ASABE Logo

Article Request Page ASABE Journal Article

Evaluation of Dicamba Drift Injury and  Yield Loss on Soybean Using Small  Unmanned Aircraft Systems (sUAS)  and Multispectral Imaging Technologies

Kelvin Betitame1, Cengiz Koparan1, Yu Zhang1, Kirk Howatt1, Michael Ostlie2, Sreekala G. Bajwa3, Xin Sun1,*

Published in Journal of Natural Resources and Agricultural Ecosystems 2(2): 63-76 (doi: 10.13031/jnrae.15686). Copyright 2024 American Society of Agricultural and Biological Engineers.

1    Department of Agriculture and Biosystems Engineering, North Dakota State University, Fargo, North Dakota, USA.

2    Carrington Research Extension Center, North Dakota State University, Carrington, North Dakota, USA.

3    College of Agriculture and Montana Agricultural Experiment Station, Montana State University, Bozeman, Montana, USA.

*    Correspondence: xin.sun@ndsu.edu

The authors have paid for open access for this article. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License https://creative?commons.org/licenses/by-nc-nd/4.0/

Submitted for review on 26 May 2023 as manuscript number NRES 15686; approved for publication as a Research Article by Associate Editor Dr. Kelly Thorp and Community Editor Dr. Kati Migliaccio of the Natural Resources & Environmental Systems Community of ASABE on 7 November 2023.

Highlights

Abstract. Many dicotyledonous weeds can be selectively controlled by dicamba application during the growing season. However, vapor or spray drift from dicamba application could cause significant yield loss on non-target species such as soybean [Glycine max (L.) Merrill] that do not contain the dicamba resistance trait. In this study, the impact of simulated dicamba drift was investigated on dicamba-susceptible soybeans during different growth stages. A field experiment was conducted employing small unmanned aircraft systems (sUAS) and multispectral image sensors to assess the dose-response relationship between dicamba drift and soybean yield. The experiment followed a randomized complete block design, with main plots representing three distinct soybean growth stages (third trifoliate, sixth trifoliate, and early reproductive stage) and sub-plots encompassing various simulated dicamba drift rates. The analysis, involving non-linear regression and normalized difference vegetation index (NDVI) regression, revealed that visible soybean injuries were observed at minimal dicamba rates. However, significant yield loss occurred during the V6 growth stage at a dicamba rate of 98 g.ai ha-1, emphasizing the critical growth stage sensitivity. Moreover, a moderate yet meaningful relationship (R2 = 0.46) was established between sUAS multispectral NDVI data and soybean yield, highlighting the potential for sUAS with multispectral sensors to predict soybean yield losses. These results contribute valuable insights into the impact of dicamba drift on soybeans and the efficacy of remote sensing technology in assessing yield outcomes.

Keywords.Dicamba-drift injury, Multispectral image, Precision agriculture, Soybean yield, sUAS, Yield prediction.

The widespread use of crops that have been genetically engineered with resistance traits to herbicide has made weed control much easier for farmers. These crops with resistance to a particular herbicide give farmers the leverage to apply herbicide over the top, giving farmers flexibility to manage the time herbicides are applied (Al-Khatib et al., 2013). However, one notable problem is the development of “superweeds,” weeds with resistance to herbicides, due to the repeated use of the same herbicide over the years on weeds (Al-Khatib et al., 2013). This high selection pressure of weeds, especially glyphosate, led to the introduction of dicamba (3, 6-dichloro-2-methoxybenzoic acid)-resistant crops from chemical and seed companies. But dicamba, too, has its problems; it does not stay at the target field of application, it is prone to volatilization and drift, and, therefore, has the potential to cause injury to susceptible crops in a broad area around the application location (Wechsler et al., 2019; Werle et al., 2018).

Dicamba is an auxinic herbicide that belongs to the benzoic acid class. They are selective herbicides primarily used to control broadleaf weeds during the post-emergence period (Huang et al., 2016; Jacob et al., 2015). Dicamba is a safe herbicide for humans if it is used in accordance with the label, and there are no research findings linking human exposure to dicamba with health issues such as cancer (Thurman et al., 2017); it is, however, slightly toxic when inhaled, ingested, or exposed to the skin (Bodnar, 2018). While dicamba herbicide has been available in the U.S. since 1967, its primary usage historically involved selective control of broadleaf weeds in grass crops during late winter or early spring, typically preceding the planting of broadleaf crops (Sterling and Hall, 1997). The recent evolution of weeds with resistance to the most widely used herbicide (glyphosate) necessitated an alternative weed management system to combat the prevailing problem. Compared to other herbicide modes of action, dicamba has slow weed resistance development (Kelley et al., 2005). Since weed resistance to dicamba has been slow and few weeds express opposition, dicamba provides an alternative option for controlling weeds that are resistant to glyphosate (Jacob et al., 2015).

Consequently, biotech companies developed genetically modified dicamba-resistant soybeans and cotton, enabling farmers to manage weeds in post-crop emergence (Olszyk et al., 2015). While dicamba-resistant seeds offer benefits such as in-crop weed control, their volatility poses challenges for producers, particularly due to the high susceptibility of non-resistant broadleaf crops. The exposure of micro-rates of dicamba to susceptible soybean crops can cause various injury symptoms, such as leaf cupping, stunted growth, necrosis, and curling of pods, all of which could eventually lead to yield loss (Egan and Mortensen, 2012; Hartzler, 2017; Osipitan et al., 2019; Robinson et al., 2013; Soltani et al., 2016). The micro-rate of dicamba as low as 0.08 g.ae ha-1 is prone to volatilization and drift, which is capable of causing injuries to susceptible soybean (Egan et al., 2014; Robinson et al., 2013; Weidenhamer et al., 1989).

Following the commercialization of glyphosate-resistant crops, by 2006, a significant portion of soybean acreage had adopted glyphosate-resistant varieties, which are susceptible to dicamba. In fact, approximately 9 out of 10 soybean acres were consistently planted with glyphosate-resistant seeds (Osipitan et al., 2019; USDA, 2019). Therefore, the off-target movement of dicamba became a significant problem for most soybean farmers between the 2015 and 2016 growing seasons, when farmers started using dicamba for weed control. Dowell (2017) reported that Monsanto (now owned by Bayer) started selling genetically modified soybean seeds with resistance to dicamba called Xtendflex in 2015-2016 without the approval of the Environmental Protection Agency (EPA) for its corresponding herbicide, Xtendimax, for weed control (Dowell, 2017). Therefore, farmers who purchased Xtendflex used previous dicamba formulations for weed control. This situation resulted in a massive drift-induced herbicide injury in soybean and other susceptible broadleaf crops across the country (Thurman et al., 2017). Although dicamba-resistant seeds and herbicides were later approved in the 2017 growing season, dicamba drift still persisted (Osipitan et al., 2019). To successfully quantify soybean injuries that result in crop losses from dicamba drift, a reliable system is needed for effective and rapid data collection.

One of the technologies mostly used for crop scouting is remote sensing. Remotely sensed images of crop fields were collected using satellites or piloted aircraft (Matese et al., 2015). These platforms, especially satellites, can provide a synoptic view of a large field at a very low cost (Stroppiana et al., 2018). Even though remotely sensed data from satellites and aircraft are rapidly collected (Duddu et al., 2019), they have inherent limitations. Open-source remote sensing data from satellites have low resolution; meanwhile, high-resolution data are also expensive (Honrado et al., 2017). Although high-resolution data can be obtained from satellites, it can also be marred by cloud cover since they are collected above clouds (Matese et al., 2015). Moreover, satellite data collection is constrained to a fixed period. At the same time, herbicide drift can occur at any time within the crop’s phenological stage, making it unsuitable for drift detection in crops. These factors necessitated the exploration of other remote sensing platforms that will offer new opportunities in addressing the problems associated with satellite images. Hence, sUAS presents an improved choice for remote sensing data collection, particularly suited for smaller research-focused field sizes.

