Can AI Predict Climate Migration?

This transcript was generated using AI and may contain inaccuracies. If you notice an error, feel free to email [email protected].

CHAPTERS 

[00:02:01]: Using survey data to measure migration intentions 

[00:04:52]: Machine learning methods and their application to migration research 

[00:07:05]: How climate shocks interact with existing migration drivers 

[00:12:51]: Comparing machine learning with traditional analytical approaches 

[00:20:01]: Bias, transparency, and risks in the use of AI for migration governance 

[00:21:33]: Potential applications of AI for monitoring and early response

TRANSCRIPT
[00:00:07.11] 

It's hard to exist in the world today without encountering artificial intelligence in ways large and small. An increasingly complex array of algorithms and data driven insights are helping to drive policy and shape the world in which we live. That's no less true for migration. A growing number of international organizations and governments and nonprofits have started to explore how AI tools tools can compile and analyze huge amounts of data to help predict future migration, including climate migration, in order to more effectively protect people at risk and manage future movement. This is Changing Climate, Changing Migration from the Migration Policy Institute. This podcast is dedicated to exploring the complex ways that climate change is driving migration. My name is Julian Hattem. I'm the host of this podcast and the editor of MPI's online magazine, the Migration Information Source. Check us out at migrationinformation.org. My guest today is here to help me resolve a big question with a complicated answer. Can AI predict climate migration? I'm Speaking today with Dr. John Aoga. John is a data scientist at the University of Louvain in Belgium who specializes in machine learning and he was the lead researcher on a project looking at how AI can help analyze existing climate related migration and predict future patterns in West Africa.

[00:01:43.20] 

John, thank you so much for coming on the podcast today. It's great to have you.

[00:01:46.21] 

Thank you for having me.

[00:01:48.19] 

In very simple terms, how did you do your study? What questions did you ask? What sort of data did you put in? And I mean, how did you analyze the numbers that came out?

[00:01:59.20] 

Thank you for that question. So this research was mainly done using a data set called Gallup. And the Gallup data set is, let's say, a survey that they do around the world in almost all the countries. And they ask more than 2,000 questions to people about many things. And so the part that I was interested in, that research was mainly related to migration, but mainly here it's not if people really move, it's mainly about intention to move. And we combine that then with climate data that we have, but we don't look at all the climate, let's say futures. We mainly focus on shocks like droughts and also fruits.

[00:02:56.11] 

And you looked at what, seven or eight countries in West Africa and the Sahel? A little bit. Right. What countries did you look at?

[00:03:02.09] 

Yeah, so we look at Senegal, Mauritania, Niger and Ivory coast countries. So we look at six countries actually. And why we were mainly looking at those countries, it was because we realized that, for example, if we look at, let's say, other countries around the world where they use technologies in agricultures we were not able to capture the real impact of climates on the migration or on the economy where those countries, mainly West African countries, has first agriculture as the main economic driver. But, but also they are still using, let's say basics techniques in agriculture, which are mainly based on climate movement, if climate based on season. Actually that's why there was interesting to look at them.

[00:04:09.04] 

Sort of climate vulnerability, especially for agricultural dependent communities. Some overlap with climate impact, I guess. And before I get to your findings, I want to ask very briefly about the machine learning aspect of this. So I acknowledge I am not a data scientist. Many of our listeners are lot data scientists in I guess very, you know, to clarify in very simplistic terms, how can you explain what you did and how the, how the correlation between migration intentions and climate shocks. Climate shocks, what you did there? Like, what is like, what is machine learning? It's a subset of like artificial intelligence. Right. And I guess how did you, how did you extricate the answers?

[00:04:48.17] 

Thank you for that question. So I can say that machine learning is something that we all do actually. When you take something, maybe a phenomenon happened, what you want is to have a way to reproduce that event. So to be able to do that. In our case we use data and then from data we want to derive what we call model. And the model will be something we can call. For example, if I'm thinking about mathematical formulas, it could be just a formulas where if I put a Y, an X, then we can get a Y. So but here X can be really complex and Y can also be really complex. For example, that can be an image that I give to my model and I want to predict if that the image contains a cat or a dog or something like that. Now in our case, what we want then is to use data to find a model that will be able to say, for example, based on the weather conditions, mainly shocks, if people will migrate or not, based on some features like age of the people and others characteristics from the populations.

