What does it take for AI in the field to actually help farmers and advisors make better decisions?
That was the question I brought to Green Minds Gather, where Rasmus Nyholm Jørgensen presented OdinsAEye in the START DIGITAL session “The Connected Agri-Food System: A Holistic View of Digital Transformation”. His talk was titled “Digitalization and AI, challenges and prospects (cases)”. Mariesminde was the main practical case, supported by experiences from RoboWeedSupport, RoboWeedMaps and Agrointelli.
The central message was simple:
Recognizing plants is one step in a longer workflow.
The value comes when reliable observations reach farmers and advisors in a form that supports decisions, practical action and learning afterwards.
Why collect images while you are already spraying?
Collecting images during spraying can still provide useful information. It can help the farmer assess whether the chosen treatment matched the observed weed population, investigate surviving plants, consider follow-up and improve planning for the next season.
This is why repeated observations during ordinary operations are an important part of OdinsAEye. The purpose goes beyond generating a prescription for the operation taking place right now. Observations support reflection throughout the season.
What does Mariesminde show us?
Mariesminde demonstrates two complementary ways of collecting data:
- A dedicated survey: a commercial RoboWeedMaps ATV survey covering approximately 1,100 hectares and roughly 450,000 images. This gives a stronger spatial basis for mapping.
- Cameras on the tractor: later, tractor-mounted prototype observations collected during existing operations. These can support monitoring through the season.
Both routes have a role, but they are not the same. Results from tractor-mounted cameras must keep the information about where and when each image was collected. And tramline observations cannot establish complete field coverage.
The overview we demonstrated combines species information, density and occurrence across the sampled images, with separate summaries for headlands and the main field. Crop history adds context. Together, this helps advisors identify the fields that need attention and decide where an inspection or additional mapping would be worthwhile.
Can AI help turn data into practical questions?
In an exploratory planning example, we grouped fields by weed problem and took the farmer’s actual sprayer and logistics into account. Examples included hemp-nettle and thistle, cornflower findings and substantial grass populations.
To be clear: these are prototypes for presenting information and comparing possible strategies. The proposed groups are not validated treatment prescriptions. Agronomic review remains essential, particularly when language models help interpret results or suggest options.
Who stays in charge?
The presentation described a direction consistent with human-centred Agriculture 5.0: AI should strengthen the farmer’s and advisor’s capabilities while allowing them to exercise judgement and stay involved.
The practical ambition is a cycle:
Observation → professional assessment → chosen plan → actual execution → follow-up
The system should help users understand why something worked or failed.
Our “glass box” example makes this concrete. If a user questions a cornflower result, they can inspect the underlying plant images. This lets them verify the evidence, although it does not explain every internal step of the model.
What are we delivering, and what can we claim?
We kept the focus on the spring 2027 MVP: a weed-population overview per field and date, initially through PerPlant. Testing with existing data in December and field operation in March were presented as targets. LLM-supported interpretation and more advanced planning remain development directions.
Earlier herbicide-saving figures were discussed as motivation, based on previous work and modelling. They are not measured OdinsAEye field outcomes. Demonstrating the benefits requires comparison with ordinary farming practice, including weed control, chemical use, workload, costs and crop performance.
One more development is worth mentioning: the emerging detection of insects and other small animals in the images. This is an additional research opportunity, but its agronomic and ecological interpretation needs further validation.
The take-home message
OdinsAEye’s contribution is to connect observations, expertise and practical decisions, helping farmers and advisors learn from what actually happens in their fields.
From images to data. From data to knowledge. From knowledge to better decisions in the field.
A big thank you to the organizers of Green Minds Gather and the START DIGITAL session for the invitation, to Mariesminde for sharing their fields with us and to Novo Nordisk Foundation and Innovation Fund Denmark for the funding.
Want to talk about connected agriculture, human-centred AI or a collaboration? Get in touch and let’s talk.