254 – Farmer Research Networks and the Future of Participatory AI

[00:00:00] David: Hi, and welcome to the IDEMS Podcast. I’m David Stern, a founding director of IDEMS, and it’s my great pleasure to be here again with Rikin Gandhi from Digital Green. It’s a slightly different episode from the last one this time, because there is a real possibility that Digital Green is going to become part of the Global Collaboration of Resilient Food Systems as a grantee, and that will lead to opportunities for collaboration, co-development, and, yeah, Rikin I hope, welcome to the family, so to speak.

[00:00:37] Rikin: Great to be back on the platform with you, David, and so excited to be part of this platform with so many other amazing organizations and people.

[00:00:47] David: Thank you. And I guess part of what I was thinking now is to actually tell you a little bit about the history, and use this as an excuse to tell you a bit about CRFS’s history, but also to then discuss with you how the work that you are doing is so relevant to what we’ve been working towards for over a decade, and why I am really excited that you are joining the family, and the implications of what this could mean.

Because, as I’ve mentioned before, I’m really a fan of the approach you’ve taken in the way you are using AI to create this “extension assistant”. Is that what you’d call it?

[00:01:31] Rikin: Yeah, I mean, I think we’re trying to flip the model of agricultural extension, maybe in a similar way as what you all are thinking about on the research side, but for long extension has been very top down in one way, from research, to extension agents, to farmers. With Farmer.Chat, what we’re trying to see is how you can flip that paradigm so that farmers have greater agency and choice to share what matters most to them.

[00:01:58] David: Exactly, and this is why you are such a great fit potentially for the CRFS, because, for a decade now, trying to flip the research process so that farmers are at the center of what we call farmer research networks has been the work that CRFS has been deeply involved in. And there’ve been some really interesting successes.

So I’ll take you back in history a little bit, about a decade. The program had already embraced agroecology, and Rebecca Nelson was the academic director at the time, and there was this effort to have what they called large-end trials, where farmers were at the center of the research process and the trials were being done in such a way that they were on farmer’s field, but at scale.

And there were interesting other groups who were doing this. Within the CGIAR, one group I love is the group working on Tricots, which have a simplified approach. They have really good, rigorous research for these participatory breeding and other trials. But the problem that Rebecca was really interested in is exactly what you are describing: how do we get farmer agency up? How do we really put farmers’ agency at the center of this research process at scale?

And Rick Coe, who I’m sure you’ll meet and you’ll interact with as part of this, and you might have met already, was the person who, for me, articulated it really well. He articulated that traditionally agricultural research was about understanding the underlying processes. And to do that, you simplified the situation, you put it on station, you created a controlled environment so you could understand and study the processes. This sort of agricultural research dates back over a century and huge advances came from understanding how plants grow, and what’s happening within agriculture. But this was happening at a low level, at low scale, small numbers, and not understanding, not in a very participatory way, it was very much isolated and, if you want, very academic and scientific.

The next innovation which sort of happened, and which became very popular, was actually saying that we shouldn’t just be looking at the sort of on-station trials, we should be looking at participatory trials, working with farmers to understand the context, to understand things in context. And so he described how the levels of participation increased. He actually had this on the Y axis, and so you were still having relatively small numbers, but a high level of farmer engagement and participation, Participatory Action Research, these sorts of things, PAR.

And this became very popular, there’s a lot of research which happens in that way, and so farmers are very much at the center of that process. It’s very much a collaboration between the researcher and the farmer, but it remains at a relatively small scale.

Then, the next big advance was what people called precision agriculture, where you now get remote sensing data and you get data from sensors of all sorts at great scale. So you get lots of data, you are able to use AI and all sorts of other modern techniques to be able to analyze that data. But you are not really, it’s not very participatory, it’s very much just observing what’s happening. And so you now have the scale of being at large scale, but not high levels of participation.

And the hypothesis he put was that the farmer research networks could be an approach which could be high participation and large scale. And this could be getting a different type of data. And I loved that visual that sort of came out of a different way of doing agricultural research, where we’re able to answer different questions.

So this is the background from over a decade ago of what the thinking behind the farmer research networks should become. And in practice, of course, that implementation is really tricky. And so there have been many cases where this has reverted back to more participatory action research and people trying to do participatory action research, maybe at slightly bigger scales. And there have been a few efforts of it being more like remote sensing, precision agriculture, but slightly more participatory.

But getting to that place where it’s highly participatory and deeply large data, this is the hope that the sort of tools that you are building could actually enable us to engage in research in that way. And so maybe I’ll let you come in and see: does this resonate? And then we can go into a bit more detail.

