Key Takeaways

• Innovative new sprayers use machine vision and artificial intelligence to identify specific types of weeds and selectively spray them with herbicide.
  
  
• Research suggests that this new technology can reduce chemical use for certain applications by 75% or more.
  
  
• Turf safety can be improved with this technology and the higher precision of these sprayers may increase the range of products that can be used for various applications.
  
  
• This technology has the potential to be applied to diseases, insects and other pests as well. Further research and development will continue to expand the possibilities.
   

Developing weed-recognition models and training machine vision sprayers

This is a complex topic that could fill multiple textbooks, but it’s worth covering the basics of how machine learning models recognize weeds and, once again, defining a few terms to understand what’s happening behind the scenes. “Model” and “algorithm” are sometimes used interchangeably, but in essence they are two different things. In general, a model is a one-time statistical prediction of whether some section of turf is a weed or not based on data analysis of sensor input. An algorithm is an automated process to take the first model and incorporate feedback to make a better model. Then you can test that new model and incorporate even more feedback to improve the model further. This is the learning portion of machine learning. It's called machine learning because there are ways to incorporate this feedback automatically.

Deep learning is an advanced subset of machine learning within AI. It is built on artificial neural networks (ANNs), which are mathematical models inspired by the way neurons connect and process information in the human brain. Convolutional neural networks (CNNs) are a specialized type of ANN designed to interpret images by recognizing spatial patterns, allowing them to automatically learn the visual features that distinguish one object from another – i.e., a weed from turfgrass. Machine vision sprayers use various real-time, multi-object detection models and algorithms – with one of the most common being different versions of the You Only Look Once (YOLO) algorithm. Another example is a Mask R-CNN model specifically designed to detect nutsedge in bermudagrass (Xie et al., 2021).

As mentioned earlier, machine vision refers to the process of turning images into a digital signal that is sent to an onboard computer to be used for automated differentiation between weeds and desirable grass using a pre-trained AI model. It only takes a fraction of a second between the images being taken and the model determining if a weed is present and activating the solenoid to fire the appropriate nozzle(s). The models are pre-trained by scientists and developers using thousands of images of weeds to help the model “learn” what a weed looks like. Most modern weed-recognition models rely on deep learning, but simpler green-on-brown models use less-complex image processing and could be used to treat weeds growing in dormant warm-season turfgrass settings.

Published research results about how this technology performs in turfgrass applications are still limited and the development of these machine learning models is ongoing at universities and companies across the globe. This includes work on expanding the number of weed species that can be identified, faster and more-consistent recognition, and refining the ability of this technology to be used on other pests, like diseases.

Weed Control and Herbicide Use Compared to Broadcast Spray Applications

Research in agricultural and turfgrass settings show weed control levels to be comparable with conventional broadcast applications – but machine vision sprayers require much less herbicide to deliver that level of control. Conventional sprayers rely on consistent and uniform coverage of the entire turf surface to deliver an effective dose of herbicide to the weeds. Machine vision and AI capabilities enable this next generation of golf course sprayers to use many more individually controlled nozzles to achieve extremely high precision. The newly released Ecorobotix sprayer named ALBA uses 108 nozzles on two booms that are spaced 0.8 inch apart, allowing spray to be applied to an area as small as 1.2 by 1.2 inches. A study conducted in China with a different machine vision sprayer prototype resulted in the same level of weed control in dormant bermudagrass turf as with a conventional sprayer (Jin et al., 2023). The study results also showed a linear response in herbicide savings based on level of weed infestation – i.e., less weeds equals more savings.

Although research in agricultural settings shows consistent and significant reductions in herbicide use, peer-reviewed research results on herbicide reduction in turfgrass are limited. Work at the University of Florida suggests about 80% savings in herbicide use is possible through machine vision technology (Petelewicz et al., 2024). Of course, herbicide savings depend on a range of factors, most notably the level of weed infestation. More weeds mean less savings as a highly precise application becomes increasingly similar in coverage to a broadcast application.

The research conducted at the University of Florida found that herbicide savings also depended on which machine learning algorithm or model is used, along with nozzle spacing. Optimizing herbicide savings and weed control efficacy with machine vision sprayers takes the right marriage of machine learning, nozzle setup, cameras, herbicide and target weeds – among other factors.

For golf course superintendents, actual savings realized will be highly dependent on several key factors. If excellent weed control programs have been in place for years and weed populations are low, savings could be significant compared to broadcast applications. If weeds are everywhere, herbicide savings will not be as great. The type of weeds also matters because some weeds are easier to differentiate from the desirable turfgrass than others. Expectations for the level of weed control is also a factor in potential savings.

