Seeding the Future: AI's Pioneering Role in Modern Farming by Sébastien Boyer - Farmwise
💡 muffin.ai: the media to understand the challenges of AI and leverage them in our society and jobs — by a collective of French engineers & entrepreneurs.
Hello everyone!
We are excited to publish a second deep dive into the fascinating world of artificial intelligence with visionaries. 🙌
Your voice matters a great deal to us. Share your thoughts, questions, and suggestions with us. 🙏
Today we are sharing our conversation with Sébastien Boyer, co-founder & chairman at Farmwise, an AgTech company leveraging AI and robotics to help farmers profitably use more sustainable practices.
Seb Boyer is also a tech investor and his impact has been recognized in media such as Bloomberg, The Economist, the TIME magazine and the Financial Times. He was named one of the 35 Innovators Under 35 for Europe from MIT Technology Review in 2018, featured in the Forbes 30 Under 30 list in 2019, received the Young Entrepreneur of the year 2019 award by the French-American Business Association and was part of the list of the 40 under 40 leaders in the US Fruit and Vegetable industry.
🧁 We love Seb’s insights on agriculture industry and how AI help our farmers feed the world more efficiently and more sustainably !
Thanks Seb for the interview, it was a real pleasure to connect 🙂.
On the menu today:
🚜 A retrospective on the agriculture industry
⚙️ The key role of hardware to collect accurate data in the fields
💎 The strategy of a AI centered-product pays off
💡 One single ML model to detect all species
👀 Some sources recommended by Seb
If anyone shared this newsletter and you’re willing to subscribe, it’s here:
🧁 MF : What are the major problems facing the agriculture industry today? Why do you think AI can have a major positive impact on this industry?
👉 Sébastien :
Farmers in all regions of the world are facing three big challenges. First, labor shortages because lower skilled people have better and better opportunities in less difficult industries. Second, chemical impact on health and environment: the overuse of chemicals that has started in the 1970s is now proven to have a significant negative effect on farm workers and consumer health as well as a major negative effect it has on biodiversity around the fields and in agricultural regions at large. Third, global climate change imposes more volatile and tougher weather conditions on farmers who therefore need to adapt faster than before. These trends are concretely affecting farmers today in most regions of the world (from North America and Europe to Asia).
People often don’t realize it but agriculture is the first major technology industry. Over the past 100 years, new technologies have dramatically changed for the better the way we produce food. In 1900, each agricultural worker fed herself and 1.5 other people on average (40% of the population therefore worked on farms and food processing plants). In 2020, each agricultural worker feeds herself and 50 other people (and therefore less than 2% of the population needs work in the food production industry). The same goes for most other productivity metrics (Corn field yield x6 from 1900 to 2020, Water used to irrigate reduced by 40% for the same yield, milk production per cow x4).
This has been made possible by four major technological revolution: first the automation of “traction” from horses and cattle to the “tractors” in the first few decades of the twentieth century. Then soon after toward the middle of the 20th century, the Haber-Bosch process made it possible to create amonia (the main fertilizer plant need to grow bigger) literally out of thin air instead of mining it out of Chile and shipping it throughout the world. Then the in 1970’s GMOs were invented and became popular when used along side more powerful herbicides. This drastically increased yields almost overnight and allowed the production of cereals to grow and to become more cost effective. In the 1990s and 2000s the GPS was popularized and made it easier to plant with equal spacing and to have consistency in all of the other tasks required on each field.
Artificial Intelligence, combined with robotics, offer the ability for machines to automate tasks that were difficult to automate before (helping with labor shortage), offers the ability to “personalize” the treatment each plant receives (”plantalize”?) by measuring exactly what each of them needs cost-effectively, and offers the ability for farmers to better understand and therefore predict what actions will perform best under what conditions (helping them adapt to volatile weather conditions). These are theoretical reasons why I am excited to see more and more AI and robotics on farms. Very practically, I have seen it being successfully deployed then used at scale by farmers and I found this very exciting.
🧁 MF : At Farmwise you use both AI software and robotics Hardware. Why do you have to use both? What is the most precious part? The most difficult?
👉 Sébastien :
When we starting the company both my cofounder and I had a strong bias toward trying to build “software only” solutions. Neither of us had built hardware before (we are both math-physics-CS engineers trained between Ecole Polytechnique in France and MIT-Stanford in the US). So naturally this is where we started. We built prototype apps to capture information from the field and organize it better for farmers. We obviously thought about using drones to get precise images of plants, hoping that if Farmers had access to more information they would be able to make better decisions and therefore solve some of the problems they are facing.
We quickly realize that the bottleneck preventing these farmers from achieving higher efficiency (cost, labor and chemical efficiency) was neither information nor analytics. It was the physical infrastructure on the field. When you go on a field today it doesn’t take long to notice that none of the equipment have the mechanical precision to differentiate what they do. So whatever the level of information you can provide them with, they will be limited by the tools they have to act on this information. This is why we decided to have a combined approach, leveraging both AI models in software and precision robotics in hardware. This allowed us to build a full solution for them.