[00:07:08] Rikin: It totally resonates. Thank you for sharing that history, David, really illuminating and amazing about the journey that you’ve all been in setting up these farmer research networks. And there’s such a parallel/intersection with what we’re trying to do or similarly trying to create farmer networks more on the information sharing side of things.

And we’ve been doing it historically, even pre Farmer.Chat and AI with videos produced by and for farmers so that peer farmers can learn from each other and where they can be seen as legitimate role models in their respective communities that other farmers can aspire to.

With Farmer.Chat and AI that unlock becomes even larger for these communities because now they can be just taking simple photographs or leaving voice notes, and they can be used for multiple purposes. One purpose is for sharing knowledge with their peers.

And in fact, we get 70% of the queries, not where farmers are articulating themselves, but where they’re learning from their proximate farmer peers where maybe a fellow farmer is experiencing a pest or disease issue, they’re nearby to them, and they wish to get that information surfaced. To say that, if some fellow farmer is experiencing some issue, then maybe I might too. And that they don’t need to kind of go through the whole process of articulating a query from scratch.

And similarly, for farmers who sometimes struggle with articulating even what some of these issues are, being able to take photographs and being able to classify them or contribute them to their peer groups, to researchers in proximity can be quite powerful to, again, flip the model of ag extension and research in the process.

The other thing I would mention is, for very long, people have been trying to create decision support models and tools for these communities, A) often separate from these communities themselves, and B), in a way that’s inaccessible for the communities themselves because they turn out to be very complex calculators that require a lot of academic training to perhaps interact with, whereas the farmer has much more practical training.

What we’ve been working on with some researchers has been how can you take unstructured types of data sets, like local language voice notes and photographs, and use those to serve as input to some of these models, both to refine them and localize them, when they’ve been created in a way that’s not tailored to a particular context.

But also so that the farmer can also just get greater access to these solutions that are like sitting on a shelf somewhere, and where with these ability of AI systems to take unstructured voice or images, parameterize that, and then interpret the response in a way that the farmer can take action upon it. Then we can make these models and systems much more useful to these communities and have feedback loops from the farmers themselves that can help to make these models a lot more contextualized and actually useful for the communities themselves.

[00:10:28] David: Absolutely, and this is where I think we have found there is this incredible overlap. What I’m gonna now tell you about is what do farmer research networks that are working well look like and how might they further enhance what you are doing.

So, what this idea has translated into in many different contexts is the idea of farmer experimentation at scale built by, co-designed with farmers, but in the way that there is real structure, good experimental design inserted into the networks. Now, the key here is that this means that experimental design is what is able to, in theory, identify causal links.

So, if you have unstructured data, you can hypothesize causation. But you need experimental data to actually be able to be confident, if it’s done well, that you are finding and you are getting scientific evidence of causation. And what the farmer research networks in different contexts have done is to find ways to really get far more experimentation out at scale in ways that these causal links can be studied.

The problem has been, how does the data come back in formats that the researchers can interact with, where you can actually get and understand the complexities of these causal pieces? Because they’re often very contextualized, they’re context specific, they work well in one case and not in another. And so there’s a lot of complexity to this.

So, I’ll give you two quick instances. One is from East Africa, the agroecology hub in Kenya based out of Manor house. They’ve developed an approach where they have a farmer training process where, through the networks of farmers in different ways, they identify problems which are common to many farmers, and they then bring those farmers together for a workshop. And the farmers selected across the different contexts are those who are most heavily affected by this problem, whatever it may be.

My favorite recent one was termites. Termites aren’t supposed to be a pest, but in western Kenya, they became a pest. And so they had a workshop where the farmers came together with researchers to discuss, well, what do we do about our termite problem? And the researcher, of course, does some research beforehand to understand, well, what could the nature of this problem be, what could some scientific solutions be, and so on.

But that is not presented to the farmers at the beginning. The nature of each day is that it starts with the farmers discussing, and the researcher can ask questions, they can engage in that discussion, but they can only bring information in in the afternoon. And so each day is structured as farmer discussion in the morning, and researcher input comes in in the afternoon.

And of course what’s amazing is that these farmers are really heavily affected by this problem and, by and large, anything the researcher had to say gets said by the farmers in the morning. So, in this particular termite case, the fact that termites shouldn’t be a pest, they’re actually a beneficial insect, came from the farmers, the farmers already knew this. And so if the termites are becoming a problem for crops, it’s because there’s a problem with the environment, there’s not a problem with termites.

And that was really fascinating, and that came out, was that the farmers already knew this between them. So there wasn’t much for the scientists to add and they already knew some of the different things that could be done. And so this then comes out as this really interesting discussion.

