In my LinkedIn and other social media feeds over the (northern hemisphere) summer, it seems there’s been a consistent stream of open-source project announcements. Perhaps it’s due to the frequency illusion (or Baader–Meinhof phenomenon), where things you are aware of you tend to notice, though I’m fairly sure my interest in open-source agtech never waned. So, in this newsletter I want to highlight a few interest-piquing projects from researchers, startups and established players I’ve seen of late. Besides the vibe, if you look at creation of new GitHub projects (repositories) tagged with agriculture as one indicator, the stats show we’re on an exponential growth curve. But I think there is more to it than that.

So, let’s dive in.

The past few months have been an incredible whirlwind. Besides some travel for my brother’s wedding and the Elevate conference at Grand Farm in Fargo, ND, we have finally launched Noktura ApS, to commercialise the OpenWeedLocator and build an accompanying DIY vision AI ecosystem, we’re currently calling Canopy. The vision is that anyone can take their AI ideas into the field and build it into a machine they need – both from a hardware and machine learning perspective. To be successful this needs to be open.

This has been a very long-term goal of mine – to test the idea that open-source agtech is not only a good idea but the foundation and critical element of a successful and sustainable company in agriculture. While many people see it as a risk, to me it seems a lot closer to a competitive edge. I’m incredibly grateful to my co-founder Peter Darket for his belief in this idea and jumping headfirst into agriculture and technology from a management consultant and engineering background. Without Peter, this would still be a pipe dream. To get us off the ground, the support from the University of Copenhagen innovation group, Lighthouse, has been fantastic.

In the few months since founding, we’ve built prototypes, sold just under 50 OWL preorders in the Founding Batch and started to get our heads around the challenge of procurement. To the folks around the world that have put their trust in us, we’re very appreciative of the support at such an early stage, and excited to start shipping your orders in December. Talking to the farmers, researchers and people interested in the OWL3.0 as part of this process, it’s only become more apparent that the future of agtech must be open-source and led by those actually using the tech. The current closed approach does not scale and misses out on a huge segment of the industry (as discussed in the previous OSA).

I should probably share the OWL3.0 launch story in its own newsletter sometime, but if you’re interested in the vision behind Noktura and supporting farmer-led innovation resonates with you, we’d love to chat.

Besides our work on Noktura, what’s been exciting is the breadth of open-source projects being worked on and announced in agtech recently. In this article I wanted to highlight three different projects:

One of the developers of OPheno at BASF, Heiko Hopke Kromminga, has been kind enough to share some perspective on the background and decision making.

But before we dive into these projects, I just wanted to set the scene with some high-level stats on open-source work in this field.

If we go back in time 10 years to 2016, there were just 25 new repositories from that year tagged on GitHub with ‘agriculture’. For those not familiar with GitHub or a repository, GitHub is the go-to online platform for hosting your code, enabling version control and the easier management of public contributions. A repository is just a project within your account for that set of files/folders to exist. It’s an easy way to share code and monitor changes over time, so a decent though far from perfect method to sample open-source projects in this space.

To put this era in context, 2016 was the year YOLO was launched (the most common open-source GoG weed detection framework), it was a time before deep learning was used to any great extent in agriculture and a time when writing deep learning code required considerable skill sets and years of training, i.e. before generative AI. Fast forward to 2024, and there were 295 agricultural repositories released, 2025 saw 660 and 2026 even with three months remaining has seen 2,112 unique repositories tagged with the keyword agriculture (see the first chart above).

Perhaps unsurprisingly, what is being built has shifted (proportionally) considerably. If we randomly sample 60 repositories per year (where possible), the hardware and robotics projects fell from making up 27% in 2019 to just 5% so far in 2026. Web apps have risen to about half of all new repositories, comprising mostly crop/agronomy recommendation and disease detection apps.

Finally, if we look at GitHub stars on each project, as a quick measure of popularity and/or usage over time, there doesn't seem to be any clear growth in projects with 10+ stars. Since 2016, the number of repositories from each year to reach this figure has stayed consistent between 15 and 25. However, we need to wait some years to see how the current surge plays out.

Bringing it together, these trends indicate that adoption and real-world traction remain the biggest barriers. The cost of code development has effectively dropped to close to zero, but as with the ‘before times’, the existence of agtech code never implied successful adoption. The opportunity now, is that we can try so many different things for much lower cost before one sticks.

Some method notes: This is a very simplistic method where I just pulled all repositories tagged with the ‘agriculture’. Use of keywords like ‘weed’ and ‘crop’ bring a lot of false positives, so ‘agriculture’ felt like a nice middle ground. Some will clearly be missed in the counts, and it is likely that AI coding tools will do a better job of tagging repositories than humans from 2016. So, there are confounding factors here that are somewhat hard to control for. In saying that, it is coarse enough to offer a form of ‘temperature check’.

To showcase some of the interesting work going on in this space, I wanted to go into a bit more detail on a few different projects I’ve come across and share more of their story. Heiko Hopke Kromminga from xarvio (BASF) was kind enough to share some more background on OPheno, which I’ve included below. If you have a project and would like to provide some commentary on it, always just reach out and I am happy to accommodate.

OPheno is an open, community-driven space for sharing research-grade weed emergence and phenology data, hosted on Hugging Face. Anyone holding the rights to weed phenology data can submit it through a submission portal, and submissions are reviewed and curated into versioned public datasets under CC BY 4.0. Heiko is the Technical Product Owner for weed management decision support at xarvio (BASF). He grew up on a mixed dairy, poultry and arable farm and has an MSc in Agronomy focused on digital farming and machine learning. Here is his take on why BASF backed an open dataset.