[00:06:08.11] 

And then using that, what we will do mainly is to pour the data to the system and then find that model. And whenever we have that model ready, then we can try them with new data, new examples that we have. We will submit new people information and see if they will move or not based on that.

[00:06:32.24] 

What did you find? What did it say? What is the connection between climate shocks and migration intentions in these countries?

[00:06:40.12] 

So thank you for that question again. So the first thing that we realize that. So even if we are not looking to climate change, we realize that many works before we Know that for example age, gender, incomes will influence the way that people we decide to move or not. But what we realize that even with the climate shocks those features are always here present and then they are like more like as I said with that they will matter more with the climate shocks. We were like studying two kind of intention. Short term intention to move like around one and two years and, and long term intention to move more than five years intention to move. And we realized that for the first one really shorts shocks, for example, let's say a harsh time of rain or something like that can influence small shorts intention of people to move. But then long term droughts for example will influence people from long term to move. But something that was really interesting is if the category of people is between 15 and 24, they always want to move. Like even if it is for short or long term, they always want to move.

[00:08:13.20] 

But then if you have people now like from 35 and 49, that will depend on if the climate condition continue to be really hard for long term.

[00:08:27.17] 

So the older people are the more likely they are to be affected by the, the longer the climate shock has to be for them to be more impacted. But otherwise the climate shock is generally exacerbates pre existing migration drivers. I guess that's interesting. Yeah. And you said you looked at six West African countries. Yeah, I guess. To what extent can you speculate do we think A that these, these findings, these correlations apply elsewhere, especially if not entirely global, certainly in other relatively low income, agriculturally dependent contexts. Question A and then question B. I mean can we also apply them forward? What, how much predictive power do you think there is in, in these kind of correlations that you've drawn?

[00:09:10.21] 

I, I think that's the, the way to, to see that young people always want to move and older people don't really want to move or that will only be if the conditions are really hard. This is kind of, kind of global, I think. Yeah, but then if we take let's say specificity here, what we realize in the, in the work is that at some point we were like asking ourselves if we can take our results to global to other countries. But unfortunately what we realized that the global model was performed really poorly. When we take the model for specific country and then specialize them, they are really efficient for each countries. But then when we want that model to be efficient for other countries, then it's performed really poorly. So the takeaways there is it's better to have a specific model for each country for at least specific let's say features, because as we realize, for example, if a country use technologies more than another, of course the result of the migration intention will be different actually.

[00:10:40.13] 

So it's really country dependent, the same connection. And Senegal does not work in South Africa or Guatemala or France or Belgium. Okay, interesting. Going forward, do you think that there is a predictive power for future climate events as well? I mean, if there were a drought tomorrow, do you think that this kind of connection would help predict migration intentions?

[00:11:02.04] 

Yeah, what I'm thinking is, for example, since for example West African countries continue to use like really basic, let's say technology in agriculture and so on, I'm pretty sure that this result, even if the data was between 2016 and 2016 today, I think that is still applicable for those countries. And that means that today, for example, if we have really long term droughts in West African countries, or at least countries that are, that have, for example agriculture as the main driver of the economy, we can use this model to at least predict in the way people will want to move. But then I think that if we want to tackle other countries, we need for example, to make new survey to see how the new climate shocks will influence those new countries.

[00:12:09.05] 

I want to kind of talk about them, if not the methodology, at least the use of AI and machine learning here for these kind of, I guess what, what are the benefits of using these technologies for drawing these correlations and making predictions versus more traditional approaches that were perhaps.

[00:12:26.02] 

Yeah, so I will say a word first. In traditional approaches. So in traditional approaches what's happening is, for example, if we have six features, for example, related to population and climate, what will happen is they will do a model for each feature and then do meta analysis to combine those features to have one result at the end. So one first advantage of using AI tools like machine learning and so on is the way that we can handle many variables together. Like we don't need to run in that case six different model and then do meta analysis. We can use only one model that can incorporate all the variables and features and more than that can also capture the influence of maybe interaction between pairs of variables or maybe more than those. So that is the first.

