Introduction
The ultimate goal of remote sensing in weed science has always been sense and spray, but identifying weeds and treating them effectively still has limitations on repeatability and reliability. It has been said the future of site-specific weed management hinges on the ability to discriminate weeds from crops in real-time (Shaner & Beckie, Reference Shaner and Beckie2014); however, our ability to achieve cost-effective, truly automated weed detection and control is still not fully mature. New possibilities for weed science have been enabled by the recent introduction of remote sensing with unmanned aerial vehicles (UAV), but their capabilities have not been fully researched and reported. Unmanned aerial vehicles have been more of a novelty that, partially due to regulations (at least in the USA), have been little employed in agricultural operations. This situation also exists within the agriculture research community, as scientists either lack access to UAV technology or the expertise to take advantage of the data collected. As precision agriculture researchers continue to move the cutting edge forward, it will be important to reach out to end-users and help them understand the innovations taking place.
Adoption of UAV technology has also been limited by the lack of information on the value and utility of this technology; this is particularly true in weed science (Lopez‐Granados, Reference Lopez‐Granados2011, Rasmussen et al., Reference Rasmussen, Nielsen, Garcia‐Ruiz, Christensen and Streibig2013). While researchers continue to chase after goals of real-time site-specific weed management, it is imperative that the user base be increased by demonstrating practical applications of UAVs. An UAV will not fix every problem on the farm, but it can be a tool that provides value when applied appropriately. Rasmussen et al. (Reference Rasmussen, Nielsen, Garcia‐Ruiz, Christensen and Streibig2013) concluded that information extraction workflows were needed to move the science forward in weed science. However, complicated extraction workflows, which may involve multiple software packages with steep learning curves, will not result in end-user adoption unless those workflows can be condensed to automated routines. In this paper, some areas where UAVs show potential in weed science are presented. These are couched within three examples of how UAVs were used for specific weed science applications, along with observations and lessons learned. It is intended that these examples are near-readily adoptable by end-users with some skill and understanding of the science, but who are not necessarily experts (e.g., crop advisors, tech savvy producers, independent UAV companies that provide a service to producers, scientists on the fringe of precision agriculture). As such, this paper does not include robust experimental design or rigorous statistics; merely proof of concept and suggestions about how the precision agriculture community might engage new end-users with simple techniques for UAVs in weed science.
Quantifying non-target herbicide injury
In this example, a small-plot research field planted to corn (Zea mays) was severely damaged by carryover from trifloxysulfuron-sodium herbicide, a sulfonylurea herbicide marketed as Envoke. The prior year, the field had been the site of a cotton (Gossypium hirsutum) research study. The cotton researcher had made two foliar applications of the herbicide for weed control at the label-specified rate of 0.003 kg AI ha−1. By late April, the seven-month crop rotation restriction for corn had elapsed; however, it became obvious by June that carryover herbicide injury was affecting the corn stand in the areas where cotton had been planted. Herbicide injury appeared greater in areas where cotton stand had been thin, and in previous alleys between plots where no cotton had been planted the previous year.
Data Collection
Imagery was collected with a multi-rotor UAV (Sky-Hero Spyder X8) equipped with a multispectral sensor (Micasense RedEdge) and a standard red-green-blue (RGB) camera (Canon SL1). Flight altitudes used produced multispectral imagery of 4 cm2 resolution and standard imagery of 0.65 cm2 resolution. The original intent was to perform other research in this c.7.5 ha field, and image collection commenced at emergence (early May). However, once injury was apparent, imagery was collected monthly to monitor stunted growth, with images collected in mid-May, June, and July. It was deemed economically disadvantageous to harvest the remaining corn, thus yield data were not available.
Applicability
The timely response of the UAV and the overhead view make this tool of use even to the novice user for assessing stand loss or stunting. Extent of damage can be visualized and measured with UAV imagery. Herbicide injury, both carryover and drift, is just one type of damage, but failed emergence, early-season flooding, or herbivory could also create the need to quantify extent and severity of damage. UAVs may become another tool for producers as they calculate economic losses and make re-plant decisions by combining image data with an estimate of yield and price per bushel.