“Weed management is getting harder, fewer effective chemistries, rising resistance, more movement to use less. I work on weed management decision support, and one thing is clear: timing is where the gains are. That means predicting weed phenology and applying herbicides when they are actually required.

That may sound simple but good quality weed data is really hard to get. Such a development is therefore tied to running intensive trials, resulting in longer development cycles and higher costs.

And yet quite some data already exists in one form or another. Most of it is from the past since running weed phenology trials is not "sexy" anymore. So it sits there locked and hidden.

We started thinking: how cool would it be if there were some initiative which unifies this data, lowering the barriers for educational modelling, enabling innovative algorithms and deepening our understanding of growth dynamics in agricultural fields?

The idea of OPheno was born: standardized, curated, easily accessible weed phenology data under CC BY 4.0. By now we have one dataset which focuses on weed phenology in untreated control plots. Though we don’t want to limit ourselves, in the future we expect more datasets as contributions come in. Building this didn't cost us much extra effort. We could reuse the data pipeline behind our commercial HEALTHY FIELDS for RITA offer in Japan, so the additional work to create OPheno was marginal. Going forward, most of the effort will go into data harmonization.

Xarvio (BASF) is not opening proprietary data here. What we contribute is the standardization layer for data that was already collected. In the short term this serves the research community more than it serves us: for product use, the data has to match specific target species and locations, and contributed data won’t always line up.

So our stake is longer-term and indirect. A larger and cleaner open pool will eventually lift everyone’s models, ours included. BASF is a research-driven company, so backing open-science infrastructure like this fits well. Is there a risk it never pays off for us? Sure. That’s a bet I’m willing to make if the trade-off is helping science and the community move forward.

Personally, I wanted the project as independent as possible so it outlasts the priorities of the company but there was also a design decision to make: a plain phenology dataset versus a full, ready-to-use one. That’s why we sacrificed some of the independence in return of environmental meta data. We use our pipeline to release such enriched versioned datasets. Modellers will hopefully thank us for that. The first version has only limited environmental data but the next one will extend it, stay tuned.

If you have any weed phenology data you hold the rights to, use the submission portal and we’ll integrate it. For the enthusiastic ones feel free to join our discord channel for updates and discussions.”

OPheno on Hugging Face

Weed-phenology dataset

DecimeterWeedMan is an open-source platform for ultra-high precision (<10cm) spot spraying. The work was published in Smart Agricultural Technology in July by a group from Denmark, primarily Aarhus University led by Rasmus Nyholm Jørgensen. It combines hardware, a vision model and a model of spray dynamics to improve these high precision systems. The boom they used is 1.25 m wide with 20 solenoid-controlled nozzles at 62.5 mm spacing. To detect weeds, a YOLOv11 model was trained on 14,390 images with 705,902 annotations across 46 EPPO-aligned weed and crop classes. What was cool for me to see was it exceeded 92.6% hit rate while also using the species-level detection to manage weed species diversity and spray only the most problematic, competitive plants. Like the OWL, the code, CAD files and configuration are all shared so you can build your own.

Paper (open access)

Code, models and CAD on Zenodo

AgML is a Python framework from the Plant AI and Biophysics Lab at UC Davis led by Mason Earles, which solves a somewhat similar problem to OPheno but for image data - standardised access to public agricultural image datasets. It provides tools for loaders for classification, detection and segmentation, benchmarks and pretrained models. Over 2026 the team has been moving the collection onto Hugging Face, where the AgML organisation now lists 499 datasets with a target of 1,000 by mid October. It also provides iNatAg, a curated set of 4.7 million images covering 2,959 crop and weed species drawn from iNaturalist. AgML has close to 300 GitHub stars and more than 134,000 downloads from PyPI.

Finding, cleaning and converting datasets is still one of the slowest parts of agricultural ML. A single loader and consistent annotation formats (COCO for detection, pixel masks for segmentation) removes much of that work. Props to Mason and the team for the vision and hard work getting it to where it is today.

GitHub

Datasets on Hugging Face

Dataset browser

NatAg paper

Clearly, generative AI is a major (only?) reason for the rapid, exponential rise in published projects. Projects like weed or disease detection web apps that would previously have been a solid semester’s worth of work for a student can now be achieved in an afternoon by effectively anyone with a laptop and an internet connection. GitHub is in a battle for its existence to manage the complexities of so many AI-driven interactions with their platform.

Yet that doesn't explain why they are published openly. Open-sourcing your software is still a decision with trade-offs. Each of those approx. 2,000 people who published their project on GitHub had to uncheck a ‘Private’ box. Perhaps it is that people value code so little these days, they see no downside it releasing it, instead only potential upside in sharing and showcasing what they built. For me, that is where this industry is headed. Code is no longer a moat, and holding it closed harms not only the end user (farmers in this case) but also the entire community. It would be fantastic to see an exponential rise in Github activity from the likes of John Deere, Case, some of the other major players and even startups out there.

There is some incredible work going on in the open-source space in agriculture that will contribute to faster innovation cycles. This work goes well beyond these three examples I mention here of course, and there are plenty more examples over on the OpenSourceAg list.

Putting this together, the biggest question for me is — does this growth mean we are turning a corner in the industry towards openness or does it just show the perceived value of generated code has declined so much that open-source is now the default? If the latter is the case, then does an open-source culture follow naturally? It’s a strange time ahead.

On a side not, on some recent ferry rides across the Baltic, I spent some time on an article covering appropriability and the case for open-by-default research funding in Australian agriculture. The article is for the Australian Farm Institute’s Farm Policy Journal along with Matthew Pryor and Sarah Nolet of Tenacious Ventures. I’ll write some commentary here about that idea once it’s published later in October.

Until next time!

Cheers,

Guy