Here’s how we used satellite data to map poverty in Thailand and the Philippines
Innovative, new methods are needed to estimate poverty due to the high costs and long time frame of traditional methods.
Accurate measurements of the economic characteristics of populations critically influence both research and policy. Such measurements shape decisions by governments about how to allocate scarce resources and provide the foundation for global efforts to understand and track progress toward improving lives.
Household and enterprise surveys, censuses, and administrative registers constitute major sources of development indicators, including that of the Sustainable Development Goals. There is usually information available at national and regional level from these sources, however, policymaking often requires more granular data in order to efficiently tackle disparities within countries.
Given the high cost of conducting surveys, increasing sample size to reflect all geographic areas and different population groupings may not be practical. On the other hand, while censuses don’t suffer from limited coverage, they provide outdated information since they are expensive to conduct and seldom done. Complementing these traditional sources of data for development with innovative methods could potentially address the issues mentioned above.
Neural networks, which use algorithms to mimic how the human brain operates and recognize relationships in a set of data, could be one of these methods to help policymakers obtain the granular poverty data they need. These networks can combine night and daytime satellite images in order to estimate economic wealth on a highly granular level.
We tested this approach using poverty data from the Philippines and Thailand. We chose these countries because they have existing initiatives to combine household survey data with census data to produce more granular yet reliable estimates of poverty, and these initiatives provided sufficient data on which machine learning algorithms could be trained.
The two countries, although both considered to be middle-income countries, have slightly different poverty profiles, with Thailand showing significantly lower poverty rates than the Philippines. This allows for an examination of how poverty distribution affects poverty estimation using machine-learning algorithms.
It is possible to predict poverty on a grid-level as small as an area 4 kilometers by 4 kilometers.
We trained the algorithm to predict nighttime light intensity based on daytime images. Intensity of night lights is a reasonable proxy for human settlements and economic wealth. The goal was to get a model that can recognize features in daytime images that lead to bright nights. These features correspond to things like street patterns, building density, and roofs, etc.
We obtained those features on the grid-level for the whole country from the daytime images. We used aggregates (from grid-level to municipal/city-level in the Philippines and tambon (township) level in Thailand) of those features to analyze their relationship with poverty rates at those levels, which represent the most granular level of our input data. Once that relationship has been established, it is possible to predict poverty on a grid-level as small as an area 4 kilometers by 4 kilometers.
We were able to predict the government-published poverty estimates using machine-learning algorithms applied on satellite imagery much better in the Philippines than in Thailand. One possible reason for this is there are significantly more tambons with low poverty rates.
The lack of variability in our poverty data may have contributed to the machine-learning algorithm’s underestimation of poverty distribution in Thailand. However, when we used proxy measures other than income poverty rates, which have greater variability, as our input data, the machine-learning approach’s predictive performance improved significantly.
Exploring the feasibility of using satellite imagery as an alternative data source for poverty estimation does not aim to replace conventional sources of poverty data: rather, data integration addresses some of the limitations associated with traditional techniques.
To achieve these objectives, which might also result in significant cost savings, government agencies that are responsible for compiling poverty statistics must make substantial investments in technology and talent. For instance, scaling up from feasibility studies to ongoing and more systematic use of nontraditional data sources requires these agencies to have access to higher resolution satellite imagery and specialist computing resources.
Further extending this to other types of non-traditional sources of data for development, any initiative to integrate these new types of data into national statistical systems will require the forging of partnerships with academic, private, and public institutions to ensure a broad platform for sharing ideas, knowledge, and solutions on how to leverage innovative data sources for the benefit of all.