Latest / What's Your Problem? / Using Satellites to Prevent Famines

Transcript

Using Satellites to Prevent Famines · What's Your Problem?

0:15
Speaker 1: Pushkin.
0:31
Speaker 2: I'm Jacob Goldstein, and this is What's Your Problem? My
0:33
Speaker 2: guest today is Catherine Nakalembe. Catherine is an assistant professor
0:38
Speaker 2: at the University of Maryland. And she runs the Africa
0:41
Speaker 2: program at NASA Harvest, which is a project that uses
0:45
Speaker 2: satellite images to track agriculture around the world. And it
0:49
Speaker 2: wouldn't be wrong to say that Catherine's problem is using
0:53
Speaker 2: satellite imagery and GoPro cameras and AI to predict seasonal
0:57
Speaker 2: food shortages in sub-Saharan Africa. But in fact, her work,
1:03
Speaker 2: the problem she's trying to solve, is problematic. ultimately more
1:06
Speaker 2: human than that, ultimately more interesting, I would say, because
1:10
Speaker 2: the problem she's really trying to solve is, how do
1:13
Speaker 2: you use evidence to convince leaders to act to prevent
1:17
Speaker 2: people from starving? Catherine grew up in Uganda. She went
1:22
Speaker 2: to the University of Maryland to start a PhD, and
1:25
Speaker 2: then she went back to Uganda to do her graduate research.
1:28
Speaker 2: And the project was combining satellite data and on-the-ground research
1:32
Speaker 2: to to predict when the crops were likely to fail,
1:35
Speaker 2: leading to widespread hunger. And she was working on this
1:38
Speaker 2: for a few years. And then in the summer of 2015,
1:41
Speaker 2: when she was still a graduate student, there was a
1:45
Speaker 2: particularly bad drought in the region she studied. And she knew,
1:48
Speaker 2: she knew from her work that the people there would
1:51
Speaker 2: not have enough to eat in a few months. And
1:53
Speaker 2: she decided she had to get that information to someone
1:56
Speaker 2: with the power to do something about it.
1:58
Speaker 1: I realized really quickly that I knew more about the severity,
2:03
Speaker 1: basically the distribution of drought across this region more than
2:07
Speaker 1: anybody because I was able to look at it from space.
2:11
Speaker 1: And one year I finished my work and I was like,
2:13
Speaker 1: I can't just go back with this information because it
2:15
Speaker 1: had been an extreme drought year. I went to the
2:17
Speaker 1: office of the prime minister to share my My data
2:20
Speaker 1: and evidence in that year had taken lots of photographs
2:23
Speaker 1: and videos, et cetera. So I brought that. And then
2:25
Speaker 1: they had a meeting with the prime minister that led to— Wait.
2:28
Speaker 2: Can I just pause? So you're, what, a grad student
2:32
Speaker 2: or a young professor.
2:33
Speaker 1: At this point? No, I'm a grad student.
2:35
Speaker 2: And so it's an unusual thing for a grad student
2:38
Speaker 2: to be like, oh, my God, I have to tell
2:40
Speaker 2: the prime minister about this. What happened? How did that happen?
2:45
Speaker 1: So I've been coming for the last three years. People
2:49
Speaker 1: are trying to do the same things and crops fail.
2:52
Speaker 1: And I'd already started to kind of figure out how to,
2:55
Speaker 1: you know, predict what's going to happen, you know, early
2:57
Speaker 1: in the season. And there's always news around this issue,
3:00
Speaker 1: you know, Karamoja drought again, etc.
3:03
Speaker 2: So Karamoja is the region in, I guess, northern Uganda
3:07
Speaker 2: where you were studying.
3:09
Speaker 1: Yeah. And so it's always contentious. Is it happening? Are they,
3:13
Speaker 1: you know, just pretending? And this year, I had so
3:17
Speaker 1: much evidence, you know. First, because I'd taken so many photographs,
3:20
Speaker 1: I'd recorded video. And I wanted this to be sort
3:24
Speaker 1: of complementary to my PhD research. So I went and
3:27
Speaker 1: I shared it to the commissioner. And then he asked me,
3:32
Speaker 1: can you stay an extra few days? Because we have
3:35
Speaker 1: this food security meeting on Thursday with the prime minister
3:39
Speaker 1: and his steering committee, which are the other ministers. And
3:43
Speaker 1: I was like, of course. So I show up and
3:45
Speaker 1: I share the evidence and they asked me to say
3:47
Speaker 1: a few words. And I was like, you see this photograph?
3:50
Speaker 1: You see all of this thing that's dead here? This
3:52
Speaker 1: map basically represents the same thing everywhere. And it's eminent
3:57
Speaker 1: and people will be hungry pretty soon. And three days
4:00
Speaker 1: later or two days later, they sent food trucks. I
4:03
Speaker 1: have this photo that was sent to me from the
4:06
Speaker 1: guy who was the secretary to the prime minister or
4:10
Speaker 1: personal assistant. He was like, the trucks left this morning.
4:15
Speaker 1: I have this email, and I was amazed that I
4:20
Speaker 1: did that, and it went that far, basically.
4:23
Speaker 2: And that fast. So when you say people will be
4:26
Speaker 2: hungry pretty soon, that is an understated way, it seems,
4:30
Speaker 2: of saying what's going on. What was going on?
4:34
Speaker 1: You have subsistence farming where you produce what you eat.