Remote sensing by sUAS enables the monitoring of plant stress through on-demand RGB and multispectral image data. This offers high spatial and temporal resolutions, facilitating the distinction between healthy injured crop features (Duddu et al., 2019). The multispectral sensors mounted on drones can capture high-resolution images with narrow wavelength bands that can detect subtle changes in plant health before they become visible to human eyes (AgEagle, 2018). Additionally, the multispectral sensor's ability to capture numerous bands, exemplified by the Micasense RedEdge-MX dual with 10-band multispectral capability, renders it suitable for image comparisons with satellite imagery (Laliberte et al., 2011). The traditional method of detecting and assessing crop injury from dicamba drift involves visual ratings and manual assessment through field sampling. Even though visual ratings of the injury levels of the plants might be practical, they can be marred with inaccuracy, resulting in an under- or over-rating of the extent of dicamba injury. Moreover, visual rating is labor-intensive and time-consuming, requiring a lot of time to assess large fields, leading to visual fatigue and inconsistencies in rating (Duddu et al., 2019).

Studies have shown that off-target movement of herbicides that results in height and crop yield reduction could be properly monitored using remotely sensed data instead of manual assessment (Henry et al., 2004). Therefore, soybean researchers and producers are interested in a cost-effective and precise method of identifying and quantifying dicamba injury in soybean. Oseland et al. (2021) used vegetative indices, which included a red-edge wavelength (717 nm), to determine yield loss in soybean from 2,4-D and dicamba at sublethal rates. Huang et al. (2015) assessed soybean injuries from glyphosate drift using vegetative indices from multispectral remote sensing. Zhao et al. (2014) used the Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), Soil Adjusted Vegetation Index (SAVI), and Difference Vegetation Index (DVI) computed from hyperspectral remote sensing for early detection of glyphosate injury in soybean and cotton. Al-Gaadi et al. (2016) predicted potato crop yield using vegetation indices from Landsat and Sentinel images.

While the aforementioned studies demonstrated the utility of remote sensing in assessing crop injuries caused by herbicides, additional research is required to ascertain the precise threshold of herbicide (specifically dicamba) drift that leads to yield reduction. This is crucial because the degree of yield loss in dicamba-susceptible soybeans is determined not only by the amount of dicamba absorbed by the plants but can also be related to the specific growth stage during which these susceptible crops are exposed to the drift. Furthermore, it is necessary to investigate whether vegetation indices can effectively characterize crop injuries that result in yield reduction. As such, the objectives of this research were to determine the dose-response of dicamba-susceptible soybeans to simulated dicamba drift at vegetative and reproductive growth stages and to assess the relationship between vegetation indices and the yield of dicamba-susceptible soybeans.

Materials and Methods

Data Collection

A multirotor UAV (Matrice 600, DJI, Shenzhen-Guangdong, China) with six propellers was used for collecting aerial images of the soybean fields at the experiment sites. The UAV was equipped with a RedEdge-MX multispectral camera (MicaSense, Inc., Seattle, WA) that had the capability of capturing five spectral bands within the range of 475-840 nm of the electromagnetic spectrum at each camera triggering (table 1). Each sensor had a digital image size of 1280 x 960 pixels with a 47° horizontal field of view, a 37° vertical field of view, and a 58° diagonal field of view operating within the temperature range of 0°C to 60°C capable of capturing an image per second. The MicaSense RedEdge camera had an embedded daylight scattering sensor integrated with GPS to calibrate the images with a reflectance panel for calibration. The camera had a ground sampling distance of 8 cm per pixel at 120 m above ground level. Image collection with sUAS was performed between 11:00 and 15:00, with a flight duration ranging from 15 to 20 minutes. Aerial image collection and visual ratings of the plants were conducted concurrently on the same day at both locations. The aerial and visual data were completed 10 days after treatment (DAT), 20 DAT, 40 DAT, and 45 DAT. This was a way of checking inter-plot drift occurrence at each stage after herbicide application.

The average temperature and wind speed during data collection varied from 20? to 30? and 1.3-3 m/s, respectively, in both locations. The sUAS mission plan altitude was assigned as 61 m for autonomous navigation and image capturing, which took about 20 minutes to complete the mission. Operating at this altitude, the sUAS could cover a large area with minimum time, ultimately saving battery life. The images were captured at a resolution of 1280 x 960 at 1.2 megapixels and covered a 0.269-ha area on the ground, with 4 cm per pixel. To achieve high-quality mosaics, flight speed, and path were synchronized with the camera parameters, including exposure time, aperture, sensitivity, and frame rate. A 75% frontal and side overlap was set for all the flights. For radiometric calibration purposes, images of the RedEdge MX calibration panel were taken at about 1 m height before each flight during field data collection.

Table 1. The specifications of the sUAS and the multispectral camera.
EquipmentSpecifications
  1. sUAS - DJI Matrice 600 Pro Multirotor
    1. Dimensions: 525 mm × 480 mm × 640 mm
    2. Max Takeoff Weight: 15.5 kg
    3. Max Ascent Speed: 5 m/s
    4. Max Decent Speed: 3 m/s
    5. Max Wind Resistance: 8 m/s
    6. Max Speed: 40 mph/ 65 kph (no wind resistance)
    7. Flight Control System: A3 Pro
  1. Camera - Micasense RedEdge MX
    1. Dimension: 8.7 cm × 5.9 cm × 4.54 cm
    2. Spectral Bands: Blue, Green, Red, Red Edge, Near Infrared
    3. Ground Sample Distance: 4 cm per pixel (per band) at 200 ft AGL
    4. Capture Rate: 1 capture per second (all bands)
    5. Output bit Depth: 12-bit RAW
    6. Focal Length: 5.4 mm
    7. Field of View: 47.2° HFOV, 35.4° VFOV
    8. Image Digital Size: 1280 × 960 (1.2 MP × 5 images)

Experiment Site

The experiment was conducted at the North Dakota State University (NDSU) Casselton Agronomy Seed Farm (CASF), 46.5317°N, -97.1403°W, Prosper Research Center (PRC), 46.5956°N, 97.0628°W, and Carrington Research Extension Center (CREC), 47.3116°N, -99.0751°W, during the summer of 2019. Compared to other states in the U.S., such as South Dakota, the climate in North Dakota exhibits distinct seasonal variations. During the summer, North Dakota experiences predominantly hot and humid conditions, while its winters are characterized by freezing temperatures and substantial snowfall. In the summer of 2019, the average temperature recorded was 19.1 °C, marking a notable difference of -16.8 °C compared to the previous summer of 2018. The average precipitation during the same summer was 236.98 mm, which was also 10.41 mm more than the 2018 summer precipitation (Akyuz et al., 2019). CASF was planted with sunflowers for the previous year with no-till as the farming practice.

In contrast, CREC was planted with spring wheat using a conventional tillage method, and the third experiment site, PRC, was planted with oats in the previous season under no-till. The yield data for the PRC site was not collected as the field was flooded with heavy rain. Weeds in the experiment areas were controlled by treating them with the preemergence herbicide “Gangster” (Amaranthus rudis Amaranthus tuberculatus Valor, Valent U.S.A. Corporation, Walnut Creek, CA 94596).

Additionally, weeds that emerged during post-crop emergence were controlled by hand weeding when required. Unintended injuries due to drift from nearby fields were avoided by careful selection of experiment sites (fig. 1). Glyphosate-tolerant soybean (ND17009GT), which the Soybean Breeding Program at NDSU developed, was planted in all three experiment locations at a seeding rate of 1,250,000 seeds ac-1. The ND17009GT variety belongs to the early maturity (00-maturity) group, which makes it suitable for the northern areas of North Dakota. Planting of seeds for CASF and PRC sites was done on May 31, 2019, while planting seeds in CREC was done on May 21, 2019. Buffers of soybean rows were created in all three locations to minimize spray drift within plots during treatment application.

Figure 1. Experimental Locations: (a) An overview of the experiment locations spanning two counties. (b) A closer inspection of the field at the Carrington Research Extension Center in Foster County, where the experiment was conducted. (c) A detailed view of the field at the Casselton Agronomy Seed Farm in Cass County, which also served as an experimental site.

Soybean injuries were visually rated using a scale of 0% to 100% based on the crops' injury level anytime aerial images were taken. An estimate of 0% for a particular plot indicated no plant injury was detected in that plot. Meanwhile, an estimate of 100% in a plot indicated all the plants in that plot were dead. One person was dedicated to collecting the visual ratings for the whole period to maintain consistency in the visual ratings. The method used to rate herbicide injury levels is explained in table 2. The ratings of treated plots were compared to those of the controlled plot, and the injury levels were assigned depending on leaf cupping, leaf puckering, and plant height. In general, the phototoxicity injury rating was performed based on the severity of the symptoms exhibited by the plant.

