work with the editor to determine the appropriate public posting location for the data from this project. The raw images are approximately 50 GB and the photogrammetry models are also very large, which limits posting locations.

Conflicts of Interest:
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript or in the decision to publish the results.