Image

Documentation

Learn how to use our API and UI and more.

graphical divider

The Wealth Data Layer

An introduction to the science behind Wealth estimation, why this matters and how you can access this data layer.

Overview

The Wealth layer displays the estimated prosperity of an area, where this value is a function of diverse indicators such as asset ownership, housing quality and infrastructure availability. The underlying data which supports this layer are the Demographic and Health Surveys (DHS) conducted in Africa from 1994 to 2016. This data layer provides an insight which is both a critical indicator of well-being and safety in its own right as well as being causally linked to many different interventions and treatments. Atlas AI’s wealth layer has been used

  • Helping the World Bank and Ethiopia target social welfare programs more accurately.
  • As a Business Intelligence tool by a major African telco to help plan business expansion.
  • Studying the relationship between Wealth and other covariates of economic or sociological interest.

Data availability

Atlas AI’s Wealth layer is trained on LANDSAT imagery and produces estimates at a 2 kilometer resolution. Estimates are provided for four year periods from 2003 to 2018 to give a total of 4 datasets. Since the Wealth layer is fundamentally a measure of human prosperity, it has been filtered to exclude areas with no or minimal permanent human presence such as lakes, deserts or rainforests. This filtering is accomplished via use of Global Human Settlement Layer maps. The layer can also be aggregated up from this pixel-level to district, provincial or national level. This aggregation is done by taking the weighted average of all pixels within the larger aggregation level: pixels filtered out due to having no population are ignored.

User Type Geographic scale Temporal scale Granularity
Demo access Kenya-subregion 2003-2006 4 years / 1912m
Registered free access Kenya 2003-2006 4 years / 1912m
Full paid access Africa 2003-2018 4 years / 1912m

Details

Wealth: Quantifying the wealth of households in a way that is consistent and useful over the entirety of Africa is a complex problem: it is not as easy as simply asking about the annual income of a household in the local currency. The methodology will need to account for self employed individuals, farmers and small shop owners as well as economies where wealth might be primarily found in non-currency means such as livestock. DHS surveys provide an excellent source to turn to since they are by their nature designed to have a consistent set of questions that are applicable across a wide set of circumstances and have been collected since 1994. The questions in the survey, many of which have remained consistent over the entire time period the survey has been collected. Broadly speaking there are two sets of variables of interest to us in the DHS surveys. The first are questions about the dwelling itself such as whether it has running water, how it is heated, whether there are separate rooms for activities such as cooking or clearing. The second are asset ownership questions such as whether the household owns livestock, a motorcycle or a refrigerator.
The results for many households in a cluster are then averaged by the survey takers for a cluster of households and then released along with an approximate latitude and longitude. The next step is to take the diverse set of categorical responses to the aforementioned questions and convert them into a single number that roughly corresponds to the wealth implied by their responses. This is accomplished by a process known as Principal Component Analysis (PCA) which allows us to capture as much of the variation as possible into a single variable. This value, a weighted sum of answers to the DHS question, is known as the wealth composite.

Machine learning: Now that we have a single label and an approximate location for each DHS cluster, we can input this data set into our machine learning pipeline. Aggregating geospatial signals for a variety of sources, including visual satellite data as well as other spectral information, provides us the features for each cluster. Having a label and feature set for each cluster in the DHS set allows us to start training using a Deep Neural Net architecture which generates a model that is able to predict the wealth of location given only geospatial information for a given three year period. We then predict this value over the entirety of the African continent for several different three years spans. Instead of returning the wealth composite, a difficult number to interpret, we return the decile of this value falls into relative to the entire distribution of wealth values from 2003 to 2018.

API access

For fuller documentation on the Atlas AI API click here

Data layers in the Atlas AI API are identified by numeric ids, the id for the Wealth layer is 1. Using this id as a parameter allows you to find all available Wealth datasets:

curl --header "Content-Type: application/x-www-form-urlencoded" \
--header 'token: yourapitoken' \
-G \
--data-urlencode 'indicatorId=1' \
https://data.atlasai.co/api/v1/datasets
                    

This will return all the distinct datasets relevant to the Wealth data layer. For example, one of the elements in this response might be the Wealth dataset for the most recent time period as shown below:

[
  {
    "id":189,
    "dataType":"vector",
    "indicator":"Wealth",
    "aggregationLevel":"District",
    "version":"v1.1.0",
    "timePeriod":"2015-2018",
    "available":true
  }
]                   

Once you have established the id of the Wealth dataset you are interested in downloading the data is the same as for any other data layer, as detailed in the API documentation.

Icon For Arrow-up