Hardware is incredibly hard but sometimes it is necessary. Most software-only companies in AgTech have either failed or are in trouble. Hardware is difficult because the iteration cycles are much much longer and more expensive. Mistakes are costly. Scaling is costly and slow. I would say that in the example of Farmwise at least, this is part of our barrier to entry making what we have built, and the learning we got along the way, very valuable.
🧁 MF : Your systems are working on 10s of thousands of acres of vegetable crops each year in the US now. Can you share real-life learnings you had when deploying an AI-centered product at scale? What surprised you? How do you make sure you achieve the reliability and predictability that so many AI systems struggle to achieve?
👉 Sébastien :
One of the first things we had to do was around the data. Because there is no publicly available datasets of plants. And because the quality of the data matters so much to our use-case. We had to build our dataset ourselves. To do so we built a 4-wheel pods and we used it on as many fields as possible to accumulate enough data (with the right quantity, quality and diversity) in order for our Deep Learning engineers and scientists to build predictive models that match the customer requirements for quality on the field. One surprises we got was how little images from a single field is necessary to get most of the learnings this field can provide (by measuring the incremental accuracy you get as you add more images from the same field you can decide how many is enough per field). It is smaller than I expected. What matters more is the number of different fields (different crops, stages, lights, soil, weed species, debris type) you are able to add to your dataset.
Another learning these AI-based products are never “done”. They stay alive and evolving forever. The farm practices change regularly, the predominant weed species change regularly, just to name two variables. This implies that the datasets and the models need to continuously evolve. One of the critical assets we have at Farmwise is our software infrastructure. After years of investment and continuous improvement on it, it allows us to iterate faster and more cost effectively on new data to improve models. The improvement of our ML pipeline (from data capture to model deployment) is more critical in my opinion than any of our ML models. Focusing on solving this “meta” problem (building a system that efficiently solves the customer problem) is a strategy that I am very proud we chose to have. This allows us not only to iterate on new data but also on new ideas coming from the cutting edge ideas coming from academic research in Deep Learning.
🧁 MF : For the more tech savvy of our readers, do you have key learnings to share about ML models, data or systems that could be useful for others trying to implement solutions in the real world?
👉 Sébastien :
Sure thing.
I have talked about the quantity versus quality / diversity trade-off in building your datasets. This is pretty obvious I think for people building ML systems in production. Another remark on building real-life dataset is that the consistency of the hardware matter quite a bit. You want to make sure you have your data-acquisition hardware design close to locked in before spending a lot of money building a large datasets. Small biases in the data coming from the data-acquisition hardware can drastically reduce the value of the data. We had to rebuild our datasets a few years ago because we had made too drastic of a change in the camera system we used to acquire the data.
Another interesting finding is that we were able to increase the accuracy of our computer vision models (that detect plant species in real-time on our machines) by using simulated data (in our case semi-simulated). By patching images one onto another, we were able to create fake-instances of fields and increase the diversity of our datasets. For instance, we patched specific weed species onto images of fields that didn’t have them in order to “simulate” future fields that might contain them. This allowed us to increase the robustness of our algorithms in the real world where the diversity can often take you by surprise.
One other finding was that over time we decided to implement a single model to detect all species as opposed to a model per crop species. This is less obvious than it sounds because in our use-case, we know in advance what type of crop is on the field (farmers know what they planted, they just don’t know the weed species that are growing next to them). So we started by building a specific crop versus weed model for each crop type. By removing the uncertainty of the crop type, we were able to achieve better results with crop-specific detection models. Over time, as we accumulated more data and as we iterated on the structure of the neural networks themselves, we found a way to build a hybrid model that essentially is trained to first detect what crop type is present on the field and then work under this assumption to discriminate between this crop type and the weed species. This increased the accuracy of our detection further.
There remain many challenges ahead of us. Particularly, we are working to handle better use cases where plants are more occluded, to run our inference even faster, and to be continue to adapt to the diversity that we encounter as we expand geographically to more regions of the world. There are a ton of challenges to solve and this, along with the real-life impact of seeing our ML algorithms and our robotics systems helping farmers everyday, is what makes working on these problems exiting.
🧁 MF : thanks Sébastien for sharing, can you leave us the sources you recommend if we want to go further on the subject
👉 Sébastien :
The Alchemy of Air by Thomas Hager. On the invention of the Haber-Bosch process and how this saved the world. It is a great mix of adventure, science, war, and scientific discovery. One of my favorite books.
Seeds of Science by Mark Lynas. On GMOs without the emotions.
Creativity Inc: By the founder of PIXAR. On how to build and manage creative teams that continuously innovate.
Enjoy your week !
— muffin.ai team
If you enjoyed the edition, consider clicking on the ❤️ button and leaving a comment so that more people can discover muffin.ai on Substack 🙏