[00:13:28.17] 

So it's both more efficient and kind of catches the interactions. Exactly.

[00:13:32.23] 

And then another benefit would be the way that using machine learning we can capture more non linear patterns that maybe statistical basic model will maybe may miss. And last thing that I can realize, since we are like doing this analysis together, we can also like quickly set up the model and then find new insights quickly. And also maybe if we have new data, we don't need for example, to do all the specific stuff, let's say specifically, but we can just run one things and have the final result based on that.

[00:14:21.21] 

Are there limitations both I guess, of using machine learning and AI and also more generally? I mean, I assume this kind of research depends on a lot of underlying data. That's available data.

[00:14:32.04] 

Exactly.

[00:14:33.01] 

I mean, are those data always available? Especially you know, in places like Niger, which are political, socioeconomic, price problems. What are the limitations that you've come across?

[00:14:42.20] 

So I can see two kind of limitations here. So the limitation about our work, but also the limitation about machine learning in general. So the limitation about our work is that we were focusing on looking into intention, migration intention, where the actual move can be something interesting to look at. But finding this kind of data was hard at that time. And also based on because of ethical maybe view. It is not maybe interesting to get this kind of data.

[00:15:18.19] 

Sorry because of an ethical view. What do you mean?

[00:15:21.22] 

So for example, if we want to understand the actual move, maybe we have to take for example some people, a sample of people in the population and then follow them for let's say five years or maybe 10 years and look at where they go, what they do and, and so on. And doing that we can maybe capture some privacy information that shots maybe interesting.

[00:15:50.08] 

It's also logistically difficult. Yeah, exactly.

[00:15:52.21] 

Yeah. But some a way to make that possible is for example to use telecommunication data. Because with cell phone information, of course we can get all those information, but telecommunication company won't give those information because yeah, GDPR and so on things. But it's anyway limitation of our study. But then when we look at machine learning in general, of course, first thing that we can see as limitation is data. Because if we need for example to capture this migration intention or even migration behavior in a country, we need for example to run new survey to get new data from those countries to be able to analyze how the migration will go and what will be the impact of that migration or an economy. But other than that, of course today using AI, we should also think we should be aware about ethical aspects of using them, like ecological aspects, energy consumptions and, and these kind of things.

[00:17:09.13] 

Because these tools, just to clarify, these tools require a lot of energy and

[00:17:13.05] 

a lot of energy especially there's an

[00:17:14.23] 

irony if you're studying climate change, but the tools you're using requires a lot of energy, which is contributing some ways.

[00:17:22.07] 

Exactly. And something else related to machine learning is also the Biases. So for example, it's happened since we are working with people data and based on the fact that we do survey and then ask people questions and so on, based on the country, it can also be in touch with some biases and our model will definitely reproduce these biases.

[00:17:53.11] 

What kinds of biases? Can you give an example?

[00:17:56.02] 

So there are many kind of biases when we use machine learning. For example, if I have a data that's, let's say, let's take a real application, let's say a biometric application to have a face recognition. If the data is poor with many white skin data, then it will be for example difficult to capture other kind of skin you see. So this kind of biases can be present and your model will reproduce it directly. So there should be like some analysis of the data upfront to make sure that at least we are like balanced in terms of representation of each class of population that we are attacking.

[00:18:42.10] 

That's a great point. And that kind of is a segue to a question I wanted to talk about which is the use of machine learning and AI, both for research but also for kind of policy making and for managing migration. I mean even in the absence of climate related impacts, there have been an increasing number of efforts by governments and international organizations to use AI software to predict and manage migration, both to predict regular migration. And you imagine who's going to come, how many people, what to do with it. And also some AI, facial recognition and other technologies to facilitate entry. Yeah, and so I guess there's also a. I don't know if it's a tension. It's one thing for researchers to do with this, to deal with these technologies and to try and learn from them is are there differences when these kind of technologies are in the hands of governments? Are there more ethical considerations then? And I guess yeah. What are the promises and perils of these sort of technologies both in, both for research but also more generally in, for governments and international organizations.