4:38
Speaker 1: Very little is left over. And at the end of
4:41
Speaker 1: the season, you're supposed to sort of start picking up
4:45
Speaker 1: or eating what will be available to you through to
4:48
Speaker 1: the end of the next season. So this is typically
4:50
Speaker 1: around September. And so by August, everything had failed. And
4:56
Speaker 1: so there wasn't going to be anything harvested. But if
5:01
Speaker 1: you've kind of been out in the field a collecting data,
5:06
Speaker 1: and you run into kids eating sorghum shoots, for example,
5:09
Speaker 1: you know, the middle part of a.
5:10
Speaker 2: It's kind of like eating grass.
5:15
Speaker 1: Exactly. So like the middle part of a corn stem,
5:18
Speaker 1: it's sugary. And so, you know, you find kids eating
5:22
Speaker 1: that and they don't have anything else. They don't have livestock.
5:26
Speaker 1: Their family does not have any other source of income.
5:29
Speaker 1: It's pretty dire. So I think the other thing that
5:32
Speaker 1: might have made a huge difference in this particular instance
5:35
Speaker 1: is the photo evidence that I had. I take thousands
5:39
Speaker 1: of photos. I think I have over 300,000 photos.
5:41
Speaker 2: That's such a low-tech thing. Do you know what I mean?
5:47
Speaker 2: It's not like you were doing some wild satellite AI something.
5:52
Speaker 2: You were just walking around and taking pictures.
5:54
Speaker 1: Mm-hmm. And I had my map, my drought analysis map,
5:59
Speaker 1: that was very simple. And then we called it NDVI,
6:02
Speaker 1: Normalized Difference Vegetation Index, the anomaly of it.
6:06
Speaker 2: Basically, how much drier than usual. Exactly.
6:09
Speaker 1: Yeah.
6:10
Speaker 2: So, okay, so you do this, and the result in
6:13
Speaker 2: that first year is that they get food aid to
6:15
Speaker 2: people sooner.
6:16
Speaker 1: Right. Almost immediately. I have this text message from the commissioner.
6:23
Speaker 1: He said, this is the first time that nobody argued
6:26
Speaker 1: with the evidence.
6:28
Speaker 2: Did they then integrate your work in an ongoing way?
6:32
Speaker 2: How did that play out in the years after that?
6:36
Speaker 1: So the project that's called Disaster Risk Financing was basically
6:40
Speaker 1: being designed the following week. So I stayed that extra
6:43
Speaker 1: week and was like, I can not only develop the methods,
6:46
Speaker 1: but I can train the office how to utilize the
6:49
Speaker 1: data that's available in order for them to do that
6:51
Speaker 1: early warning, as well as write summaries, reports, et cetera.
6:56
Speaker 2: Take pictures. It sounds like one big lesson is take pictures.
6:59
Speaker 1: It's a validation for sure, but they'd already understood. So
7:03
Speaker 1: I kind of created this method to understand, compared to
7:06
Speaker 1: the history, if we hit this threshold, things are going
7:10
Speaker 1: to fail. Considering what has happened in the last 30
7:13
Speaker 1: years or last 20 years, If we hit this threshold,
7:16
Speaker 1: vegetation won't recover and we need to start planning.
7:19
Speaker 2: Basically a threshold of drought. If rain is below some level.
7:23
Speaker 1: So initially it was like if vegetation conditions drop below
7:27
Speaker 1: this level within this period, a few months, months within
7:32
Speaker 1: the growing season, then we need to do something.
7:35
Speaker 2: So it's interesting. I mean, I might naively have thought
7:38
Speaker 2: that the constraint, the fundamental constraint was just money.
7:43
Speaker 1: Right.
7:43
Speaker 2: That there's just not the money to get the food
7:45
Speaker 2: to the people or infrastructure or something. But it's information
7:49
Speaker 2: at some level, like reliable evidence.
7:52
Speaker 1: Reliable evidence that is accessible. But I think money and
7:58
Speaker 1: the willingness to use that money is also really important.
8:03
Speaker 1: So that region where I was studying, even though it
8:07
Speaker 1: had the lowest economic indicators, you know, household income was
8:11
Speaker 1: very low. Infrastructure is very poor. The cost of food
8:16
Speaker 1: was so much higher than it is in the capital. Yes.
8:20
Speaker 2: Well, that goes with poor infrastructure, right? Exactly. When it
8:23
Speaker 2: costs a lot to move food, the food costs a lot.
8:27
Speaker 1: Exactly. Now, add the fact that there is a drought
8:31
Speaker 1: and traders can do whatever they want. And so it
8:34
Speaker 1: becomes much more expensive. And when the government is trying
8:37
Speaker 1: to acquire this water and food, et cetera, they end
8:40
Speaker 1: up paying a lot more. Uh-huh. And so if you're
8:44
Speaker 1: doing it in anticipation, it means that then you can
8:47
Speaker 1: acquire things at a much lower cost.
8:49
Speaker 2: Oh, so if they know in August that there is
8:51
Speaker 2: going to be great need in September, they can buy
8:55
Speaker 2: before there's a run.
8:56
Speaker 1: Exactly. And the other thing was, because it was in anticipation,
9:00
Speaker 1: what ended up happening is they would have projects that
9:06
Speaker 1: could hire people to work on that were community development projects.
9:10
Speaker 1: It could be tree planting. It could be road construction,
9:13
Speaker 1: it could be improving wells, and then they will be
9:16
Speaker 1: paid for those jobs. Yeah, that then they could use
9:19
Speaker 1: that money to buy food and other things.