Experiment Design

The experiment was laid out in a random complete block design, with the whole plot being the soybean growth stage and the treatment rate as the subplot with three replicates for the two locations (fig. 2). The main plot treatment was three growth stages of the soybean: third trifoliate (V3), sixth trifoliate (V6), and first reproductive stage (R1), while the subplot was the simulated dicamba drift. The growth stages of the plants within a plot were recorded when at least more than half of the plants within the plot had reached the expected growth stage (Robinson et al., 2013). The gap between V3 and V6 stages was about 12 days, and between V6 and R1, it was 14 days. Each plot, along with the buffers, measured 8 m in length and 3 m in width (fig. 2).

The sub-plots in this experiment received treatments of Engenia (dicamba N, N-bis-(3-aminopropyl) methylamine salt, 600 g/L, BASF Corporation, Research Triangle Park, NC) at varying rates. For CASF, the treatment rates were 0.49, 4.9, and 49 g.ae ha-1 of dicamba, equivalent to 0.086%, 0.875%, and 8.75% of the recommended 560 g.ae ha-1 application rate for Engenia in dicamba-resistant soybean. Additionally, a non-treated plot served as a control in CASF. In the case of CREC, the treatment rates were 0.98, 9.8, and 98 g.ae ha-1 of dicamba, corresponding to 0.175%, 1.75%, and 17.5%, respectively, of the recommended 560 g.ae ha-1 application rate for Engenia in dicamba-resistant soybean. Again, a non-treated control was included. No adjuvants were added to the tank mix, as the intended approach was to simulate drift in the field, and therefore, reducing herbicide particle movement was not factored during formulation. A carbon dioxide (CO2) backpack sprayer (Turbo Teejet Induction, Spraying Systems Co., Wheaton, IL) was used for spraying the dicamba at CREC. The sprayer was calibrated following a standard procedure to deliver 0.2 L m-2 at a pressure of 179 kPa through nozzles (TTI8002) suitable for the selected treatment rate to simulate drift. The average wind speed at the time of treatment for the three stages was approximately 1.4 m s-1, with a cloud cover of 20%, 40%, and 25% in the first, second, and third treatments, respectively. The corresponding air temperatures were also 27?, 18?, and 21? at CREC. In CASF, the same CO2 backpack sprayer was also used, but with a different delivery pressure of 276 kPa through nozzles (TT11001). The nozzles were selected based on the dicamba simulated rate for each location for effective delivery at the specified rate. The wind speed in CASF at the time of treatment was 0.313 m s-1 for the first treatment, 3.57 m s-1 for the second treatment, and 4.02 m s-1 for the third treatment. The corresponding cloud cover was approximately 10% for all three stages of treatment application.

Table 2. Visual rating of injury levels of soybean treated with dicamba.
Rating
(%)		Description
0Normal plant growth, no visual injury seen.
10Slightly visible plant injuries, like height reduction, leaves cupping, and slightly bend petioles.
20Moderate visible symptoms of plant injury such as height reduction, curled petioles, area covered by plant canopy, and crinkled leaves.
30Transition from moderate to high in height reduction, canopy cover, etc.
40Comparatively, about 40% of the plants become stunted, bending or stems start to curl,  top leaves show severe malfunction, and new leaves have been suppressed.
50Twisting of stem and petiole, plant height is highly reduced, with the suppression of new growth.
60An estimated 60% of plants exhibit severe canopy and height reduction; cupping of leaves is severe with twisted stems.
70About 70% of the plants exhibit a transition from severe to very severe leaf cupping, height reduction, and twisted stems.
80An estimated 80% of soybeans show severe injuries like dropping leaves, twisting stems, and severe height reduction.
90Plants show symptoms of death, with 90% dropping of leaves with twisted petioles and necrotic stems.
100Total death of plants
Figure 2. Field experimental layout at Casselton Agronomy Seed Farm (Casselton, ND, USA). The rates are a percentage of 560 g.ae ha-1 of dicamba Engenia.

Image Processing and Analysis

The RedEdge MX sensor captured approximately 700 images per flight in 5 bands. Each image had a resolution of 2064 × 1544 in 16-bit Tag Image File (TIF) format. After image collection, image processing software Pix4Dmapper (Pix4D SA, 1015 Lausanne, Switzerland) was used to preprocess the images for ortho-mosaicing. The Pix4Dmapper software package employed photogrammetry and computer vision algorithms to convert aerial and oblique images acquired by sUAS into aligned, geo-referenced 2D orthomosaics, 3D surface models, and point clouds. Typically, the Pix4D algorithm looks for hundreds of tie-points between overlapped images and stitches the individual images together to construct one ortho-rectified image of the entire crop field in a study location (Gross and Heumann, 2016). The software was employed to generate an orthomosaic image for each field. The orthomosaic generation process encompassed several systematic steps. In the initial processing phase, the foremost measure involved the selection of a full keypoints image scale to uphold the utmost precision in producing the mosaic image. Subsequently, the workflow progressed to the point cloud and mesh generation step, a pivotal phase contributing to the creation of a highly detailed representation of both terrain and vegetation. Finally, the sequence advanced to the digital surface model, orthomosaic, and vegetation index generation stages. This is the stage where the imported reflectance panel images are used for calibration of the images. This calibration process, a fundamental component, preceded the subsequent creation of the digital surface model, reflectance maps, and orthomosaics. Such calibration ensured the accuracy and reliability of our geospatial data analyses by aligning the outputs with the genuine ground conditions and accounting for illuminance factors (Pix4D, n.d.).

Table 3. Equations used in this study to calculate Vegetative Indices (VIs) in ArcGIS Pro with related bands.
Vegetation
Index		AbbreviationEquationReference
Normalized Difference Vegetative
Index		NDVIRouse et al. (1973)
Green
NDVI		GNDVIGitelson et al., (1996)
Enhanced
NDVI		ENDVIMaxmax,
(2015)
Normalized Difference
RedEdge		NDREBarnes et al., (2000)
Soil Adjusted Vegetative
Index		SAVIHuete,
(1988)
Green
SAVI		GSAVISripada R.,
et al., (2006)
Optimized
SAVI		OSAVIRondeaux
et al., (1996)

Vegetation Index Calculations

During image preprocessing, individual bands' mosaics and reflectance maps were generated. These mosaic bands were used to calculate seven vegetative indices: Normalized Difference Vegetation Index (NDVI), Green NDVI (GNDVI), Enhanced NDVI (ENDVI), Soil Adjusted Vegetation Index (SAVI), Optimized Soil Adjusted Vegetation Index (OSAVI), Green Soil Adjusted Vegetation Index (GSAVI), and Normalized Difference Red Edge Index (NDRE) (table 3). ArcGIS Pro software (Esri, Redlands, CA, USA) was used to calculate the VI's, plot segmentation, and image thresholding to delineate treated areas effectively. In this study, the simulated dicamba drift from the sprayer was mostly centered in the middle two rows in each plot. Therefore, the middle two rows in the plot were masked using a polygon feature and segmented (fig. 3) for further analysis and vegetative index calculation. A field shapefile was created before the image was segmented with the projection set to World Geodetic System (WGS) 1984 with Universal Transverse Mercator (UTM) Zone 14 North for North Dakota. Then, the Create Feature tool in ArcGIS Pro was used to create polygons representing the area of interest in each plot.

After segmentation, the Raster Calculator tool in the ArcGIS Pro software platform was used to calculate the VIs. The seven VIs selected to estimate dicamba injury are listed in table 3, along with the formulae used to calculate them.

In this experiment, the SAVI and OSAVI were utilized to compare plant visual ratings. These indices were chosen due to the fluctuating vegetative cover during various growth stages. The selection was particularly relevant when vegetation cover dropped below 40%, as soil background reflectance can impact indices like NDVI and GNDVI (Huete, 1988). Therefore, the SAVI was developed to minimize the influence of soil brightness during index calculation (Duddu et al., 2019). The OSAVI used a value of 0.16 as the canopy background adjustment factor (Steven, 1998). OSAVI has been shown to accommodate more significant soil variation than SAVI when the plant vegetative cover was low, and it was extremely sensitive when the vegetative cover exceeds 50% (Steven, 1998). An image threshold operation was performed to separate the visible soil background from the crop vegetation. In CASF, minimal weed presence and plant shadows were observed during image collection, as these factors often introduce noise in an image. The initial data analysis involved importing the required image bands into ArcGIS Pro. The raster calculator was employed to compute the desired VIs. The next step was to use the natural breaks classification method to separate the VIs into different classes based on the vegetation index colors. Subsequently, the extract by attribute tool facilitated image segmentation. After segmentation, the ‘Zonal Statistics as Table’ tool was utilized to extract the vegetation index (VI) values within the plots from the raster images.