Approach
Multiple approaches were pursued to quantify losses, and each had inherent error. One advantage of the cloud-based processing services offered in the marketplace is the ability to easily view and download geospatial layers of normalized difference vegetation index (NDVI) and other indices. It would be expected that NDVI would decrease because of plant senescence due to injury. Dicke et al. (Reference Dicke, Jacobi and Büchse2012) previously used multispectral imagery to determine NDVI, and coupled with yield data, to quantify injury to corn from other sulfonylurea herbicides. The easiest approach was to do a zonal comparison in a geographic information system between NDVI values in unaffected and affected areas. The mid-June and July images were divided into zones and average NDVI values were taken. Average NDVI values in unaffected zones averaged 0.65 for mid-Jun, and 0.88 for mid-July. Within the affected area, the 95% quantile for NDVI values begins at 0.54 for mid-June and 0.76 for mid-July, indicating a reduction in potential of c.15% when compared to the unaffected area. The assumption here is that the unaffected or “control” areas would have yielded similar to the affected areas and be otherwise homogeneous to the affected area in terms of soil, irrigation, etc. Considering only the research plots, and using them as experimental units, in mid-June, the median plot NDVI value was 0.33, and in mid-July the median value was 0.61. A Wilcoxon signed rank test on plots confirmed that these values were statistically different (p<0.0001), and histograms showing the frequency of NDVI values in mid-June and mid-July also showed a general shift in values; despite the overall increase in median plot NDVI there were still many areas of missing plants. To get away from generalization, a third approach involved use of a plant stand count algorithm. Much attention has been given to this type of analysis and this tool is available from various cloud-based processing services for use with UAV data, making it easy to access by end-users. A template matching algorithm (Tanimoto, Reference Tanimoto1981) was attempted with the May image to get a count of emergence, with the thought that this value could be compared to the known quantity of plants in each plot for accuracy estimation and overlaid with the mid-July image to see where plants no longer covered the soil. It was determined that the algorithm was not sensitive enough, and this resulted in overestimation of initial and remaining plants, generally because the algorithm classified a high amount of tire tracks within the field as plants. Further refinement was done to remove misclassified objects that were not within the planted row (Pena et al., Reference Pena, Torres-Sanchez, de Castro, Kelly and Lopez-Granados2013), but error still persisted; accuracy varied, but the algorithm missed a median value of 25% (m.a.d. 5%) of the actual plants in a subset of 200 plants (20 plants per 10 rows). In a research field, this is easily addressed with a bit of time, but as this would never be time efficient in a production field, the method was abandoned. In conclusion, the simplest approach may likely prove the most transferable, and perform with similar accuracy as more sophisticated options, given the potential sources of error observed in these approaches.
It is worth noting a lesson learned. Photos were taken weekly from a pole mounted at the field margin at a height of 1.5 m; this was to simulate the “truck window” view so that it could be shown that more information is generally available from above. The pole was mounted at the field margin where no damage occurred. From the field photos taken on the mid-July date, the corn plants are at c.R1 maturity level (when corn has almost reached full height) and it is not possible to see that behind several rows of healthy corn lies row after row of stunted plants.
Calculating spray thresholds
The second application is calculation of spray thresholds based on weed populations. In this example, a c.16 ha production rice (Oryza sativa) field with significant infestation of hemp sesbania (Sesbania exaltata) was analyzed using nearest neighbor analysis of weed density to determine an economic threshold for herbicide application. Hemp sesbania is one of the most common and problematic weeds in rice production. It can make harvesting more difficult and contaminate harvested rice with its easily-noticeable black seeds. Mississippi’s Rice Growers’ Guide (Miller et al., Reference Miller, Street, Buehring, Kanter, Walker and Bond2008) recommends treatment of hemp sesbania at a density of 1–2 plants m−2. A standard treatment for this situation is acifluorfen, marketed as Ultra Blazer. The rice producer intended to make an application at the label rate of 0.14 kg AI ha−1.
Data Collection
UAV imagery was obtained in early August. For reference, in Mississippi, rice is planted in April or May; harvest is in September or October. Data were collected using a fixed wing aircraft (Precision Hawk V3) equipped with a multispectral sensor (Micasense RedEdge). The sensor provides a digital surface model in the delivered cloud-processed output, which was considered potentially informative as mature hemp sesbania can be up to 4 m tall; considerably taller than mature rice stands.
Applicability
Spray thresholds can vary between whole field and site-specific application (Longchamps et al., Reference Longchamps, Panneton, Simard and Leroux2014). Unmanned aerial vehicles have great potential in this area because they provide near-real-time data on weed density and distribution at low cost (Lopez‐Granados, Reference Lopez‐Granados2011) and can be deployed on-demand, facilitating comparisons of whole field versus site-specific analysis in the overall threshold evaluation. While pre- and mid-season herbicide applications are a given, late season applications are less certain and must be justified; advanced tools and technologies are needed to facilitate spot spraying of late season weeds, which is seldom warranted by economic thresholds (Bagavathiannan & Norsworthy, Reference Bagavathiannan and Norsworthy2012).
