Back to cover

Poverty Mapping: Innovative Approaches to Creating Poverty Maps with New Data Sources

Chapter 5 | Daytime and Nighttime Remote Sensing Imagery Data

This section discusses the methodological implications of using remote sensing imagery as a data source for generating poverty maps.

Definition

Remote sensing is the process of observing, monitoring, and detecting the physical characteristics of an area from a distance, using sensors located in aerial platforms (such as satellites, aircrafts, and drones). For generating poverty maps, the most commonly used type of remote sensing data are satellite data, including both daytime and nighttime data. Both types of satellite images have a raster data format, and therefore consist of a numerical gridded representation, where each cell of the grid is associated with a specific geographical location.

A considerable advantage of this type of data is its large temporal and geographic reach. Daily satellite images are available from reliable sources for the whole world from approximately the 1980s to the present. Furthermore, satellite data can be disaggregated into small areas—such as cities or villages—depending on the image resolution. This results in consistent and comparable data suitable for conducting meaningful comparisons across time and geographies.

Data Sources

Low- or medium-resolution daylight satellite imagery data are available from government programs such as Landsat (National Aeronautics and Space Administration / United States Geological Survey) and Sentinel (European Space Agency) from 1972 onward. Although these data are collected daily, images with high cloud coverage are typically discarded. Higher-resolution satellite data are available for purchase from specialized vendors, but publicly available satellite data are typically sufficient to generate poverty maps.

Most common sources of nighttime satellite data include the Defense Meteorological Satellite Program, available for 1993–2017, and the Visible Infrared Imaging Radiometer Suite, available for 2012 to the present. These data sets provide an estimate of radiance, defined as a stable measure of brightness as seen from space, filtered to remove extraneous features such as biomass burning and aurora. Nighttime light data are available in different series, depending on the sensor that was used to capture the image. Although data from different series can be combined to perform time series analysis, this requires prior calibration of the data.

Methods

There are essentially three types of models that use machine-learning techniques to derive poverty proxies using remote sensing (for example, satellite image) data. The simplest isa feature-based prediction model that uses quantifiable geospatial features (such as the number of buildings in a region, the length of roads, or points of interest). This type of model typically relies on random forest regression to generate a prediction.1

The second type of model is an image-based prediction model that extracts geospatial features directly from remote sensing imagery. These models rely on deep-learning algorithms such as CNNs, which are particularly suited to working with images.2 The third type of model combines the two approaches described above: geospatial features are combined with satellite imagery to estimate a poverty-related proxy.

Applicability Considerations

There is strong potential for the use of satellite imagery to derive poverty maps in the context of evaluation. Both daytime and nighttime satellite imagery data are freely available from reliable sources at a low or medium resolution, which is sufficient to create poverty maps. However, a higher level of expertise is needed to develop poverty maps using satellite imagery. This includes experience in manipulating Earth observation data, machine learning (specifically deep learning and CNNs), and programming.

Additional computing resources (such as access to graphics processing unit clusters) might be needed to use this approach, especially when using CNNs. Access to graphics processing unit clusters must be purchased, with plans typically priced per hour of use.

All parts of the analysis can be completed with a combination of Excel, Python (open source), and geospatial software such as QGIS (open source). Different research centers have also made publicly available the Python code used to create their poverty maps, which would be an excellent starting point for developing new maps.3

Examples

Example 1: Poverty Maps Using Geospatial Features

Tingzon et al. (2019) developed a poverty map of the Philippines using geospatial features extracted from OpenStreetMap (a crowdsourced online mapping platform). The features used in the model included roads, buildings, and points of interest (such as parks, schools, hospitals, and cinemas). These features were extracted within a 5 kilometer radius for rural areas and a 2 kilometer radius for urban areas.

A random forest regression model was trained on these features, both separately and jointly, to predict socioeconomic well-being. The authors found that using roads, buildings, or points of interest alone already explained 49–55 percent of the variance, with roads being the best predictor (R2 = 0.55). Training a model on all three types of OpenStreetMap features resulted in a slightly higher R2 value (0.59). Furthermore, the authors tested the performance of the model by adding nighttime lights data. This resulted in an increased R2 (0.63) for wealth prediction.

Example 2: Poverty Maps Using Satellite Imagery

Yeh et al. (2020) trained deep-learning models to predict survey-based estimates of asset wealth across approximately 20,000 African villages using temporally and spatially matched multispectral daytime satellite imagery and nighttime lights data (30 meter per pixel Landsat and < 1 kilometer per pixel nighttime lights imagery). The authors trained a CNN—ResNet 18 architecture—to predict village- and year-specific measures of wealth. The main objective was for the machine-learning model to identify those features present in daytime and nighttime satellite imagery that are predictive of asset wealth.

The study found that a deep-learning model trained on this type of imagery data can explain approximately 70 percent of the spatial variation in asset wealth across Africa and up to 50 percent of the changes in wealth over time when aggregating the village-level data to the district level. Notably, CNNs trained only on nighttime lights or only on multispectral daytime imagery performed similarly to each other and almost as well as the combined model, suggesting that these two inputs contain similar information, at least for predicting spatial variation in wealth in Africa. This model also outperformed simpler models based only on geospatial features.4

But deep-learning models tend to be less interpretable than other machine-learning approaches. Deep neural network constructs typically combine multiple hidden layers and neurons, resulting in millions of features. From the perspective of human interpretability, it is very difficult to track the complex interactions that occur among the features underpinning the model’s output. Although there have been advances in the academic literature toward more interpretable models (through fields such as interpretable artificial intelligence), deep-learning models continue to be considered a “black box.”

Example 3: Poverty Maps Using a Combination of Geospatial Features and Satellite Imagery

Puttanapong et al. (2020) developed a model to predict the spatial distribution of poverty in Thailand based on the integration of multiple data sources, including geospatial features and features extracted from satellite imagery. Specifically, the study used the following sources: land surface temperature, normalized difference vegetation index, intensity of lights, geocoded data on built-up and non–built-up areas, geocoded human settlement data, land cover maps, and crowdsourced data from OpenStreetMap (road count, road length, points of interest, and built-up areas).

The study applied several computational techniques to examine the relationship between geospatial features and the proportion of people living below the poverty line using conventional methods of estimating poverty levels. The methods applied in the study included generalized least squares, neural networks, random forest, and support vector regression. Results suggested that intensity of night lights and other variables that approximate population density are highly associated with the proportion of an area’s population who are living in poverty. The random forest technique yielded the highest level of prediction accuracy among the methods considered in this study, perhaps because of its ability to fit complex association structures even with small- and medium-size data sets.

  1. Random forest is a supervised machine-learning algorithm that combines predictions from multiple decision trees using an ensemble approach.
  2. Deep learning encompasses supervised machine-learning algorithms, which mimic the structure of the human brain by using multilayered neural network architectures.
  3. See, for example, the data available at the following webpages: https://github.com/sustainlab-group/africa_poverty, https://github.com/nealjean/predicting-poverty, https://github.com/jmather625/predicting-poverty-replication.
  4. A similar approach has been developed by Stanford University’s Sustainability and Artificial Intelligence Lab, which has been tested across several African countries (Jean et al. 2016). Other studies have also built on this approach, such as Babenko et al. (2017), Tingzon et al. (2019) and Heitmann and Buri (2019).