Statistical Analysis

The data from CASF and CREC were presented separately since the field treatment rates in their respective locations were different. Weibull type 2 with a 3 parameter model (Knezevic et al., 2007) in the dose-response module package of R (R statistical software, R Foundation for Statistical Computing, Vienna, Austria) was used to analyze the biological response of simulated dicamba micro drift rates applied at the V3, V6, and R1 growth stages of soybean, and the various parameters (yield, visual injury, and plant height). The regression analysis was instrumental in estimating the dicamba doses (ED) values responsible for various injury levels (5%, 10%, and 50%) threshold losses in yield, visual injury, and reductions in plant height (Osipitan et al., 2019). The 3-parameter module was used because the response variables for the experiment conducted cannot be negative, and therefore the lower limit was fixed to zero (Knezevic et al., 2007). The estimated dose of dicamba (ED) from the analysis was used to identify the growth stage of the soybean that was most sensitive to dicamba drift. Smaller ED values for a stage signified that a lower drift of dicamba was capable of causing soybean injuries. A split-plot analysis of variance (ANOVA) linear model was also performed in R to determine further the simulated dicamba rates and the growth stage of soybean that were significantly different from one another in yield. As stated in section 2.3 in the experimental design, the soybean growth stage was taken as the whole plot design, while the three simulated dicamba rates, in addition to the non-treated control, were taken as the subplot in the analysis. Mean separation was done if any of the factors were significant in the analysis to determine which treatments differed significantly (this process is shown in fig. 4).

Figure 3. Segmented image of Casselton Agronomy Seed Farm (CASF), indicating area of interest within each experimental plot.

After analyzing the biological response of soybean to simulated dicamba drift at the V3, V6, and R1 growth stages, a stepwise regression model was developed to assess the reliability of vegetation indices (VIs) as estimators for both crop yield and herbicide injury in soybean plants. The PROC REG forward selection method was used to implement stepwise regression and identify significantly contributing VIs. Once the VIs were identified based on the significance level in contributing to yield loss, the entire dataset was divided randomly into training and testing data subsets. Approximately 70% of the total dataset of all treatments was used to develop the training model, and the remaining 30% of the data subset was used to test the regression model. The performance of regression models in estimating yield and injury parameters was evaluated by calculating the root mean squared error of each location data.

Figure 4. Flow chart depicting the steps taken to obtain desire results.

Results and Discussion

Biological Response of Dicamba-Susceptible  Soybean to Simulated Dicamba Drift Applied  at the Three Growth Stages of Soybean

The split-plot ANOVA conducted on CREC data indicated that both the simulated dicamba rate and soybean growth stage significantly influenced soybean yields, with their interaction also being significant for yield outcomes (table 4). Yet, from the data presented, a distinct trend in the interaction between dicamba rate and soybean growth stage was not evident. Nevertheless, the V3 growth stage displayed the least yield reduction impact when applying dicamba at 0.98 g.ai ha-1. In comparison, the most significant yield reduction happened at the V6 stage with a dicamba rate of 98 g.ai ha-1 (fig. 5a).

The yield reduction was highest when dicamba was applied at a rate of 98 g.ai ha-1 at the V6 growth stage because dicamba is a systemic herbicide that is absorbed through the roots or foliage and primarily affects the vegetative growth stage of the soybean plant (Werle et al., 2018). It may take some time for dicamba symptoms to appear after a plant has been exposed. In fact, it can take up to seven days for susceptible plants to begin showing signs of injury. This delay occurs because the herbicide moves to the actively growing areas of the soybean plant first, and injury symptoms only begin to appear in the vegetative growth stage after that. As such, when the soybean plant is exposed to dicamba drift in the V6 stage, this results in abnormal pod formation in the R1 stage, leading to yield reduction. So, the exposure of soybean to dicamba drift at the R1 stage will indicate that pod formation in the R1 stage will not be affected but rather pods in the R2 stage. Therefore, soybean plants exposed during the V6 stage had a greater yield loss than those exposed during the R1 stage, even though both were exposed to the same rate of 98 g.ai ha-1.

Table 4. Estimate of soybean yield in Carrington Research Extension Center that resulted from the combination of different dicamba rates applied at V3, V6, and R1 growth stages of soybean.[a],[b]
Soybean Growth Stage
and Rate Interaction		Soybean Yield
(kg ha-1)		Groups
V3:0.982698.3a
R1:0.982676.1a
V3:02652.5a
V3:9.82588.4a
V6:0.982458.5ab
R1:02438.0abc
V6:02424.8abc
V6:9.82410.2abc
R1:9.82397.4abc
V3:982226.9bc
R1:982097.4c
V6:98896.4d
LSD0.05349.6

    [a]    V3, V6 are the third and sixth vegetative trifoliate leaf stages of soybean, respectively; R1 is the first flower stage of soybean.

    [b]    Treatments with the same letter are not significantly different.

(a)(b)
Figure 5. Non-linear dose-response curves of soybean to simulated dicamba drift in Casselton Agronomy Seed Farm and Carrington Research Extension Center at V3, V6, and R1 growth stages with (a) soybean yield response to simulated dicamba in Carrington Research Extension Center and (b) Soybean yield response to simulated dicamba drift in Casselton Agronomy Seed Farm.

The response of the V6 growth stage to dicamba exposure being the most susceptible stage is somewhat contrary to what Osipitan et al. (2019) reported, where they recognize V7/R1 as the most sensitive stage to be exposed to dicamba drift. However, while our soybean was treated at the V3, V6, and R1 growth stages, they treated their soybean at the V2, V7/R1, and R2 growth stages. Reynolds (2014) conducted a simulation of dicamba drift on dicamba-susceptible soybean during the vegetative and reproductive growth stages. The simulation revealed that the V6 growth stage was the most vulnerable to dicamba exposure. During the present experiment, the exposure to dicamba drift was most sensitive at the V6 stage. However, no reduction in yield was observed until the exposure rate reached 98 g.ai ha-1 (fig. 5a). On the part of soybean growth stage that was least susceptible to dicamba drift, Robinson et al. (2013) recognized the early stages (V1-V3) of dicamba susceptible soybean as the growth stage that results in the least yield reduction when treated with dicamba. Out of the V3, V6, and R1 growth stages, V3 experienced the smallest decrease in yield in our experiment (fig. 5a). The V3 stage experienced a lower yield reduction as the crop had sufficient time to recover from dicamba exposure before entering the reproductive stage (Robinson et al., 2013). Even though V6 was the most critical stage to be exposed to dicamba drift, it required the highest amount of dicamba estimated dose (ED), 22.471 g.ai ha-1, to cause a 5% yield reduction (ED5), while R1 and V3 stage required 11.611 and 20.828 g.ai ha-1, respectively, to cause the same 5% yield reduction (table 6). Hence, with a 5% yield reduction, R1 emerged as the most susceptible stage to dicamba drift. Nonetheless, this reduction will not substantially impact yield because the yield loss was not significant at 95% significance level; however, visible injury symptoms may manifest. The situation was, however, different with ED10 and ED50, where V6 was most vulnerable with 27.634 and 70.990 g.ai ha-1.

However, interestingly, when the same analysis was performed on CASF data, only the simulated dicamba rate (0, 0.49, 4.9, and 49 g.ai ha-1) was significant on the yield. The comprehensive mean separation analysis of the CASF data unveiled critical insights into the impact of dicamba drift exposure on soybean yields. Remarkably, the findings underscore the absence of a significant yield difference when soybeans were subjected to dicamba drift at the rate of 0.49 g.ai ha?¹ in comparison to the untreated check field. This outcome was established with a significant level of 95%.