9:22
Speaker 2: So it's interesting. I thought we were going to be
9:24
Speaker 2: talking about satellites and AI, and I guess we will be.
9:28
Speaker 2: But it's interesting that this conversation starts in such a
9:32
Speaker 2: low-tech way. I mean, obviously, digital cameras were not quite
9:35
Speaker 2: as ubiquitous as they are today. But this is 2016.
9:37
Speaker 2: It's not 2000 or something. Lots of people already had
9:40
Speaker 2: digital cameras.
9:44
Speaker 1: Why? Yeah, why? I'd say people believe what they see.
9:50
Speaker 1: And the NASA administrator in 2019, Jim Bernstein, he used
9:57
Speaker 1: to share my photos in his presentations. He'd talk about
10:02
Speaker 1: from NASA to the moon and beyond. But he would
10:06
Speaker 1: still use the photos and not the maps or the images.
10:09
Speaker 1: the numbers because it tells it you know we can
10:12
Speaker 1: see it from above and this is what it means
10:14
Speaker 1: on the ground this is the reality of what it
10:17
Speaker 1: is on the ground and that's i think much more
10:19
Speaker 1: accessible and understandable.
10:21
Speaker 2: Um so I feel a little weird now after you're
10:25
Speaker 2: talking compellingly about how important photos are. But I do
10:28
Speaker 2: want to talk to you about satellite imagery and AI.
10:31
Speaker 1: But they're photos.
10:32
Speaker 2: They're photos. Very good. They're photos. I find them moving
10:36
Speaker 2: in a different kind of way. I mean, you know,
10:40
Speaker 2: I've read this sort of big hand-wavy thing, which is,
10:44
Speaker 2: You know, there's lots of satellite imagery of the Earth
10:46
Speaker 2: now and lots of AI models designed to analyze crops,
10:49
Speaker 2: but those are largely designed for kind of big monocultural agribusiness,
10:54
Speaker 2: and they don't apply so well, certainly, to the kind
10:57
Speaker 2: of smallholder farms in the places you have worked. Like,
11:01
Speaker 2: I get that that's kind of the big picture. But
11:05
Speaker 2: more narrowly, like, tell me about your work to deal
11:08
Speaker 2: with that. Like, first of all, what is it that
11:09
Speaker 2: you want to know from satellite imagery?
11:13
Speaker 1: We want to understand what's growing where and how it's doing.
11:19
Speaker 1: The other is the complexity of the agriculture system. So
11:23
Speaker 1: when we're looking for crop fields, we're looking for shapes
11:28
Speaker 1: and then we're looking for some sort of seasonal pattern.
11:32
Speaker 1: So if you had 10 images over a growing season,
11:36
Speaker 1: you should see something like it's cleared now. it browns,
11:41
Speaker 1: it's like brown and wet, and then something green pops up,
11:45
Speaker 1: it matures, and then it's harvested. So it sort of
11:48
Speaker 1: like becomes, you know, dry. And so we should be
11:51
Speaker 1: able to see that signal. However, how low the brownness
11:56
Speaker 1: might be, like how dry something might be, is contextual.
12:02
Speaker 1: And while a reduction might seem like minor in some places,
12:07
Speaker 1: it's actually a very serious problem. And so the first
12:11
Speaker 1: thing was create a cropland map, basically exclude everything that's
12:15
Speaker 1: not cropland, estimate, understand vegetation conditions historically, like what happens
12:21
Speaker 1: every season, what months matter the most in terms of vegetation,
12:25
Speaker 1: in terms of crop production, and then try to understand
12:28
Speaker 1: at what point recovery doesn't happen. So to be able
12:32
Speaker 1: to do that at a really large scale, satellites are
12:35
Speaker 1: a phenomenal tool for this. At the time, even today,
12:39
Speaker 1: the only satellite that allows us to be able to
12:41
Speaker 1: do this at a much higher frequency is MODIS. And
12:45
Speaker 1: MODIS is a 250-meter resolution. That could be multiple crop
12:49
Speaker 1: fields in a single pixel.
12:52
Speaker 2: So the farms are too small or too varied.
12:56
Speaker 1: Too varied, yep.
12:58
Speaker 2: Tell me, what is a typical farm in this region like?
13:02
Speaker 1: So in Karamoja, people grow sorghum, millet, a little bit
13:07
Speaker 1: of maize. There's a lot of sunflower, too.
13:10
Speaker 2: And will they grow all that in one place, more
13:13
Speaker 2: or less?
13:13
Speaker 1: So you will find sunflower mixed in with sorghum, for example.
13:17
Speaker 2: And how big is a typical farm?
13:20
Speaker 1: So farm plot... On average, less than 2.5 hectares. So
13:26
Speaker 1: a football field, a soccer field is about an acre.
13:31
Speaker 1: That's the average over most of sub-Saharan Africa, but they're
13:37
Speaker 1: much smaller. So like in a single soccer field, you
13:39
Speaker 1: might find maybe four plots as an example.
13:43
Speaker 2: Like four different people's little farms?
13:45
Speaker 1: Four different people's plots with hedges, and they're not like
13:49
Speaker 1: distinct perfect shapes like you would see.
13:51
Speaker 2: Quite small. Hard to see from a satellite, presumably, especially
13:56
Speaker 2: if there's a bunch of different crops there.