But at the end of the week worth of interactions – and this is all held in an agroecological farm where the farmers get introduced to other agroecological practices along the way, so they’re learning while they’re there on things alongside this specific problem that they’re trying to dig into, to experiment on together – by the end of the week, they’ve designed or they’ve co-designed an experiment, which might be different for different groups, but that the farmers are going to go back and do, not just themselves, but with many other people in their community. And so it tends to be a very simple experiment.

The biggest problem we’ve had related to this process is exactly the problem that you’ve solved. How do you actually collect useful data to come back in terms of photos and voice notes and things which farmers are able to and happy to do for the experimental protocol that they’re implementing? You know, when they’ve had to change it, why have they changed it?

These are the things which are really important. But getting that information back in a way that it’s analyzable is too much, because you’ve got hundreds and hundreds of farmers doing this. Listening to all those voice notes is too much for the research group. There’s a small number of researchers with lots of farmers, this becomes an impossible task unless we have tools, like the tools that you are building, to be able to be part of this process.

That’s one instance. And I just very quickly go into an instance in West Africa where they’ve gone a step further. This is a farmer federation in West Africa who used their first trials, which I was involved in over a decade ago, on human urine as a top-dressed fertilizer. And this became something where they got thousands of farmers trialing this in different ways in a very simple experimental design.

My small contribution to this was to simplify their experimental design and increase their plot sizes so that the results came out more clearly. That was a very small contribution in that particular case, but it meant that their results were mind blowing and really powerful.

And they now have got to the stage where they’ve built an app, which they use in the farmer federation and they collect the data. And now when other researchers want to come and experiment with them, they don’t come and design the baseline survey or anything like this, they come and they say, “well, this is the information we want, do you have it?”

And by and large, the farmer federation says “yes, that’s the sort of information we maintain on the farmers that we have, and so we can help you with that. You can use it for your sampling if you want”. Because they have information on their farmers and they can then say, “well, okay, we’d like to do this experiment”. And they then go to their farmers and say, “is this an experiment which is of interest to you?”

And if they get enough people who are interested in that experiment, then it becomes a collaboration. But the power dynamics have totally changed. This has gone to the extent that the farmer experimentation is no longer part of the research project, it’s part of the farmer federation activities, it’s built into and deeply ingrained in what the federation is doing.

And that means that the interactions with researchers have totally flipped. But, at the moment, and this is something I’ve been discussing with them for years, what they’ve done to do this is they’ve got very good quantitative data, which they’re getting out and they’re getting at scale. But I’ve been sort of saying, “what if we could add qualitative data so you could get richer data?”

And the point is that they’ve tried this, but there’s just too much data, there’s nothing they can do with it. And again, this is where the sort of advances you’ve been making really could transform this so that they could now get farmers in the different areas to be doing voice notes and photos so that the data they’re getting gets much richer and they are able to help researchers to now draw out much more subtle and interesting learnings from the interactions they’re having as part of their groups.

So here are two concrete examples where, you know, there’s already real work to build from. My hope is this is gonna be exciting to you now to be engaging with these sorts of partners. 

And I guess, maybe just take a step back, we have done episodes about this in the past. You know, the distinction between good data science, which is born out of data abundance, when you have lots of data, but you don’t tend to have experimental design, so you can’t do causation, and your traditional small scale research where you have good experimental design, which means you can get causation.

Here, there’s a possibility to do both, to have large scale data with rigorous experimental design, meaning we can actually investigate causation, but at a large scale with volumes of data, which are fascinating and so imaginative. This changes research. This is what I’m really excited about.

So anyway, keen to get your reaction on this.

[00:19:43] Rikin: I am super excited as well, David, that’s amazing and thank you for sharing those examples. I think there’s such parallels between the work that you’re doing with these farmer federations and then trying to see how these farmers own innovations that they’re making every single day on the front lines of climate change, how they’re using that info for themselves, and then how some of that can be shared on their own choice to particular researchers to inform their agendas.

And the parallel with our work is that even the videos that we were previously doing, they would feature local farmers as innovators, experimenting and trying out new practices. And there would be new applications of information that researchers had not come up with themselves, that fellow farmers would learn from and say that, oh, this is an interesting new approach that this peer of mine has applied, let me consider it too if he or she is someone similar to myself.

And now, as you said, with AI, the richness of the interactions become even more. We’ve created these non-monetary leaderboards for farmers to be seen as role models within their respective communities with the videos that they submit. And the first use of those videos is not for research, it’s for the farming communities themselves, just as you said, for the farmer federations themselves.

And then secondarily, if researchers want to be able to tap into this, yeah, let’s ask the farmers for permission to be able to gain access to that kind of data. And then with the power of AI, I can then be able to mine that data, parameterize it, and be able to make sense of it in a way that was much more impossible in the past where it had to be manual curation or annotation of every single person’s voice note or image, and then I’d be able to make sense of it.