Equally striking are the results obtained from fields exposed to dicamba drift at rates of 4.9 g.ai ha?¹ and 49 g.ai ha?¹. In both of these cases, the soybean yield reduction exhibited no statistically significant variance, maintaining a 95% significance level throughout the analysis. These findings collectively emphasize the remarkable similarity in yield reduction between the treated fields, where dicamba drift was simulated at varying rates, and the untreated check field. Such outcomes provide valuable insights into the impact of dicamba drift on soybean cultivation and underscore the importance of precise drift rate management in agricultural practices (table 5).

Table 5. Estimate of soybean means yield from Casselton Agronomy Seed Farm due to different treatment from dicamba simulated drift.[a]
Simulated Dicamba Rates
(g.ai ha-1)		Soybean Yield
(kg ha-1)		Groups
01414.7a
0.491349.7a
4.9967.7b
49797.4b
LSD0.05330.3

    [a]    Treatments with the same letter are not significantly different.

However, as expected, increasing the active ingredients of dicamba resulted in decreasing susceptible soybean yield, which is the same as reported by Brown et al. (2009) and Osipitan et al. (2019). Increased levels of active ingredients amplify the severity of injuries like leaf cupping, reduced height, epinasty, and delayed flowering in susceptible soybean plants. Consequently, abnormal pod formation occurs, leading to diminished pod development and reduced yields (tables 6 and 7).

The variation in dicamba-susceptible soybean response across the two locations stems from the difference in simulated drift rates. Notably, the CASF simulated dicamba application was twice that of CREC. Despite CREC higher dosage, the interaction of rate and growth stage was influenced by V6 at 98 g.ai ha-1 and R1 at 98 g.ai ha-1. In conclusion, the susceptibility of soybeans to dicamba drift became a concern when exposed to rates exceeding 98 g.ai/ha in the late vegetative stage.

Table 6. Simulated dicamba drift dose that causes 5% (ED5), 10% (ED10), and 50% (ED50) soybean yield reduction in Carrington Research Extension Center.[a]
Growth
Stage		Soybean Yield Reduction
ED5 (S.E.)
(g.ai ha-1)		ED10 (S.E.)
(g.ai ha-1)		ED50 (S.E.)
(g.ai ha-1)
V320.828 (32.600)44.878 (44.500)1489.337 (3216.200)
V622.471 (7.100)27.634 (16.900)70.990 (2361.200)
R111.611 (14.900)31.903 (28.800)3209.408 (6809.100)

    [a]    Comparison of estimated dose (ED) was done using standard error (S.E).

Table 7. Simulated dicamba drift dose that cause 5% (ED5), 10% (ED10), and 50% (ED50) soybean yield reduction in Casselton Agronomy Seed Farm.[a]
Growth StageSoybean Yield Reduction
ED5 (S.E.)
(g.ai ha-1)		ED10 (S.E.)
(g.ai ha-1)		ED50 (S.E.)
(g.ai ha-1)
V30.001 (69.500)0.016 (138.800)73.379 (7336.000)
V60.092 (504.600)0.285 (3957.000)50.030 (657.700)
R10.584 (1.660)1.204 (2.916)32.815 (69.872)

    [a]    Comparison of estimated dose (ED) was done using standard error (S.E).

Visual Rating of Crop Injury  for Simulated Dicamba Drift

The analysis of visual injury for soybean was taken 30 days after treatment (DAT) because any crop injuries due to simulated dicamba drift will be visible to the eye at that time. The injuries ranged from 0% to 40%, with various symptoms such as leaf cupping, epinasty, growth reduction, and crinkle of leaf tips. The analysis of soybean injuries was done using a non-linear regression model. Both locations showed an increase in visual injuries with increasing doses. This was quite obvious, as other research conducted by Brown et al. (2009) and Silva et al. (2018) both reported the same trend. Figures 6a and 6b distinctly demonstrate that the severity of visual injuries correlates with the growth stage of treatment application. However, the disparity in the curves originates from the varying simulated drift between CASF and CREC.

In figure 6a, CREC visual injuries were notably higher during the R1 growth stage compared to V3 and V6 stages. For the simulated dicamba drift at 10 g.ai ha-1, injuries recorded were 5%, 8%, and 21% for V3, V6, and R1, respectively. At 98 g.ai ha-1, the injuries were 18%, 37%, and 40% for the same growth stages. Likewise, consistent outcomes emerged as 43% of dicamba-susceptible soybeans, visually assessed 4 weeks post-treatment, exhibited injury during the R1 growth stage. Meanwhile, the V4 and V6 growth stages recorded 26% and 42% injuries, respectively. This, however, was not the case in Schneider et al. (2019), when they observed the vegetative stage (V3) growth stage to be more susceptible to a sub-dose of dicamba than the reproductive stage (R2) when they rated the phytotoxicity symptoms at 28 DAT. Notwithstanding that, the sub-dose of dicamba applied at a rate of 2.8 g.ae ha-1 and below was not statistically different in the V3 and R2 growth stage (Schneider et al., 2019). In this experiment, however, the V6 growth stage experienced the least visual injuries observed in CASF and CREC. This scenario is evident in table 8, which shows that the R1 stage needs 0.3789 g.ai ha-1 of the estimated dose of dicamba to exhibit 5% visual injury, whereas V6 needs 9.795 g.ai ha-1 of the estimated dose of dicamba to exhibit the same 5% of visual injuries. ED10 and ED50 for V6 need 36.309 and 1119.8 g.ai ha-1 of effective dicamba dose to cause 10% and 50% of visual injuries, respectively, making the V6 growth stage the least sensitive stage to visual injury. Reynolds (2014) also observed V6 to be less sensitive to visual injury when the field was rated for visual injuries at 28DAT.

In CASF, the data also showed that the severity of visual injuries depended on the stage at which the treatment was applied. The visual injuries observed ranged from 5% to 35% with increasing simulated dicamba drift. More visual injuries were observed in the V3 growth stage compared to V6 and R1. This was similar to the results of Schneider et al. (2019), as explained above, when they visually rated the sub-dose of dicamba at 28DAT, with the V3 stage being more sensitive. Silva et al. (2018) also reported similar results when the vegetative growth stage (V5) recorded a little over 60% injuries, and the reproductive growth stage recorded just below 50% injuries when the effect of 7 g.ae ha-1 were visually rated at 4 weeks after treatment. The maximum visual injury recorded for the R1 growth stage was 25%, while the V6 growth stage recorded 15%, with V3 being 35% at the highest dose of 49 g.ai/ha. The injury symptoms shown in CASF confirmed that low doses of dicamba drift are capable of causing injuries to soybeans that are susceptible to dicamba. This is so because there was still an increase in visual injury with an increasing rate of simulated dicamba drift; meanwhile, the rate applied in CASF was half of what was applied in CREC. Table 9 shows the estimated dose of dicamba at three stages capable of causing 5%, 10%, and 50% of visual injuries. It is important to note that signs of visual injuries to soybeans do not correlate to yield loss unless there is persistence in the injuries through the growing season (Weidenhamer et al., 1989). Although the highest visual injury recorded for the experiment conducted in CASF was relatively low when compared to other research results, such as Osipitan et al. (2019), it is crucial to understand that visual assessments of crop injuries, particularly in soybeans, are subjective despite existing rating guidelines. As a result, consistent visual injury values should not be anticipated, even when treatment rates are identical. Nevertheless, a comparable trend is expected, where higher herbicide rates correspond to increased visual injuries.

(a)(b)
Figure 6. Non-linear dose-response curves of soybean to simulated dicamba drift in Casselton Agronomy Seed Farm and Carrington Research Extension Center at V3, V6, and R1 growth stages with (a) Visual rating of soybean response to simulated dicamba drift in Carrington Research Extension Center observed at 30 DAT (b) Visual ratings of soybean response to simulated dicamba drift in Casselton Agronomy Seed Farm observed at 30 DAT.
Table 8. Simulated effective dicamba drift dose needed to cause 5% (ED5), 10% (ED10), and 50% (ED50) of visual injury in Casselton Agronomy Seed Farm.[a]
Growth
Stage		Visual Injury
ED5(S.E.)
(g.ai ha-1)		ED10(S.E.)
(g.ai ha-1)		ED50(S.E.)
(g.ai ha-1)
V33.871 (66.445)10.306 (196.780)133.660 (3238.100)
V69.795 (375.220)36.309 (145.300)1119.800 (5003.300)
R10.379 (2.690)5.268 (34.749)5165.800 (3232.200)

    [a]    Comparison of estimated dose (ED) was done using standard error (S.E).