13:58
Speaker 1: So hard to see their boundaries. So a single modus
14:01
Speaker 1: pixel might include four of those fields that might have
14:04
Speaker 1: different things, for example.
14:06
Speaker 2: Yeah. Let's talk about the GoPros. So fun. Tell me
14:12
Speaker 2: about your work with GoPros.
14:13
Speaker 1: So one of the issues is in order to get
14:17
Speaker 1: from cropland, where all the crops are growing, to go
14:20
Speaker 1: to crop type. So the USDA, for example, produces a
14:23
Speaker 1: cropland data layer, which if you take a look at it,
14:27
Speaker 1: you can see where corn, wheat, soy is growing. And
14:30
Speaker 1: now that's been changing over the last couple of years.
14:32
Speaker 1: In the US. In the US, yes. This also exists
14:35
Speaker 1: for Europe. They have something similar. There's, of course, some simplicity.
14:40
Speaker 1: There's a single crop. The fields are homogeneous. They're kind
14:44
Speaker 1: of easy to map. In East Africa, in addition to
14:47
Speaker 1: fields being small, I already mentioned the intercropping, mixed cropping,
14:51
Speaker 1: and all the heterogeneity, how people plant at different days.
14:56
Speaker 1: To be able to get a crop type map, you
14:58
Speaker 1: need a lot of training data. So, you know, one
15:01
Speaker 1: of the biggest advances, I think, you know, the reason
15:04
Speaker 1: why we have huge leaps in terms of AI is
15:07
Speaker 1: the availability of training data. You know, same as face detection,
15:10
Speaker 1: all people's faces on Facebook, et cetera, makes that technology
15:13
Speaker 1: a lot easier to develop. And so this is true
15:16
Speaker 1: also for crop type mapping. So it's like face detection,
15:19
Speaker 1: but for crops.
15:20
Speaker 2: Why do you want a crop type map?
15:22
Speaker 1: There's so many benefits to knowing exactly what crop is
15:26
Speaker 1: growing where. You want this map in order to understand,
15:29
Speaker 1: for example, what kind of fertilizer will be relevant in
15:33
Speaker 1: a particular region. So if you know there's places dominated
15:37
Speaker 1: by rice, just like in Ghana, a particular place in Ghana,
15:41
Speaker 1: where could irrigation infrastructure make a difference? The other and
15:45
Speaker 1: most important, most basic thing is, if you want to
15:48
Speaker 1: understand how much will be produced at the end of
15:51
Speaker 1: the season, you need to know the total area by
15:54
Speaker 1: specific crop. So if we have 25,000 hectares of corn planted...
16:00
Speaker 1: And the average condition or the average yield is 2.5
16:03
Speaker 1: tons per hectare. You multiply that and you have your production,
16:07
Speaker 1: which will allow you to understand how much corn is available.
16:10
Speaker 1: If it's a critical crop in a particular area, you
16:13
Speaker 1: would know that, you know, food security would be okay.
16:17
Speaker 1: So we had a big issue in Illinois with corn,
16:21
Speaker 1: et cetera. If Germany had a big issue with wheat production,
16:25
Speaker 1: bread being so important.
16:27
Speaker 2: Yeah.
16:27
Speaker 1: They would have to plan accordingly. So you need to
16:30
Speaker 1: figure out, what do I need to import to have
16:32
Speaker 1: some sort of stability?
16:34
Speaker 2: And so for Sub-Saharan Africa in general, or the regions
16:38
Speaker 2: you study, are there crop-type maps now?
16:42
Speaker 1: There are some crop-type maps, yes. Some are done on
16:47
Speaker 1: a very small scale, regional scale. But what this project
16:50
Speaker 1: was trying to do was try to, one, just figure
16:54
Speaker 1: out the fastest possible way we can collect data over
16:57
Speaker 1: really large areas. One, because in some countries there are
17:00
Speaker 1: three seasons. So Rwanda has three growing seasons. So you're
17:03
Speaker 1: looking at three crops. So to cover the whole country
17:07
Speaker 1: and get an understanding of did they grow corn more generally,
17:13
Speaker 1: bananas more generally? No. You need to be able to
17:16
Speaker 1: do the survey continuously.
17:18
Speaker 2: And the satellites don't have the resolution, essentially, to do it.
17:22
Speaker 1: To be able to do it, yeah.
17:23
Speaker 2: You can't tell from satellite data.
17:24
Speaker 1: So we need to train a model to tell the model,
17:27
Speaker 1: this pixel, this particular spot is banana. Find me banana
17:31
Speaker 1: everywhere else in the image, yeah.
17:34
Speaker 2: So you have this idea of like, okay, we need
17:39
Speaker 2: more on-the-ground data. So what do you do? What is
17:42
Speaker 2: the GoPro project?
17:44
Speaker 1: GoPros are perfect because they have a GPS. So when
17:49
Speaker 1: we take a photo, the photo has a location. So
17:52
Speaker 1: typically when I'm collecting field data, I need a photo, evidence,
17:56
Speaker 1: but also I need to be clear when somebody labels
17:59
Speaker 1: in the form that maize is in this growing stage
18:03
Speaker 1: and you have to stand in the middle of the field.
18:06
Speaker 1: Because I want to use that location to train my
18:08
Speaker 1: model to predict where else maze might be.
18:10
Speaker 2: This is the old school, slow, never going to scale
18:14
Speaker 2: technique that you're trying to kind of quasi-automate here.