[00:19:45.10] 

Yeah, thank you for that question. So we talking about ethical consents. I think that we can see for let's say components here. Biases. We already talk about biases. Privacy. We also talk about privacy, but we can also talk about transparency and misuse. For example, transparency will be like at some points we have models that will predict something, but behind the model we don't know how the model were able to do that prediction. And that comes with some kind of transparencies like making sure that what we have in, in the data is what should be used to have that exact Prediction that we are, we are getting and misuse, for example, doing that research, it was at some point to help us to help at risk communities. Right? But then for example, if a government gets those information or maybe this model, it can use it for example to restrict movement. Right? Because they understand now how orientation work and so on, they can use it to restrict movement instead of helping us to help actually at risk communities. So misuse is also something really important in this doing this kind of research.

[00:21:10.22] 

That's great and that is a great transition to I guess probably my final question. We're about out of time, I guess. What is the positive view, Steve? What is your hope for the kind of, the applications of these kind of technologies of your study in general and other studies like it, relying on AI? What are the possible benefits?

[00:21:30.05] 

Yeah, thank you for that question. So for me, doing this prediction actually for policymaker in general and also ad organizations would be for example, to have a general monitoring of what is going on, for example per country or maybe per region where we can use those models over time to adjust intervention to see where we can prioritize, for example, intervention where we can prioritize, for example funding. It could also help for example to flag some district or maybe region, for example, to say that okay, these cities or maybe this region has at risk communities that we should mainly care about and they may be maybe more sensitive about things. Having, for example an observatory of the climate conditions on some areas can help to see, okay, based on if we are having more than, let's say four months of droughts, we can say, okay, okay, if we have more than four, four months here, that means that we should care more about these, we should take this into account. Maybe if not, maybe we'll have a disaster here and so on. So for me, so that means we could have observatory to be able to monitor things in order to take good decision.

[00:23:02.16] 

We can also have a seasonal, seasonal plant cash for the founding. Based on those conditions, we could use this research also to prioritize where to, to put more money where we need to help people and so on.

[00:23:19.02] 

That's great. I think we've got to wrap it up there for timing. We're going to have to say goodbye. But John, thank you so much for coming on the program today. This was another day. I'm not a data scientist, as I've said several times and this was very helpful for me in drawing these connections. Thank you for your time.

[00:23:34.09] 

Yeah, thank you for having me. And I hope that this also help you to understand a bit the what this domain work and I hope I didn't use, let's say, jargon and to be understandable.

[00:23:50.02] 

You were great. Thank you. Dr. John Aoga is a postdoctoral researcher at the University of Louvain in Belgium. He is a lead author of a paper entitled Impact of Weather Factors on Migration Intention Using Machine Learning Algorithms, which was published in the journal Operations Research Forum. Thank you for listening to Changing Climate, Changing Migration from the Migration Policy Institute. Find all of the episodes in our archives on MPI's website at migrationpolicy.org/podcasts. You'll find conversations with guests about various predictions for future climate, migration impacts in certain regions of the world, and a lot more. If you liked today's conversation with John, make sure you listen to the episode Before the Storm: Getting out in Front of Climate Displacement, in which I explore how humanitarian organizations are trying to anticipate climate displacement and plan accordingly. Subscribe to the podcast so you don't miss an episode, and if you like what you hear, please leave us a review in which makes it easier for other people to find us. We're on Spotify, Apple Podcasts, and all the other major podcast platforms. This podcast is just one part of MPI's work trying to understand climate migration. We also have a special collection of articles in the Migration Information Source magazine and lots of research and policy briefs from top experts.

[00:25:13.23] 

All of that is on our site at migrationpolicy.org/climate. Elizabeth Navarro produced this episode, Michelle Mittelstadt provided editorial oversight, and Lisa Dixon offered additional assistance. Our theme music is Touch by Patrick Patrikios. Thank you again for tuning in. Once again, my name is Julian Hattem. I'll catch you next time.