Getting Into the Weeds: the AI data dilemma

William Gibson’s quote about the future being here but not well distributed applies to weed management technology. Despite 'see & spray’ variable rate spraying only starting to gain popularity, the underlying technology has been around for a long time.

The future is already here – it's just not evenly distributed.
William Gibson

Early versions used light filtering and reflectance to detect weeds. But the demand for selective spraying has increased, requiring faster and more accurate capabilities. Modern approaches use low-cost image sensors with machine learning and artificial intelligence. These models are accurate and can handle complex scenarios, moving from simple green-on-brown to green-on-green detection.

The models’ accuracy has improved due to commoditized hardware and software, large open source datasets for training, and increased use of these models in production, allowing for better fine-tuning for specific crop, weed species, and treatment scenarios.

In our latest podcast, we dig into the status quo of non-chemical weed management technologies and business models. In this post, I want to double click on the role of training data in fueling these systems.

Is the move from open to proprietary inevitable?

Increased adoption of precision weed management technologies is highlighting tension about who owns and operates training data sets and large language models (LLMs) in agriculture. The foundation for these tools is open source data. Vast, open image data sets have been used to create models that excel at weed detection. 

But as the commercial market has evolved, it has moved into closed, proprietary environments with unclear or restricted sharing rules. This evolution is creating a conundrum for farmers, advisors, operators, and researchers. 

How much benefit should go to commercial entities using open systems? What are the data sharing rights and responsibilities for commercial operators leveraging open systems during training and fine-tuning? What are the roles for research and industry, and how do we ensure value flows both ways?

Better data sharing regimes will unlock farmer involvement

Right now, it’s clear that there’s a strong value proposition for adopting precision weed management, but it’s unclear which models (LLMs) are better or how the market will evolve. This may mean that farmers using one system today will switch to another in the future (e.g., because it’s cheaper, better, etc.). 

The concern is that, without more clear sharing regimes, their on-farm data may be locked away and not available for training alternative systems. This is a bad outcome for all, as it will create barriers to adoption and slow our collective learning rate. 

Fortunately, companies have a stronger incentive now than ever to improve data sharing because having farmers and agronomists involved in labeling and annotation for training and fine-tuning models will make their technologies better, faster. 

Relationships and business models for ag retail will evolve

Data access and ownership aren’t the only big issues. Enhanced weed detection opens up various weed control methods. Selective spraying is just the beginning; targeted management will include optical, electrical, mechanical, and other approaches, reducing farm inputs and driving economic changes. 

For example, highly tailored LLMs, operating at farm scale and fed with localized data from more sources, will create new business opportunities for retailers and advisors. If advisors become custodians of these farm data lakes, challenges around ownership, access, and commercial benefit could open up, but may also have farmer-friendly solutions. In the same way that a farmer trusts (and pays) their account to collect and keep sensitive financial information, advisors may become responsible for collecting, retaining, and tuning highly specialized LLMs on behalf of farmers.

The path ahead

AI will continue to speed up acceptance and adoption of new weed control methods that have significant economic and environmental benefits. However, balancing commercial development, industry benefit, and farmer protections is tricky. Finding new ways to support adoption and business models will present challenges and opportunities for farmers, their advisors, and innovative solution developers.

No items found.

Want more content like this? Sign up for our weekly insights.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

Key takeaways

  • Balancing the need for open systems with private innovation
  • Data use in training and fine-tuning of LLMs
  • New roles for advisors as AI & data use scale

Get this report