18:17
Speaker 1: Yes. At this point, even if all I had were
18:21
Speaker 1: images that had a longitude and latitude, and I know
18:27
Speaker 1: where the camera is facing, I could go through all
18:30
Speaker 1: these individual photos in Google Earth Pro and add the
18:34
Speaker 1: label myself.
18:36
Speaker 2: So at least you would have a bunch of images
18:38
Speaker 2: with a very specific geolocation on latitude and longitude.
18:43
Speaker 1: And I know what the crops look like. I know
18:46
Speaker 1: what the crops are.
18:47
Speaker 2: So this is phase one. But you don't want to
18:50
Speaker 2: have to label a thousand things corn, whatever.
18:54
Speaker 1: Yeah, I would have been okay with that too, though. Considering,
18:58
Speaker 1: so I did this in Mali. I did it in Senegal.
19:02
Speaker 1: In Tanzania.
19:02
Speaker 2: Wait, this meaning driving around with the GoPro? No, the phone. Oh,
19:07
Speaker 2: the walking around.
19:08
Speaker 1: Walking around, yeah.
19:09
Speaker 2: Yeah, okay. I mean, it sounds kind of charming in
19:12
Speaker 2: a way.
19:12
Speaker 1: I learn a lot from doing field work. You learn
19:16
Speaker 1: what it means for someone's field to be washed out
19:19
Speaker 1: when you go. But if the idea is to try
19:22
Speaker 1: and create a map that will be useful for this task,
19:26
Speaker 1: there's only so much you can do when you cover
19:29
Speaker 1: a very small area. Yeah. So 2020, I call this
19:32
Speaker 1: project a COVID-safe project. 2020 was like, oh, well, now
19:36
Speaker 1: we have this funding. We have to figure this out.
19:38
Speaker 1: I go buy a GoPro at a store, a fashion
19:43
Speaker 1: sort of amount with my dad because I couldn't order
19:47
Speaker 1: it from Amazon. And then we make this magnetic thing
19:51
Speaker 1: on the car and just drive. I go to, there's
19:53
Speaker 1: an agricultural research station. And the director was like, I
19:57
Speaker 1: described to him what I'm trying to do. He was like, oh,
19:59
Speaker 1: course you can drive around here, whatever. So we try
20:02
Speaker 1: it out. So the first thing was, if we get
20:04
Speaker 1: the images, are they useful? Is the GPS good enough?
20:07
Speaker 2: And it was.
20:09
Speaker 1: And so then we were like, oh, this could actually work.
20:17
Speaker 2: We'll be back in just a minute. That's the end
20:29
Speaker 2: of the ads. We're going back to the show. And
20:31
Speaker 2: as you probably recall, when we left off before the break,
20:34
Speaker 2: Catherine had just tested out a GoPro herself and decided
20:37
Speaker 2: that GoPros could be a good way to gather crop
20:40
Speaker 2: data from small farms.
20:42
Speaker 1: So I have a friend in Kenya. I asked him
20:46
Speaker 1: if he can buy, if he can order things on
20:48
Speaker 1: Amazon and to work with a team there to see
20:53
Speaker 1: if they can collect data in Kenya. So we get
20:56
Speaker 1: a team of two. So basically four. So it's a
21:00
Speaker 1: team of two people. One team drives south, the other
21:03
Speaker 1: drives north. And we cover all of Western Kenya in
21:08
Speaker 1: a week.
21:09
Speaker 2: Wow. And they're just driving everywhere they can drive, in
21:12
Speaker 2: a car, on a motorcycle?
21:13
Speaker 1: So you can see from some of the original data
21:15
Speaker 1: that we have from Kenya that we're kind of covering
21:18
Speaker 1: these big... There are lots of fields, but they're proximal
21:22
Speaker 1: to these big roads, which... It's great, but not good enough.
21:28
Speaker 1: And so it became pretty quickly obvious that a motorcycle
21:31
Speaker 1: might be better because a motorcycle can go to a
21:33
Speaker 1: level two, level three, you know, smaller paths and go through.
21:37
Speaker 1: And in Tanzania, they went a little bit crazy because
21:39
Speaker 1: I feel like they're driving into rice paddies and going
21:42
Speaker 1: like this.
21:42
Speaker 2: Well, people live far off of roads that you could
21:45
Speaker 2: drive a car on. Exactly. The kinds of farms you're
21:48
Speaker 2: talking about exist far from drivable roads.
21:50
Speaker 1: Exactly. So... But just from the Kenya team, we had over,
21:55
Speaker 1: I think the first iteration might have been maybe 300,000
22:01
Speaker 1: images or more. Then the team of two in Uganda,
22:06
Speaker 1: I kind of equipped them. And during COVID, there's nothing
22:11
Speaker 1: else you could do and asked them to kind of
22:12
Speaker 1: just go drive around. And they went crazy because then
22:16
Speaker 1: we had over 2 million images from them. Okay. And
22:19
Speaker 1: so then it's just too many images. Even I got
22:22
Speaker 1: tired of just like trying to figure it out. But
22:24
Speaker 1: I had an undergraduate computer science student who worked over
22:27
Speaker 1: this and developed Street to Sat, the first iteration of it.
22:32
Speaker 1: And remove clouds, adjust images automatically so that they're kind
22:38
Speaker 1: of like straight. And then remove faces and ensure that
22:44
Speaker 1: they're actually crops in the image.
22:48
Speaker 2: Crop and not crop, you called it in the paper.