Approach
Multiple approaches were used to estimate weed density, which is the driver for a spray threshold decision. The first approach was a color segmentation analysis based on the bands collected with the multispectral sensor to separate hemp sesbania from rice. This approach required finesse and easily over or underselected weedy patches. The second approach, which proved easier and more reliable, utilized the digital surface model information provided by the sensor’s cloud processing service. Using focal statistics, the elevation recorded in each image pixel was compared to the value of the neighboring pixels (Chapman et al., Reference Chapman, Merz, Chan, Jackway, Hrabar and Dreccer2014). Those pixels which were significantly different from adjacent pixels were marked as potential hemp sesbania plants. The cutoff between crop and weed was based on the histogram of difference values, in this case 0.025 m, and the image was reclassified into two classes. This method proved satisfactory but captured field edges at levees and thin spots/holes in the crop canopy, in addition to weeds. This is because these areas also exhibited large differences in elevation between adjacent pixels; this could be fixed with masking. A 1 m x 1 m vector grid was created which was overlaid on the two-class image. The empty vector grid cells were assigned a value based on the underlying class, creating a map of grid cells meeting the criteria for herbicide treatment, which was 1 plant per grid cell based on the recommendation from Miller et al. (Reference Miller, Street, Buehring, Kanter, Walker and Bond2008); this criteria could be adjusted for other weeds in other crops. A decision on whether to spot-spray can be based on the count of grid cells relative to the total count of grid cells; cell counts can also be used to justify making an application for less competitive weeds.
Nearest neighbor analysis was performed to estimate the degree of clustering for treatment cells to determine if not only application, but site-specific application, was warranted. Economic analyses must consider the fixed cost of making an application, and also the variable costs of herbicide and related additives (e.g., surfactants). The producer in this example spent US$42 ha−1 to treat the entire field, of which US$20 ha−1 went to the actual cost of the herbicide. Based solely on the treatment grid, over 50% of the field did not require treatment. However, another factor in the equation is the ease which with site-specific application can be performed and the potential savings. For highly dispersed weeds, whole field application will likely be the preferred option. In this field, weeds were too dispersed and the potential savings too low to warrant site-specific application. Thus while the UAV can provide information on the amount of treated versus untreated area and the degree of clustering or dispersion of weedy areas, these are only decision aids for producers and trusted advisors. Having this type of information will allow users to determine where economic thresholds occur for making a herbicide application later in the season, and if whole field versus site-specific application would be the most financially advantageous.
One item of note relates to elevation trends. The digital surface model from the sensor’s cloud-based processing service gives elevation in meters above sea level. Because rice fields are generally sloped to facilitate movement of water into the outflow drain, this gradient was detected by the sensor. The use of relative heights helped overcome this issue, but could prove more challenging in terrains that are not laser-leveled.
Site-specific herbicide application
While there is more attention given to production agriculture, remote sensing applications have also been examined in turfgrasses. Bell and Xiong (Reference Bell and Xiong2008) reviewed optical sensing in turf and noted that precision weed control in turf could produce significant economic savings while reducing the environmental impact of herbicide use, but also noted that the necessary sensors made precision weed management undesirable to end-users. At the time of this publication, available sensors were either not appropriate for turfgrasses (e.g., resolution mismatch) or cost prohibitive due to the need for multiple sensors on spray booms for sense-and-spray operation. Still, most of the published research in this area relates to nitrogen status and quality of turf, rather than weed science aspects, with almost no work on site-specific herbicide application. Spot spraying of weeds allows a manager to use a more costly, but more selective, herbicide because it will be applied in limited amounts. This can create better control with less risk of injury to the desired turfgrass. In this example, imagery collected with UAVs was used to identify and spatially define annual bluegrass (Poa annua) present as a weed within maintained hybrid bermudagrass (Cynodon dactylon × C. transvaalensis) turf. A second example of the same workflow was applied to bermudagrass encroachment in Zoysia japonica. These represent two of the major turfgrass species used for golf course fairways in the Southeastern USA.
Data Collection
A multi-rotor aircraft (Sky-Hero X8) equipped with a multispectral sensor (Micasense RedEdge) was deployed in January over research plots of dormant maintained hybrid bermudagrass. Data collection for the second example was performed with a multi-rotor aircraft (DJI Phantom 2 Vision Plus) equipped with a thermal sensor (FLIR Tau 2), which was deployed in-season over experimental fairways of established zoysia turfgrass.