Table 9. Simulated estimated dicamba drift dose needed to cause 5% (ED5), 10% (ED10), and 50% (ED50) of visual injury in Carrington Research Extension Center.[a]
Growth
Stage		Visual Injury
ED5(S.E.)
(g.ai ha-1)		ED10(S.E.)
(g.ai ha-1)		ED50(S.E.)
(g.ai ha-1)
V32.394 (48.586)6.517 (143.420)89.571 (238.080)
V666.793 (238.840)170.880 (614.700)199.710 (7310.200)
R10.174 (1.698)2.403 (22.934)232.200 (2297.200)

    [a]    Comparison of estimated dose (ED) was done using standard error (S.E).

Plant Height Reduction from Simulated  Dicamba Drift at the Three Growth Stages

The non-linear regression model used to evaluate the soybean height reduction showed an immediate reduction of plant height for the V3 stage, followed by the V6, and no reduction at the R1 stage in CASF and CREC (fig. 7a and 7b). There was a 24.7% and 11% plant height reduction at the V3 and V6 growth stage in CREC compared to the non-treated control plant height. The height reduction in CASF for the V3 and V6 growth stages was 24.3% and 15.6%, respectively.

The absence of crop height reduction during the R1 stage and the delayed reduction in height during the V6 stage indicate the absence of drift between plots during treatments. If any drift occurred, it was not substantial enough to induce plant height reduction. Both locations exhibited heightened plant height reduction as the treatment rate increased for crops treated at the V3 stage. However, in the V6 stage, plant height reduction in CREC began at a rate of 9.8 g.ai ha-1, whereas in CASF, it commenced at a rate exceeding 9.8 g.ai ha-1. The ED5, ED10, and ED50 values for CASF and CREC (tables 10 and 11) show the estimated dose of dicamba required to cause 5%, 10%, and 50% height reduction.

Schneider et al. (2019) and Silva et al. (2018) reported similar results in which they recognized V3 as the most perilous stage to be exposed to dicamba in terms of height reduction. Schneider et al. (2019) reported a 53% height reduction in the V3 stage compared to a 30% height reduction in the R2 growth stage when 28 g.ae ha-1 was applied, and the data was collected at 28DAT. However, Meyeres et al. (2021) reported the contrary: the R1 growth stage experienced the highest height reduction of 36% compared to the V3 growth stage, with a 5% yield reduction. The results reported by Meyeres et al. (2021) were not entirely different from the results obtained by Kelley et al. (2005), where they also reported a 22% height reduction for V3 and a 28% height reduction for V7.

(a)(b)
Figure 7. Non-linear dose-response curves of soybean to simulated dicamba drift in Casselton Agronomy Seed Farm and Carrington Research Extension Center at V3, V6, and R1 growth stages with (a) Soybean height response to simulated dicamba drift in Carrington Research Extension Center and (b) Soybean height response to simulated dicamba drift in Casselton Agronomy Seed Farm.

Plant height reduction primarily occurred during the V3 growth stage due to dicamba-induced injury to soybean leaves and petioles, as well as the curtailed growth of the plant's apical meristem (Marques et al., 2021). During the V6 and R1 stages, the delayed onset of dicamba injury symptoms renders these phenological stages less prone to height reduction. This phenomenon arises because, by the time the crop begins to display injuries, the plant's height has nearly reached its final value.

Table 10. Simulated estimated dicamba drift dose that causes 5%, 10%, and 50% of plant height reduction in Casselton Agronomy Seed Farm.[a]
Growth
Stage		Soybean Height Reduction
ED5(S.E.)
(g.ai ha-1)		ED10(S.E.)
(g.ai ha-1)		ED50(S.E.)
(g.ai ha-1)
V33.0154 (5.379)7.672 (10.401)543.320 (328.250)
V636.469 (13.332)43.099 (11.675)92.342 (22.630)
R1131.610 (123.410)288.460 (205.620)502.540 (157.420)

    [a]    Comparison of estimated dose (ED) was done using standard error (S.E).


Table 11. Simulated estimated dicamba drift dose that causes 5%, 10%, and 50% of plant height reduction in Carrington Research Extension Center.[a]
Growth
Stage		Soybean Height Reduction
ED5(S.E.)
(g.ai ha-1)		ED10(S.E.)
(g.ai ha-1)		ED50(S.E.)
(g.ai ha-1)
V35.7727 (9.528)15.189 (19.057)125.410 (6.93.800)
V654.637 (58.625)96.381 (78.611)128.400 (4.61.330)
R10.014 (0.082)0.014 (0.064)0.061 (1.204)

    [a]    Comparison of estimated dose (ED) was done using standard error (S.E).

Yield Prediction Model

A multicollinearity test was made for all VIs before stepwise regression analysis to determine collinearity among the predictor variables. The study considered a regression model with a higher R2 value and lower RMSE as the best predictive model for estimating the yield. Five indices of NDRE, GNDVI, GSAVI, OSAVI, and SAVI exhibited multicollinearity out of all seven VIs. Therefore, these five indices were dropped, and a stepwise regression analysis was then performed on only NDVI and ENDVI to determine their significance in predicting the yield. The stepwise regression showed that NDVI was the only index significant in predicting the yield for the first and second flight days. The regression for the third flight day showed that neither index predicted the yield significantly. The coefficient of determination for CASF and CREC after the first flight day was 0.42 and 0.15, while it was 0.12 and 0.13 on the second flight day. A summary of the coefficients of determination for the CASF and CREC are given in table 12. The results showed that the data obtained in CASF performed better in using VI’s to predict yield when compared to CREC. The disparity in performance between the two locations could be due to environmental effects in our aerial data collected and also due to differences in soil.

The VI’s obtained in our case heavily relied on crop foliar coverage, which depended on soil fertility. Fields with high soil fertility will have high canopy coverage and, by extension, will have their VI’s performing better in yield estimation than fields with low soil fertility. This was clearly illustrated in results obtained by Oseland et al. (2021) when VI’s obtained in the V3 stage performed better in predicting yield loss as compared to those obtained in the R2 stage when soybean was responding to sublethal rates of dicamba exposure. Similar results showing the moderate performance of VI’s in predicting yield were observed by Al-Gaadi et al. (2016) when they used NDVI and SAVI (vegetative index) derived from Landsat-8 and Sentinel-2 to predict potato yield. Huang et al. (2015) also used NDVI to predict soybean yield treated with glyphosate, and the results were not different from the present study (Huang et al., 2015).

Table 12. Soybean growth stage vegetative index and their relationship with actual yield.[a],[b]
Image
Date		LocationVegetation
Index		p-
value		R2Prediction
Equation
07/03/2019CRECNDVI0.020.15y = 54.16NDVI+23.54
07/12/2019CRECNDVI0.040.12y = 31.73NDVI+34.2
06/28/2019CASFNDVI0.020.42y = 21.03NDVI-2.82
07/06/2019CASFNDVI0.040.31y = 6.80NDVI-1.85

    [a]    CREC which stands for Carrington research extension center.

    [b]    CASF which stands for Casselton agronomy seed farm.

This demonstrates that NDVI served as a moderate estimator of yield. However, its performance can be influenced by high biomass, introducing potential noise into the collected imagery. As previously emphasized, the vegetation indices (VIs) in our study relied on crop foliar reflectance, and their efficacy hinged on canopy coverage. It is crucial to note that NDVI may become saturated under dense foliage conditions (Allawai and Ahmed, 2020), leading to potentially unreliable results. Therefore, GNDVI was incorporated to address this issue, as it accounts for situations with thick foliage that could result in NDVI oversaturation. It's worth noting that there was no instance of NDVI oversaturation that necessitated GNDVI compensation; thus, GNDVI was excluded from the analysis due to its high correlation with NDVI.

Although determining the relationship between vegetative indices and yield loss is one of the objectives made for this research, it is important to recognize that the yield loss observed in this research was not the result of only simulated dicamba drift. Other factors, such as rainfall and soil fertility, also adversely affect crop yield. It was, therefore, not surprising that NDVI moderately predicted the crop yield in our model figure 8. To improve the model's performance for future research, it will be prudent to incorporate abiotic factors that lead to yield loss. Another vital factor to note is that if the plant cover is sparse and the soil surface is exposed, the soil background can introduce noise to the reflectance data in the case of foliar-dependent indices like NDVI (Ali et al., 2019).