22:50
Speaker 2: It reminded me, I don't know if you watched the
22:52
Speaker 2: show Silicon Valley, but there's a guy in that show
22:57
Speaker 2: who invents AI.
22:59
Speaker 2: And they show it a hot dog and it says
23:02
Speaker 2: hot dog. And everybody's like, wow, it can detect food.
23:04
Speaker 2: And then they show it a hamburger and it just
23:05
Speaker 2: says not hot dog.
23:07
Speaker 1: Yeah.
23:07
Speaker 2: And all it does is say hot dog, not a
23:09
Speaker 2: hot dog. This is where you are now.
23:10
Speaker 1: This is where we are. It's like, is there a
23:12
Speaker 1: crop in this field? Yes. If there is, then that's phenomenal.
23:15
Speaker 1: And then we kind of did a couple of iterations
23:19
Speaker 1: of what we call labeling. Yeah. Basically, go through a
23:25
Speaker 1: couple of images to create a label dataset, which is,
23:29
Speaker 1: this is, we draw bounding boxes, maize, maize, this is
23:32
Speaker 1: what maize looks like, this is what banana looks like.
23:34
Speaker 2: So now you're saying you're, this is training, you're creating training.
23:37
Speaker 1: We're creating a training dataset, yeah.
23:38
Speaker 2: Training data, yeah.
23:39
Speaker 1: And then with the first iteration, then we run it
23:42
Speaker 1: through the initial data from Kenya to do sugarcane maize. Okay. Okay.
23:48
Speaker 1: And it was pretty good. It was pretty good. It
23:51
Speaker 1: was like, you know, when it's actually working and you
23:54
Speaker 1: can't believe that it's working, but it is working and
23:57
Speaker 1: it's unbelievable. So you're like, you're in this phase where
24:00
Speaker 1: you're like, oh my God, is this actually happening?
24:05
Speaker 2: What are people doing now with this basic technique of,
24:09
Speaker 2: you know, driving around with GoPro cameras where there are
24:11
Speaker 2: these small farms? Like what is one specific thing someone
24:15
Speaker 2: is doing that is helping them take action in the
24:18
Speaker 2: world in a way they would not have been able
24:20
Speaker 2: to do?
24:21
Speaker 1: So I give one example. So in Kenya, when we
24:25
Speaker 1: did the first iteration of it as a test pilot,
24:30
Speaker 1: and then were able to actually work with the Ministry
24:32
Speaker 1: of Agriculture for them to go out and collect the
24:34
Speaker 1: data themselves, they were like, do we get to keep
24:37
Speaker 1: the cameras? I was like, yeah. And I asked, you know,
24:42
Speaker 1: what sorts of things are you thinking about? I was like, well,
24:44
Speaker 1: you know, that would be kind of just Phenomenal photographic
24:48
Speaker 1: evidence because a person can go back whenever and send
24:53
Speaker 1: me those photos while they're out there in the middle
24:57
Speaker 1: of nowhere. So remember going back to the low-tech technology
25:00
Speaker 1: is that when the ministry reports, so we have these
25:03
Speaker 1: WhatsApp groups where extension agents share constantly what's happening. If
25:08
Speaker 1: they share that there was a flood and the flood
25:10
Speaker 1: was serious, everybody knows a flood is serious, but how
25:13
Speaker 1: serious is a different story. when it comes to things
25:17
Speaker 1: like beans, water staying on the surface for a long
25:20
Speaker 1: time and the beans turn yellow is because all the
25:23
Speaker 1: nutrients have leached. And so for the extension agents to
25:27
Speaker 1: be able to go through and collect these images without
25:29
Speaker 1: needing to get off, you know, take a photo, et cetera,
25:32
Speaker 1: et cetera, is something that is used as kind of
25:35
Speaker 1: evidence and as part of their reporting in the ministry.
25:38
Speaker 2: So it gives you a baseline.
25:40
Speaker 1: It's like a baseline, but also people believe what they see.
25:44
Speaker 2: It goes back to you walking around taking pictures, but
25:47
Speaker 2: just a lot more pictures.
25:48
Speaker 1: Yeah.
25:50
Speaker 2: So if you sort of distill it down, what is
25:53
Speaker 2: your big project? What are you trying to do in
25:56
Speaker 2: the world in your work?
25:58
Speaker 1: So data is an essential, right? Evidence is essential. But
26:04
Speaker 1: try to use it to improve, you could say, the
26:08
Speaker 1: human condition through better decision making. So If I know
26:12
Speaker 1: for sure that something is coming and it will be
26:16
Speaker 1: this bad, it means that the persons or people responsible
26:20
Speaker 1: for decision-making, they have no excuse to not do the
26:24
Speaker 1: right thing. I think that's kind of what kind of
26:28
Speaker 1: drives me. And I was going to talk about the
26:30
Speaker 1: 2.0 of the disaster risk financing program is because we've
26:33
Speaker 1: gone so far now since then that what I've built
26:37
Speaker 1: out now gives a three-month forecast of what things will
26:43
Speaker 1: be three months from now based on a model that
26:47
Speaker 1: has learned. So because I can process more data, I
26:51
Speaker 1: can combine multiple complex data sets than before, thanks to compute,
26:58
Speaker 1: and increase access to data. But I understand the problem
27:02
Speaker 1: so deeply that that threshold now is specific to the
27:06
Speaker 1: specific research It's no longer one threshold for all of them.