Now, this is like a matter of minutes that I can be able to make sense of some of this. And of course it can then become a two-way exchange. Just as you also said that if these farmers are giving this kind of consent and interest in terms of partnering with other types of researchers and stuff. And then the researchers can also share what they’re also working on, what they’re learning about.

And you can actually then have a network. Because in a network, it’s not about bottom up only or top down only. It’s actually about an exchange in both directions. And if both parties are willing to exchange and share with each other, then now I have a powerful way that even these models and research agendas can be shaped based on what the farmers actually want and need. And the farmers too can also gain value and share amongst their respective community and network that they also most value and trust.

[00:22:35] David: And the thing which I love about what’s emerging from this is that there really is this need for this, it is not about the technology, it’s not just about the human infrastructure. It’s how the technology is enhancing the human interactions, it’s making possible human interactions that weren’t possible before.

The example, I think you’ve hinted on, which I think is so important in what I’ve seen, is that researchers are often constrained by what they think they can analyze. So their design is often really a highly constrained design, not because they’re not interested in more, but because they don’t have the capacity to do more.

Whereas these interactions are now changing, you know, these are proper synergistic interactions where the researcher can do more because of the infrastructure that they’re tapping into, and the farmers get more out of it because they’re able to get the research input where certain things are known from scientific processes.

But if you now have a context which emerges from the experimentation, people know what the result of the experimentation should be, but it doesn’t happen, you can now investigate. And I’ll give a very concrete example of this from Western Kenya.

There are soils in Western Kenya which have acidity and aluminum toxicity problems, which means that in those areas they don’t respond to fertilizer and to fertilization as you would have expected because of the soil properties which have emerged. And this is something which I know has happened in other parts of the world as well.

But imagine now you are part of a farmer research network where the advice, the scientific advice, which is coming says, well, if you do this, you’ll get these results, and you are able to say no, in our context, when we do that, we get different results. This is now an identification of a scientific question where we can actually get proper research happening to support the realities on the ground, because part of what we know is these systems are complex.

There’s a lot of complexity. Different contexts do behave differently because of differences in the soil composition, or whatever it may be. And being able to now draw that out in ways that when the researcher or the scientific knowledge is contradicted, now you’ve got the basis for a researcher to come in and actually investigate. But if it’s just a farmer saying, no, it doesn’t work for us, it’s difficult. But if you have it as part of these systems with good structures in place, it becomes really powerful.

The thing which I want to, again, draw out on this is this observation from the group I mentioned in Niger. Once they started doing these large scale trials and they said, actually, we’ve noticed something, the biggest factor determining the quality of the experimental data coming from the farmers is how long they’ve been doing experimentation.

Now wait a second, with hindsight, this is obvious. But can you tell me of a single researcher who includes in their experimental structures who they’re working with and their experience in working as part of these experimental processes in the design? I don’t know anyone, and I’m ashamed to say I never thought of it myself.

And yet it’s obvious. If you have farmers who are actually learning what the experimentation, and what the scientific process is bringing for them, then they get more engaged, they get interested, it is a useful process for them in their decision making. And this is something where once you start doing this at scale, it comes out naturally as something which farmers understand and they have the capability.

We’ve built this in Kenya for those trainings. There are simple elements of the scientific process, which are now coming out as part of what the scientist brings in the afternoon so that people understand, well, if you’re going to do a comparison, why is it useful to have rigorous comparisons, what does this give you, and so on? And it’s something which is so appreciated as part of that collaboration.

Anyway, it’s a fascinating subject, we could talk about this for hours. 

[00:27:20] Rikin: But it is incredible. There’s such a power when you are able to kind of see this very large and diverse community of individual farmers who can have greater voice and agency. Certainly I see that in like even our more focused work on extension, which for long has been a more blanket approach, one size fits all. But agriculture never has been, and it’s only getting a lot more diverse and variable and dynamic. And we need to accept that reality and create systems and processes that can surface that diversity and not try to just floss over it.

[00:28:01] David: Exactly. And our research needs to be serving that diversity, not creating blanket solutions. So this is something which isn’t just about changing extension, it’s not just about changing research, it’s not just about changing farmer practices, it’s about creating a more systemic approach where you actually have better symbiosis between these different activities.

And that’s what I think we are all dreaming of. So, yeah, excited to have you coming on board. These discussions are gonna carry on for years, I’m sure. But it’s been great to have this discussion and to keep going with these discussions.

[00:28:38] Rikin: Thank you so much for having me. It was really great to have this conversation and learn so much about the history, about the platform. I’m so excited about seeing how we can best plug in.

[00:28:48] David: Thank you.