Figure 8. Relationship between predicted yield and actual yield in Carrington Research Extension Center and Casselton Agronomy Seed Farm with (a) NDVI from first aerial sUAS data collection at Carrington Research Extension Farm, (b) NDVI from second aerial sUAS data collection at Carrington Research Extension Farm, (c) NDVI from first aerial sUAS data collected in Casselton Agronomy Seed Farm, and (d) NDVI from second aerial UAV data collected in Casselton Agronomy Seed Farm.

A summary of the performance indicators used in validating the prediction model is shown in table 13. The relationship between the predicted yield and the actual yield was higher at the CASF, resulting in an R2 of 0.46 compared to the CREC of 0.17. The difference in performance could be due to variations in soil type and other external factors, for example, cloud cover during data collection with the sUAS, which were not considered during data collection. Nevertheless, the accuracy of validating the prediction model for the CASF suggested similar results reported by Huang et al. (2015), when they reported 0.43 as their coefficient of determination on assessing soybean injury from glyphosate using multispectral imagery.

Conclusions

The results from the experiment provided evidence that the V6 growth stage of dicamba-susceptible soybean was the most vulnerable stage to exposure to dicamba herbicide; however, it will take the highest dose of 98 g.ai ha-1 for the yield loss to be significant. Among all the three growth stages, the V3 growth stage was less perilous to yield reduction upon exposure to dicamba drift. Regarding visual injuries, the inconsistency of the growth stage response to visual observation of injuries at the two locations was due to the difference in rates and environmental factors. This is supported by other research, such as Meyeres et al. (2021) and Osipitan et al. (2019), reporting R1 as the vulnerable stage upon visual observation, while Schneider et al. (2019) reported V3 as the most perilous stage after visual observation. The difference can also be attributed to the observer rating the injuries since the injury rating of crops is subjective and not objective.

When soybean fields are exposed to dicamba drift, crop injuries may occur, but they may not necessarily result in a decrease in yield. This was shown through a non-linear dose-response analysis of dicamba susceptible soybean, which measured the biological response (yield, plant height, and visual injury) to three different rates of simulated dicamba drift. The visual injury curve indicated that as the dosage increased, so did the level of visual injury at both locations. Additionally, a reduction in plant height was observed during the V3 and V6 growth stages at both locations when exposed to the highest rate. However, the experiment conducted in Casselton revealed that the growth stage at which dicamba-susceptible soybean are exposed to the herbicide is irrelevant if the dose is 50 g ai/ha and below, as there was no interaction between the growth stage and dicamba rate.

Table 13. A summary of the performance indicators used in selecting the best prediction model for yield estimation.[a],[b]
Image
Date		LocationRMSE
(%)		Std.
Dev		Pearson
R2		MBE
(%)
07/03/2019CREC11.230.760.1710.37
07/12/2019CREC5.360.510.154.09
06/28/2019CASF4.110.420.464.10
07/06/2019CASF3.840.290.413.84

    [a]    CREC which stands for Carrington Research Extension Center.

    [b]    CASF which stands for Casselton Agronomy Seed Farm.

The yield prediction model results also demonstrated that vegetation indices (such as NDVI) from images collected with the sUAS sensor can be used to estimate the yield of crops from dicamba drift. Despite the moderate performance of the model in predicting yield, this can be improved by incorporating other factors, such as environmental factors, into the model. One practical application of this result is settling dicamba drift injury cases in court or getting insurance claims from dicamba drift.

Acknowledgments

This material is based upon work partially supported by the U.S. Department of Agriculture, agreement number 58-6064-8-023. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the view of the U.S. Department of Agriculture. This work is/was supported by the USDA National Institute of Food and Agriculture, Hatch project number ND01487. This research is also partially supported by the North Dakota State University Agricultural Experiment Station Precision Agriculture Fund FARG080010. The authors would also like to thank Joseph Mettler, Dr. Ted Helms, and his team for supporting this experiment.

References

AgEagle. (2018). Why Narrow Bands Matter. Retrieved November 14, 2022, from https://ageagle.com/micasense-blog/why-narrow-bands-matter/

Akyuz, A., Kupec, R., Schlag, A., & Gust, G. (2019). North Dakota Climate Bulletin, 13, 1-14.

Al-Gaadi, K. A., Hassaballa, A. A., Tola, E., Kayad, A. G., Madugundu, R., Alblewi, B., & Assiri, F. (2016). Prediction of potato crop yield using precision agriculture techniques. PLoS One, 11(9), e0162219. https://doi.org/10.1371/journal.pone.0162219

Ali, A., Martelli, R., Lupia, F., & Barbanti, L. (2019). Assessing multiple years’ spatial variability of crop yields using satellite vegetation indices. Remote Sens., 11(20), 2384. https://doi.org/10.3390/rs11202384

Al-Khatib, K., Hanson, B., Miller, T., Peachey, E., & Boydston, R. (2013). Managing glyphosate-resistant weeds in glyphosate-resistant crops. 8494. ANR Publication. https://doi.org/10.3733/ucanr.8494

Allawai, M. F., & Ahmed, B. A. (2020). Using remote sensing and GIS in measuring vegetation cover change from satellite imagery in Mosul City, north of Iraq. IOP Conf. Ser. Mater. Sci. Eng., 757(1), 012062. https://doi.org/10.1088/1757-899X/757/1/012062

Bodnar, A. (2018). Dicamba drift - Part 1 What is dicamba? Biology Fortified Inc. Retrieved from https://biofortified.org/2018/11/dicamba-drift-1/

Brown, L. R., Robinson, D. E., Nurse, R. E., Swanton, C. J., & Sikkema, P. H. (2009). Soybean response to simulated dicamba/diflufenzopyr drift followed by postemergence herbicides. Crop Prot., 28(6), 539-542. https://doi.org/10.1016/j.cropro.2009.02.004

Christoffoleti, P. J., Alves de Figueiredo, M. R., Peres, L. E., Nissen, S., & Gaines, T. (2015). Auxinic herbicides, mechanisms of action, and weed resistance: A look into recent plant science advances. Sci. Agric., 72(4). https://doi.org/10.1590/0103-9016-2014-0360

Dowell, T. (2017). Dicamba update (Part I). Texas A&M Agrilife Extension. Retrieved from https://agrilife.org/texasaglaw/2017/11/13/dicamba-update/

Duddu, H. S., Johnson, E. N., Willenborg, C. J., & Shirtliffe, S. J. (2019). High-throughput UAV image-based method is more precise than manual rating of herbicide tolerance. Plant Phenomics, 2019. https://doi.org/10.34133/2019/6036453

Egan, J. F., & Mortensen, D. A. (2012). Quantifying vapor drift of dicamba herbicides applied to soybean. Environ. Toxicol. Chem., 31(5), 1023-1031. https://doi.org/10.1002/etc.1778

Egan, J. F., Barlow, K. M., & Mortensen, D. A. (2014). A meta-analysis on the effects of 2,4-D and dicamba drift on soybean and cotton. Weed Sci., 62(1), 193-206. https://doi.org/10.1614/WS-D-13-00025.1

Gitelson, A. A., Kaufman, Y. J., Merzlyak, M. N. (1996). Use of a green channel in remote sensing of global vegetation from EOS- MODIS. Remote Sensing of Environment, 58(3), 289–298. https://doi.org/10.1016/S0034-4257(96)00072-7

Gross, J. W., & Heumann, B. W. (2016). A statistical examination of image stitching software packages for use with Unmanned Aerial Systems. Photogramm. Eng. Remote Sens., 82(6), 419-425. https://doi.org/10.14358/PERS.82.6.419

Hartzler, B. (2017). Dicamba off-target injury update. Integrated Crop Management News, Iowa State University Extension and Outreach. Retrieved from https://crops.extension.iastate.edu/?cropnews/2017/08/dicamba-target-injury-update

Henry, W. B., Shaw, D. R., Reddy, K. R., Bruce, L. M., & Tamhankar, H. D. (2004). Remote sensing to detect herbicide drift on crops. Weed Technol., 18(2), 358-368. https://doi.org/10.1614/WT-03-098

Honrado, J. L., Solpico, D. B., Favila, C. M., Tongson, E., Tangonan, G. L., & Libatique, N. J. (2017). UAV imaging with low-cost multispectral imaging system for precision agriculture applications. Proc. 2017 IEEE Global Humanitarian Technology Conf. (GHTC), (pp. 1-7). https://doi.org/10.1109/GHTC.2017.8239328