One final but important point is that the rate of false hits or missed weeds depends on the algorithm and many other factors. Research is limited, but in a study at Texas A&M University, researchers achieved a false-negative (missed weed) rate of 0.4% (Xie et al., 2021) when spraying nutsedge in bermudagrass. That means almost no nutsedge plants were missed in the test area. Work at Virginia Tech using an earlier prototype of the ALBA sprayer showed 96% accuracy when targeting broadleaf weeds in bermudagrass fairways (Jakhar et al., 2025). However, a balance needs to be struck between avoiding too many false positives (seeing a weed that isn’t there) while also limiting false negatives (missing a weed that is there) when training machine learning models and ground-truthing the results.

Dr. Shawn Askew is a professor and weed scientist at Virginia Tech and led early research on machine vision sprayers in turf. He says that most superintendents he talks to want to use this technology to gain the upper hand on resistant or troublesome weeds. For weeds that are hard to differentiate from turf, like annual bluegrass, this will not happen after the first treatment but is possible with a few treatments. 

Benefits of AI-Powered, Machine Vision Sprayers

Reduced herbicide use

The reduced herbicide use that may be possible with this technology has a lot of benefits, with lower pesticide costs high on the list. For example, goosegrass is an increasingly problematic weed and topramezone (Pylex) is a common and effective herbicide option. A superintendent spraying 30 acres of moderately infested bermudagrass fairway turf could require up to 24 ounces of topramezone for a single application. If you estimate a 75% reduction in herbicide by using a machine vision sprayer, a savings of over $1,600 in product could be achieved from one application at today’s prices.

Another example provided by Dr. Askew would be using methiozolin (PoaCure) in a much broader area than just putting greens. Methiozolin costs about $2,500 per acre per year on greens and twice that on fairways. Thirty acres of fairways treated with a year-long methiozolin program would cost $150,000 for broadcast treatment, but only $15,000 using a machine vision sprayer – when treating turf with a 10% population of annual bluegrass. According to Dr. Askew, only a few golf courses globally have the budget to use methiozolin course-wide due to cost, but the savings and weed control benefits on this product alone could justify the investment in a machine vision sprayer for some courses.

Weed mapping and product usage tools

Significant savings may also be realized through advancement of weed mapping and predictive tools. Over time, the ability of machine vision sprayers to map and analyze weed distribution across the golf course during postemergence applications may lead to more-targeted (i.e., better) preemergence applications, which should reduce overall weed pressure year after year. Improved mapping may also improve the efficiency of selective postemergence applications as target areas are better understood. This can lead to time and product savings. Precise tracking of how much product will be needed for future applications and how much was used is also possible, which makes purchasing and mixing more accurate and reduces the risk of waste.

Less risk of turf injury

Improved turf safety is one of the biggest potential benefits of machine vision sprayers. Since you are (theoretically) only spraying the weeds, there is less risk of injury to the desired turfgrass. If spraying precision is high enough, it also opens the possibility of using higher rates, new combinations and/or nonselective herbicides – even on actively growing turf – to cut costs or combat resistance issues.

This technology can also increase investment in and success of organic weed control programs, something Dr. Askew is quick to point out. “Some courses require investment in organic weed control and results are often poor because organic products control weeds less effectively, a trait exacerbated by the fact that most are nonselective.” He continued, “When used at rates that effectively control weeds, the products kill turf too – and they are also very expensive.” Organic products also exhibit alternate modes of herbicide action for targeting resistant weeds. So for turfgrass managers that want to or have to use organic products, machine vision sprayers could be a powerful tool.

Regulatory and environmental benefits

Spraying less area in total to achieve the same results is desirable from the standpoint of environmental responsibility. This capability could also be pivotal for managing acreage limits on pesticides that are currently in place or may arise in the future. Older effective herbicides that cannot be used for broadcast applications due to label restrictions could rejoin the conversation because the low amount of overall product used by machine vision sprayers will meet “spot treatment” thresholds.

A significant factor driving innovation in sprayer technology is herbicide resistance. The Herbicide Resistance Action Committee (HRAC) reports that 273 different weed species are confirmed to be resistant to at least one herbicide mode of action or target site. By increasing sprayer precision and reducing the potential of turf injury from off-target application, a greater range of herbicide options becomes available.

The current and future regulatory landscape is also evolving. Recent court rulings regarding the Endangered Species Act could potentially have a big impact on where and how products can be used going forward.  

