Satellite imagery resources

Satellite imagery is an exciting resource for development economists. It allows to study a variety of outcomes, such as agricultural land use, economic activity, household welfare or crop yields. While one cannot hope to get the depth of information that a household survey provides, imagery lets one study outcomes over large regions and over extended periods of time. Below are some resources that I find useful when developing projects.

Satellite Imagery

There are many providers of raw satellite imagery. They differ by their spatial resolution, the frequency of images, the number of spectral bands, the depth and density of their archives, and their cost.

  • Landsat and Sentinel-2 have been used to study relatively large scale issues like urbanization or deforestation or land-use. Their archives are deep (from the 1970s and 2015, respectively) and they are for free, but their relatively low resolution (30m and 10m) and frequency (18 and 5 days) limits their applicability for many questions.

  • Planet offers two sets of imagery with much higher resolution: PlanetScope imagery covers the globe every day at 3m resolution, and has been captured since 2016. Researchers can apply for a very generous quota, and the resolution is high enough to be used for crop yield estimation (if you know the crop and the plot), detecting regular events (such as markets) or rooftop material identification. Skysat has an even higher resolution at <1m, but - apart from some great sample imagery in Google Earth Engine - is only acquiring imagery upon request. If you're lucky, someone else has collected imagery in your area of interest before. Researchers can (sometimes) apply for a quota of new or archived acquisitions.

In the image above, I used Skysat imagery and Random Forest classification to find market areas and vehicles at a sample location in Kenya.

Software & Processing

A satellite image is basically a raster with a spatial registration and information on the value of each spectral band in each pixel. GIS software, such as ArcGIS or QGIS can handle small amounts of images easily, but may struggle with analyses on a larger scale. Google Earth Engine is great for processing larger stacks, has a lot of inbuilt functions for e.g. machine learning and is very well documented. Using it is for free and some of the datasets listed above are easily accessible from within the tool.

If your project grows and you need to handle larger data quantities or host processing on a server, Google offers free access quotas to the Cloud through its Education program.

Other datasets

Besides raw imagery, there are also a couple of datasets that are derived from satellite imagery and may be useful for development economists:

  • Facebook uses machine learning to predict for each 30m pixel in Africa whether it contains a house.

  • NASA FIRMS detects fires - natural or man-made, for example for clearing land - at 375m pixel resolution every night