Speaker: Adel Daoud at the Geo for Good Lightning Conference

Despite the progress that humanity has made over the last century, many children still die globally, and that is because of poverty and unequal access to resources. Part of the problem is that communities might be trapped in so-called “poverty traps.”

To analyze poverty traps, we need a lot of data across different dimensions of poverty. For example, here you see access to drinking water that we can characterize into different levels. The very poorest exist at level one, where they have no access to water in their household but have to move or travel far to find clean and safe water. If you compare it to other dimensions, you see the same type of differences over time. If we look at the global population distribution, we see that there are about seven and a half billion individuals on this planet—a bit more today. One-seventh of them live in the level one conditions you see listed here, whereas the richest, which is what we are used to in the Western world, experience level four living conditions.

To analyze the causes and consequences of poverty, we need, as I said, a lot of data. We need these Y-variables that exist in different dimensions. Because of the lack of geo-temporal data on poverty, scholarship is generally limited regarding why people are stuck in these poverty traps.

What we do at the AI and Global Development Lab is combine AI, Earth observations, and social sciences to analyze the causes and consequences of global development historically, geographically, and globally. Our lab is currently generating such data by combining these different disciplines. What you see here is the City of Cape Town, and we are capitalizing on this variation of development within cities. From 1984 to 2018 and 2020, you see large developments occurring within cities, and this is exactly the variation that we are capitalizing on.

How do we do this? We have points that we gather from existing surveys. We stack them on top of this gridded map and designate different levels of poverty or wealth. The goal is to create the heat map you see to the right. We then temporally stack these slices of data on top of each other. We train a machine learning algorithm on the points where we know there is survey data. We then use the algorithm to impute where the red line is, to see how development has occurred over time and space.

Developing an algorithm for poverty measurement is one of the goals we have embarked on. What we have here—this ‘F’—is the algorithm of interest. For countries such as Burundi, where we don’t have data at certain time points, we input satellite images for this function, and then we generate maps that are hosted on Google Earth Engine.

The structure of the algorithm is roughly what you see here: there are temporal stack models, and you give it input images—daylight images and nightlight images. It processes them through the algorithm, out comes a prediction, and then we connect these with different layers so that the algorithm learns what happens over time in space.

Without our project, we’d be stuck with only two time points where household surveys exist for Burundi (and we’ll do this for the whole of Africa). But with the model that we train, we are able to impute this series of data, which then becomes the data product that we will release to the community.

In the end, we’ll create these living condition maps, or poverty maps, that are hosted on Google Earth Engine. We can also download them for further analysis in different statistical packages. The goal is to train this model and use it for analyzing the causes and consequences of poverty. The approach combines deep learning and different models from demographic and health surveys, combined with Landsat satellite images. Currently, we have generated data that lives on our website and on Google Earth Engine, and we are in the process of developing these models further.

Thank you very much.