A team of researchers at MIT has recently studied how satellite imagery can aid international development projects. Their work has been published in Big Data. The team developed algorithms with machine-learning techniques in order to automate the identification of houses in rural areas. They also created a system for mapping villages and determining the optimum location for the placement of solar panels. This work has been previously conducted by field teams, but the researchers’ technological advancements have improved the efficiency of the process and lowered its cost. The researchers applied their algorithms to two projects.
The first project involved choosing villages in sub-Saharan Africa for a cash grant program aimed at improving living standards in low-income rural areas. The grant agency prioritized the poorest villages, and the metric it employed to determine the poorest villages is the proportion of houses with thatched roofs to metal roofs. The agency gave higher priority to villages with a greater proportion of thatched roofs to metal ones, as the metal roofs are more expensive. The researchers expanded on their house identification algorithm by having it distinguish metal roofs from thatched ones. They were able to accomplish this by incorporating reflectance measurements, which are greater for the metal roofs.
The second project involved choosing villages in rural India for installing microgrids to provide electricity from battery-storage systems and solar panels. The researchers looked for optimum sites for the panels by finding the most efficient network configuration for distributing power. The team’s computers ran thousands of different possibilities of where to place the solar panels and distribution wires. By looking at the variations, they could determine which configurations would give electricity to the greatest number of houses with the least amount of wiring needed.
In order to develop their algorithms, the researchers had to start by manually selecting houses from satellite images to enter training data for a machine-learning system. Such a system generalizes the criteria for determining what is a house and what is not a house, so when it is shown a new image, it can recognize the houses in that image. As the system receives more examples of houses, it detects houses better. The team notes that their algorithms could potentially be used for a variety of other purposes. For example, as there is not much data on population changes in India, the algorithms could help determine which areas in India have gained or lost population over time.