27:11
Speaker 1: It is a district, like a county-specific threshold that could
27:14
Speaker 1: even be lowered to be sub-county that could be zip
27:19
Speaker 1: code-specific threshold that in this zip code, this threshold is it.
27:26
Speaker 1: It doesn't apply to everywhere else.
27:28
Speaker 2: It's sort of zip code by zip code in Uganda.
27:32
Speaker 2: You can tell how much drought is going to be
27:34
Speaker 2: really bad. Three months in advance.
27:37
Speaker 1: Exactly. And it's not only in Uganda. The method is
27:40
Speaker 1: largely scalable. I could apply the same methodology to the U.S.
27:43
Speaker 2: Do you have the data? Yeah. Can you do it anywhere,
27:47
Speaker 2: or do you need some amount of data collection in
27:49
Speaker 2: some places?
27:50
Speaker 1: So for the forecasting component, all I need is satellite rainfall,
27:56
Speaker 1: satellite temperature, satellite vegetation conditions. And then the validation component
28:02
Speaker 1: is where I would need people to actually go and confirm,
28:05
Speaker 1: you know, am I actually predicting the right thing? As
28:09
Speaker 1: an example, you could use auxiliary things, you know, you
28:13
Speaker 1: could use images people are posting on social media about
28:16
Speaker 1: events to sort of as an in-between. But in order
28:19
Speaker 1: for the evidence, you know, to be strengthened, then you
28:23
Speaker 1: want to go, you know, take those photos. And then
28:26
Speaker 1: the other thing is, you know, have a routine component
28:28
Speaker 1: around collecting and gathering that evidence so that it's not
28:32
Speaker 1: whenever you feel like or whenever you have money, but
28:35
Speaker 1: it's just that at this point, we need to check
28:37
Speaker 1: in and see how things are going.
28:41
Speaker 2: Like, if I were going to say, Catherine's problem is this, like,
28:45
Speaker 2: what would the sentence be?
28:48
Speaker 1: There are lots of quite simple things that we could
28:53
Speaker 1: do already, but we just don't do them because they
28:57
Speaker 1: seem so Maybe we overcomplicate them. I always think about these,
29:02
Speaker 1: a minister will not, you know, read your F1 score table,
29:07
Speaker 1: you know, the confidence of your model, et cetera, but
29:10
Speaker 1: the minister will read a memo that's usually like three paragraphs.
29:14
Speaker 1: What I learned in the risk financing work is that
29:18
Speaker 1: it has to be distilled into a half-page memo. The
29:21
Speaker 1: memo is evidence points that we've hit the thresholds, about
29:29
Speaker 1: 70,000 people will need emergency aid, and we're going to
29:34
Speaker 1: release funding on this day. That's it. And then the
29:38
Speaker 1: minister says, I agree. Right, so whether or not it
29:44
Speaker 1: took me 100 months, $ 5 billion... 25 analysts to create
29:53
Speaker 1: the evidence or whatever it is, usually it's something that
29:59
Speaker 1: is accessible, concise, that helps that decision making. And accessibility
30:05
Speaker 1: to that information is so critical in this case. So
30:08
Speaker 1: I'd spend a lot of time, you know, training and
30:11
Speaker 1: building capacity so people can reproduce things.
30:13
Speaker 2: On their own.
30:15
Speaker 1: Like you can run the helmet's work all on your own.
30:18
Speaker 1: All the methods, everything you need in the kit. And
30:21
Speaker 1: a lot of people have done this as a team
30:23
Speaker 1: that's done it in Nigeria. People have done it in
30:25
Speaker 1: India on their own.
30:26
Speaker 2: Basically building crop-type maps using GoPro cameras?
30:30
Speaker 1: Creating crop-type labels, yeah, using GoPros. But also the workflow itself.
30:37
Speaker 1: We have a notebook available that anybody can run with
30:40
Speaker 1: a Jupyter notebook without anything else. So the problem is
30:47
Speaker 1: some things are a lot simpler than we make them
30:49
Speaker 1: out to be. And if the thing is to solve hunger...
30:54
Speaker 1: I think those solutions for solving hunger already exist. And
30:58
Speaker 1: they have nothing to do with my model or my technology.
31:02
Speaker 1: It has to be moving food where it is produced
31:05
Speaker 1: to where it's needed, making it available when it's needed
31:09
Speaker 1: the most, for example. And I can provide the evidence
31:13
Speaker 1: to be like, okay, it's absolutely needed here. So I
31:18
Speaker 1: think that's my problem.
31:24
Speaker 2: We'll be back in a minute with The Lightning Round.
31:37
Speaker 2: We're going to finish with The Lightning Round, which is
31:41
Speaker 2: just a little bit more playful than the rest of
31:43
Speaker 2: the show. What's one thing your dad taught you about
31:47
Speaker 2: fixing cars?
31:49
Speaker 1: Oh, my God, that guy.
31:52
Speaker 2: My dad my dad told me um.
31:56
Speaker 1: We had my car battery in 2020 of it was
31:59
Speaker 1: overheating whatever and he was like how do you drive
32:03
Speaker 1: a car without opening the hood and i was like
32:07
Speaker 1: like seriously and then i said this to him uh
32:10
Speaker 1: because his phone is runs out of memory and i
32:14
Speaker 1: was like How do you work with a phone that
32:16
Speaker 1: has no memory? So, you know, I was trying to
32:18
Speaker 1: kind of like paint this problem. My dad, what I've
32:20
Speaker 1: learned from him is you can figure it out. And
32:23
Speaker 1: he has this terminology. It's, I think, a Luganda thing.