Challenges and What the Future Holds

Machine learning algorithms and models have proven quite effective at weed identification; however, scenarios like identifying annual bluegrass in a closely mown creeping bentgrass fairway can be tricky, and algorithms are still being refined to handle challenging identifications. Weeds can also have different morphology based on geography. Plants such as spotted spurge, clover or annual bluegrass can adapt to a wide range of mowing heights and may look different to a machine vision sprayer’s cameras based on management practices, climate and other local factors.

Additional research is ongoing and includes work led by Dr. Navdeep Godara, assistant professor in turf and forage at NC State University, exploring ways to quantify how weed density and spatial distribution influence herbicide savings from targeted postemergence applications. For instance, savings are expected to differ between species such as annual bluegrass, which exhibits a bunch-type growth habit and sedges, which form dense patches. These contrasting growth habits may affect nozzle activation frequency during targeted applications, with individually occurring weeds likely triggering more spray events than creeping or patch-forming species. Dr. Godara and his team’s goal is to provide superintendents with ways to maximize herbicide savings on problematic weeds, particularly when using high-cost chemistries like methiozolin. His research will also explore incorporating residual herbicides into targeted applications to reduce weed seed bank from historical infestations to maximize the potential of this technology.

There is also a good deal of work happening with drone sprayers using machine learning algorithms. Research in the drone-spraying space is ongoing, both with machine learning and aerial imagery to identify, spray and map weeds. Some intrepid weed scientists are even working on developing machine learning models that can identify and spray only herbicide-susceptible plants, and won’t spray herbicide-resistant plants (Jin et al., 2022).

Dr. Askew at Virginia Tech doesn’t see machine vision sprayers replacing what superintendents currently use, but rather as a tool to supplement and improve what they are able to do. “Current machine vision sprayers are smaller than a conventional sprayer and take more time to cover the course,” said Askew. “Superintendents would not invest in a technology that takes more time if it did not give them a clear advantage in return. That advantage is more power in killing difficult weeds.” Superintendents that battle herbicide-resistant annual bluegrass may not be that attracted to saving 70% on a relatively small herbicide budget, but greatly improving control of annual bluegrass, that would be of great interest.

In terms of the newly released ALBA sprayer, Ecorobotix plans to continuously incorporate new features such as RTK technology for more-precise GPS positioning. Currently, users can select a variety of weed options to target, but the company is also working on developing disease-identification machine vision models and plans to integrate the capability to make fungicide applications by the end of 2027. The sprayer, which costs around $125,000, can also be used like a conventional sprayer to make broadcast applications when needed, without the machine vision turned on. Competing products will also surely become available in the years to come.

We can’t say for certain what the future will hold, but machine vision spraying is poised to play a key role in the ever-advancing world of golf course maintenance.

References

Avent, T.H., Norsworthy, J.K., Patzoldt, W.L., Schwartz-Lazaro, L.M., Houston, M.M., Butts, T.R., & Vazquez, A.R. (2024). Comparing herbicide application methods with See & Spray^TM technology in soybean. Weed Technology, 38, e74. doi:10.1017/wet.2024.70

Jakhar, A., Askew, S.D., Romero, J.R., Crawford, S.E., & Godara, N. (2025). Economics, pesticide load, and turfgrass quality following organic or synthetic herbicides applied via machine-vision sprayer [Abstract]. CANVAS 2025, Salt Lake City, Utah. https://scisoc.confex.com/scisoc/2025am/meetingapp.cgi/Paper/168342

Jin, X., Bagavathiannan, M., Maity, A., Chen, Y., & Yu, J. (2022). Deep learning for detecting herbicide weed control spectrum in turfgrass. Plant Methods, 18(1), 94.

Jin, X., Liu, T., Yang, Z., Xie, J., Bagavathiannan, M., Hong, X., Zhengwei, X., Xin, C., Jialin, Y., & Chen, Y. (2023). Precision weed control using a smart sprayer in dormant bermudagrass turf. Crop Protection, 172, 106302.

Petelewicz, P., Zhou, Q., Schiavon, M., MacDonald, G.E., Schumann, A.W., & Boyd, N. S. (2024). Simulation-based nozzle density optimization for maximized efficacy of a machine vision-based weed control system for applications in turfgrass settings. Weed Technology, 38, e25. doi:10.1017/wet.2024.7

Xie, S., Hu, C., Bagavathiannan, M., & Song, D. (2021). Toward robotic weed control: detection of nutsedge weed in bermudagrass turf using inaccurate and insufficient training data. IEEE Robotics and Automation Letters, 6(4), 7365-7372.