Applicability
Technological advances now provide the opportunity to have georeferenced weed maps on mobile platforms, which Carrow et al. (Reference Carrow, Krum, Flitcroft and Cline2010) believed was crucial to driving so-named “precision turf management” forward. Many of the cloud-based processing services for UAV data offer mobile capabilities, such as locations services. For weeds that are highly visible, using a smartphone or tablet with GPS capability, this map alone would be enough to navigate to weeds which are clearly visible to make a spot spray application for low occurrences of weeds; this approach requires no processing expertise. With some understanding of the weed and turfgrass lifecycles, data collection can be appropriately designed to achieve detection. For example, bermudagrass has periods of dormancy in the winter, but the problematic weed, annual bluegrass does not. At the time of data collection, the only actively growing vegetation was annual bluegrass, making January an ideal time for collecting image data for a spot spray map. In instances where these types of differentials occur, this approach could be easily implemented by most users. With a spot spray map to consult, the area (i.e., spots) can be measured, and the costs of spot spray can be weighed against the cost of treating the entire fairway, making it possible to calculate the economic differences.
Approach
As stated, annual bluegrass was selected because it remains green at a time when bermudagrass is dormant. However, dormancy differentials are just one option. Zoysia fairways are observed to have encroachment from weedy bermudagrass, so it is desirable to sense and spray these encroachments. Bermudagrass has been observed to have longer dew retention than other species of turfgrass, including zoysia. For this example, thermal imagery was collected in the morning when dew had evaporated from the zoysia but was still clinging to the leaves of the bermudagrass. This allowed identification of the areas of bermudagrass infestation due to the differential temperature between the species. Although traditionally thermal sensors have had more challenging calibration protocols, the marketplace is seeing improvements in these sensors and they are becoming more accessible to less-skilled (and less affluent) users. Two drawbacks to thermal imagery are that cloud processing and mobile capabilities are not possible with the thermal sensor utilized, and thermal resolutions are considerably coarser than many RGB and multispectral options. However, a processed image could be loaded onto a mobile platform and used in a similar fashion to cloud-based options.
This “navigate to point” approach is limited to low occurrences of weeds; therefore, it was logical to develop a spray controller file for a site-specific application. Although not widely utilized, the technology does exist for variable rate application with turf management equipment. A controller file was produced by performing supervised classification on the UAV imagery to separate the image into two classes, weeds and non-weeds. The resultant image was further processed to remove areas which fell outside the treatment area that were collateral capture within the UAV imagery. A vector grid was created to match the dimensions of the available spray boom. Attributes were assigned to the vector grid cells based on the underlying classified UAV image, producing a shapefile which could be converted into a controller file for a variable rate sprayer. It should be noted that this solution does require more steps than a typical end-user might be comfortable with and requires some knowledge of controller file structure.
Looking at a bigger picture, the turfgrass industry is also experiencing difficulties with herbicide resistant weeds. Annual bluegrass is among those resistant weed species of concern, resistant to multiple modes of action, and with resistant specimens currently identified in ten countries (Heap, Reference Heap2016). Judicious use of herbicides will be warranted in the future for limiting development of resistance. At the same time, many regions are reducing chemical inputs to maintain environmental quality; this technology can be a useful tool in efforts for both.
Conclusions
Unmanned aerial vehicles offer what University of Nebraska’s Dr. Wayne Woldt called “an unparalleled and powerful view of the landscape that would be very difficult to get from the ground when monitoring large parcels of land.” There remains much to be learned about how this technology can benefit precision agriculture, particularly in weed science. Near-real-time data are an asset for weed management, and UAVs can provide on-demand data collection; the challenge will be identifying how to best capitalize on this advantage. Areas where UAVs can play a role in weed management include evaluating herbicide injury, calculating spray thresholds, and planning site-specific application of herbicide. All of these efforts enable economic decision making opportunities for users, which is ultimately how UAVs will show their value. If UAVs are really the tool of the future, it is imperative that they be accessible to a broad end-user base, with a range of non-technical to technical applications. While precision agriculture specialists move the state of the science forward, they should endeavor to bring others with them.
Acknowledgements
The authors would like to thank the Mississippi Agriculture and Forestry Experiment Station, the Mississippi State University Extension Service, the Geosystems Research Institute for financial and technical support.