32:28
Speaker 1: It says chito chige, which is kill it to learn it. So,
32:34
Speaker 1: you know, in terms of like fixing a radio, you
32:36
Speaker 1: take it apart and then you have to put it
32:38
Speaker 1: back together.
32:39
Speaker 2: It's like break it, break it and fix it and
32:41
Speaker 2: you'll understand it.
32:42
Speaker 1: Break it to fix it. basically, because then you will
32:45
Speaker 1: understand it. And so he always took apart, you know,
32:49
Speaker 1: car engines, and then he will put them all back together.
32:51
Speaker 1: So I think what I learned is that you can learn,
32:54
Speaker 1: you can figure it out, whatever it is.
32:58
Speaker 2: What's one thing you learned from your mom about running
33:00
Speaker 2: a restaurant?
33:02
Speaker 1: My mom is a workaholic. She's told me recently that
33:09
Speaker 1: She works, she likes, she keeps working because she likes
33:12
Speaker 1: to work. She's no longer, it's no longer profitable, but
33:14
Speaker 1: she likes to work and she believes she stops working.
33:17
Speaker 1: You know, she'll just like fall over and die. It's
33:20
Speaker 1: just an extreme, it's an extreme statement. But what she
33:23
Speaker 1: means is that whatever the work is, it's important and
33:27
Speaker 1: do it to the best of your ability. And so
33:30
Speaker 1: she does, you know, has done so many different things
33:32
Speaker 1: to kind of sustain it and keep it running. And so...
33:37
Speaker 1: Keep on moving, keep on working, and you can figure
33:41
Speaker 1: it out, I guess, also in this case.
33:43
Speaker 2: Who's a better badminton player, you or your sister?
33:46
Speaker 1: My sister, she's left-handed.
33:48
Speaker 2: Oh, is that a big advantage?
33:49
Speaker 1: Of course, you have no idea where her left from
33:51
Speaker 1: her right is. Her backhand is your, what you think
33:54
Speaker 1: is her backhand is her forehand. And then she worked
33:59
Speaker 1: so hard to improve her backhand that her backhand feels
34:02
Speaker 1: like a forehand, right? It's like the one when you
34:05
Speaker 1: hit it and it comes back. You're like, oh my God.
34:09
Speaker 2: Yeah. That's the sound it makes too, right? In badminton.
34:13
Speaker 2: What's one thing I should do if I visit Uganda?
34:21
Speaker 1: Eat a mango.
34:22
Speaker 2: Okay. Are they better there?
34:24
Speaker 1: Absolutely. And an avocado.
34:26
Speaker 2: And an avocado. Also better?
34:28
Speaker 1: Yeah. Never refrigerated. Fresh. Maybe harvested yesterday.
34:34
Speaker 2: Um... What about GoPros? Are you dreaming of other uses
34:40
Speaker 2: of GoPros for work or not for work or things
34:42
Speaker 2: other people could be doing with GoPros?
34:44
Speaker 1: Yeah. It's funny. I always thought about them. I used
34:49
Speaker 1: to wear one. The first, first iteration of GoPros, when
34:52
Speaker 1: I went to field work, I wanted to record everything.
34:55
Speaker 1: But about GoPros, if I was to design one, I'd
34:59
Speaker 1: include one that has a distance measurement system.
35:02
Speaker 2: I would do that.
35:05
Speaker 1: So I can know how far my maze is from
35:08
Speaker 1: where I'm standing.
35:11
Speaker 2: You're trying to do that with software, essentially.
35:13
Speaker 1: Yes. But I'd like for it to give me an
35:16
Speaker 1: estimate of how far something is. It's kind of easy
35:21
Speaker 1: to derive that if you're in a city because you
35:24
Speaker 1: have corners and things, very strict structures. But you can't
35:28
Speaker 1: do that when you're doing it for crops. I did
35:30
Speaker 1: win a GoPro for good, but they sent me a GoPro.
35:34
Speaker 1: which I already have a lot of GoPros.
35:37
Speaker 2: You won an award, and the award was another GoPro?
35:39
Speaker 1: Yes.
35:40
Speaker 2: What do you wish they'd sent you?
35:41
Speaker 1: I wish they had a conversation and I'd suggest what
35:47
Speaker 1: I want in the GoPro.
35:48
Speaker 2: Yeah, right. A GoPro that could measure distance.
35:52
Speaker 1: Yeah.
35:53
Speaker 2: I appreciate your time. It was lovely to talk with you.
35:56
Speaker 2: Thank you so much.
35:57
Speaker 1: Lovely talking with you, too. Thank you for having me.
36:07
Speaker 2: Catherine Nakalembe is an assistant professor at the University of
36:10
Speaker 2: Maryland and Africa program director at NASA Harvest. Please email
36:15
Speaker 2: us at problem at pushkin.fm. Tell us what kinds of
36:18
Speaker 2: shows we should do more of or less of or
36:20
Speaker 2: particular people you think would be good on the show.
36:23
Speaker 2: You can also find me on X and on LinkedIn.
36:26
Speaker 2: Our show is produced by Gabriel Hunter Chang and Trina Menino.
36:30
Speaker 2: Our editor is Lydia Jean Cott and our engineer is
36:32
Speaker 2: Sarah Bruguere. I'm Jacob Goldstein, and we'll be back next
36:35
Speaker 2: week with another episode of What's Your Problem?