Huang, Y., Reddy, K. N., Thomson, S. J., & Yao, H. (2015). Assessment of soybean injury from glyphosate using airborne multispectral remote sensing. Pest Manag. Sci., 71(4), 545-552. https://doi.org/10.1002/ps.3839

Huang, Y., Yuan, L., Reddy, K. N., & Zhang, J. (2016). In-situ plant hyperspectral sensing for early detection of soybean injury from dicamba. Biosyst. Eng., 149, 51-59. https://doi.org/10.1016/j.biosystemseng.2016.06.013

Huete, A. R. (1988). A soil-adjusted vegetation index (SAVI). Remote Sens. Environ., 25(3), 295-309. https://doi.org/10.1016/0034-4257(88)90106-X

Kelley, K. B., Wax, L. M., Hager, A. G., & Riechers, D. E. (2005). Soybean response to plant growth regulator herbicides is affected by other postemergence herbicides. Weed Sci., 53(1), 101-112. https://doi.org/10.1614/WS-04-078R

Knezevic, S. Z., Streibig, J. C., & Ritz, C. (2007). Utilizing R software package for dose-response studies: The concept and data analysis. Weed Technol., 21(3), 840-848, 9. https://doi.org/10.1614/WT-06-161.1

Laliberte, A. S., Goforth, M. A., Steele, C. M., & Rango, A. (2011). Multispectral remote sensing from unmanned aircraft: Image processing workflows and applications for rangeland environments. Remote Sens., 3(11), 2529-2551. https://doi.org/10.3390/rs3112529

Marques, M. G., da Cunha, J. P., & Lemes, E. M. (2021). Dicamba injury on soybean assessed visually and with spectral vegetation index. AgriEngineering, 3(2), 240-250. https://doi.org/10.3390/agriengineering3020016

Matese, A., Toscano, P., Di Gennaro, S. F., Genesio, L., Vaccari, F. P., Primicerio, J.,... Gioli, B. (2015). Intercomparison of UAV, aircraft and satellite remote sensing platforms for precision viticulture. Remote Sens., 7(3), 2971-2990. https://doi.org/10.3390/rs70302971

Maxmax. (2015). Enhanced Normalized Difference Vegetation Index (ENDVI). Retrieved January 8, 2022, from https://maxmax.com/endvi.htm

Meyeres, T., Lancaster, S., Kumar, V., Roozeboom, K., & Peterson, D. (2021). Response of non-dicamba-resistant soybean (Glycine max) varieties to dicamba. Weed Technol., 35(5), 718-724. https://doi.org/10.1017/wet.2021.4

Moran, M. S., Inoue, Y., Barnes, E. M. (2000). OpportunitiesLimitationsImageBasedCropManagement, 319346. Retrieved from https://pubag.nal.usda.gov/?pubag/downloadPDF.xhtml?id=11029&content=PDF

Olszyk, D., Pfleeger, T., Lee, E. H., & Plocher, M. (2015). Glyphosate and dicamba herbicide tank mixture effects on native plant and non-genetically engineered soybean seedlings. Ecotoxicology, 24(5), 1014-1027. https://doi.org/10.1007/s10646-015-1442-8

Oseland, E., Shannon, K., Zhou, J., Fritschi, F., Bish, M. D., & Bradley, K. W. (2021). Evaluating the spectral response and yield of soybean following exposure to sublethal rates of 2,4-D and dicamba at vegetative and reproductive growth stages. Remote Sens., 13(18), 3682. https://doi.org/10.3390/rs13183682

Osipitan, O. A., Scott, J. E., & Knezevic, S. Z. (2019). Glyphosate-resistant soybean response to micro-rates of three dicamba-based herbicides. Agrosyst. Geosci. Environ., 2(1), 180052. https://doi.org/10.2134/age2018.10.0052

Pix4D. (n.d.). Radiometric calibration target. Retrieved from https://support.pix4d.com/hc/en-us/articles/206494883-Radiometric-calibration-target#label4

Reynolds, D. (2014). Determinig the effect of dicamba rate and application timing on soybean growth and yield. Project No. 42-2014. Mississippi Soybean Promotion Board.

Robinson, A. P., Simpson, D. M., & Johnson, W. G. (2013). Response of glyphosate-tolerant soybean yield components to dicamba exposure. Weed Sci., 61(4), 526-536. https://doi.org/10.1614/WS-D-12-00203.1

Rondeaux, G., Steven, M., Baret, F. (1996). Optimization of soil-adjusted vegetation indices. Remote Sensing of Environment, 55(2), 95–107. https://doi.org/10.1016/0034-4257(95)00186-7

Rouse, J. W., Hass, R. H., Schell, J. A., Deering, D. W. (1973). Monitoring vegetation systems in the great plains with ERTS. Third Earth Resources Technology Satellite (ERTS) Symposium, 1, 309–317. https://doi.org/citeulike-article-id:12009708

Schneider, T., Rizzardi, M. A., Rockenbach, A. P., & Peruzzo, S. T. (2019). Subdoses of dicamba herbicide on yield componentes in function of the soybean growth stage. J. Agric. Sci., 11(6), 407. https://doi.org/10.5539/jas.v11n6p407

Silva, D. R., Silva, E. D., Aguiar, A. C., Novello, B. D., Silva, Á. A., & Basso, C. J. (2018). Drift of 2,4-D and dicamba applied to soybean at vegetative and reproductive growth stage. Ciência Rural, 48(8). https://doi.org/10.1590/0103-8478cr20180179

Soltani, N., Nurse, R. E., & Sikkema, P. H. (2016). Response of glyphosate-resistant soybean to dicamba spray tank contamination during vegetative and reproductive growth stages. Can. J. Plant. Sci., 96(1), 160-164. https://doi.org/10.1139/cjps-2015-0175

Sripada, R. P., Heiniger, R. W., White, J. G., Meijer, A. D. (2006). Aerial color infrared photography for determining early in-season nitrogen requirements in corn. Agronomy Journal, 98(4), 968–977. https://doi.org/10.2134/agronj2005.0200

Sterling, T. M., & Hall, C. J. (1997). Mechanism of action of natural auxins and the auxinic herbicides. Rev. Toxicol., 1(3-4), 111-141.

Steven, M. D. (1998). The sensitivity of the OSAVI vegetation index to observational parameters. Remote Sens. Environ., 63(1), 49-60. https://doi.org/10.1016/S0034-4257(97)00114-4

Stroppiana, D., Villa, P., Sona, G., Ronchetti, G., Candiani, G., Pepe, M.,... Boschetti, M. (2018). Early season weed mapping in rice crops using multi-spectral UAV data. Int. J. Remote Sens., 39(15-16), 5432-5452. https://doi.org/10.1080/01431161.2018.1441569

Thurman, B. C., Lee, P. O., Gay, M. M., & Mccants, J. C. (2017). Off target?: Dicamba drift issues ensnaring farmers.

USDA. (n.d.). USDA ERS - Biotechnology. Retrieved November 4, 2019, from https://www.ers.usda.gov/topics/farm-practices-management/biotechnology/

Wechsler, S. J., Smith, D., McFadden, J., Dodson, L., & Williamson, S. (2019). The use of genetically engineered dicamba-tolerant soybean seeds has increased quickly, benefiting adopters but damaging crops in some fields. Amber Waves, USDA-ERS. Retrieved from https://www.ers.usda.gov/amber-waves/2019/october/the-use-of-genetically-engineered-dicamba-tolerant-soybean-seeds-has-increased-quickly-benefiting-adopters-but-damaging-crops-in-some-fields/

Weidenhamer, J. D., Triplett Jr., G. B., & Sobotka, F. E. (1989). Dicamba Injury to Soybean. Agron. J., 81(4), 637-643. https://doi.org/10.2134/agronj1989.00021962008100040017x

Werle, R., Proost, R., & Boerboom, C. (2018). Soybean injury from dicamba.

Zhao, F., Huang, Y., Guo, Y., Reddy, K. N., Lee, M. A., Fletcher, R. S., & Thomson, S. J. (2014). Early detection of crop injury from glyphosate on soybean and cotton using plant leaf hyperspectral data. Remote Sens., 6(2), 1538-1563. https://doi.org/10.3390/rs6021538