Every year natural disasters displace millions of people from their homes. Humanitarian organizations often lack accurate data to quantify how many people have been displaced, as well as where these population could be located. This hinders the capabilities of humanitarian organizations to respond effectively to the needs of internally displaced people. As climate change increases the frequency and severity of natural disasters, response organizations are eager for improved data sources to better understand how many people are displaced, where they have been displaced to, and when they are able to return home. Facebook is uniquely positioned to this data gap and help increase the effectiveness of humanitarian response of food, medical and housing aid while preserving user privacy. Displacement Maps estimate how many people were displaced by a crisis longterm, and where those people are in the period after the event. In particular, we identify a person as displaced if we see an anomaly after the event compared to that person’s pre-crisis movement patterns at night. The methodology has three steps: a) detect those people that live in the affected area and analyze typical locations and movement patterns at night, b) understand who of those people are displaced two weeks after the onset of the disaster, c) estimate the number of displaced people each day at the city level.
Image: Displacement Map from Japan
Detecting users’ home and normal movement patterns
In the month before a natural disaster we estimate a user’s modal location (most likely their home) during night hours; this is our best estimate of what represents their home. Additionally, we calculate the mean-squared displacement (MSD) of a person’s activity with respect to their modal location, which represents, on average, how much they travel away from their home. These two metrics allow us to understand on average what people’s most common locations are, as well as how much on average they travel away from home in the absence of a crisis. This step is central in order to help us identify if a later date the crisis disrupts a person’s normal movement, and forces them to relocate long term. In order to have greater accuracy in our estimations, we restrict our sample to those in the bounding box that have used the Facebook mobile app with Location History turned on for at least 15 days in the pre-period.
In order to measure the number of people in the affected area that have been displaced we start by looking at the two-week period right after the crisis. Then we estimate, similar to the pre-period, a user’s home and mean-squared displacement with respect to their pre-crisis home. From these metrics we define people’s status as:
- Displaced: (1) If their post-crisis location is at least 2 kilometers away from their pre-crisis home AND (2) If their average distance from their pre-crisis home at night is double their usual value on more than half the days we observe them in this two week period.
- Never displaced: (1) If their post-crisis location is less than 2 kilometers away from their home OR (2) If their post-crisis movement patterns are less than double than their pre-crisis movement patterns on more than half the days we observe them in this two week period.
- Unknown: If they have connected fewer than 3 days since the crisis.
Starting on day 15 after the crisis, we produce daily updates of the population status to count the number of people displaced and returned, both overall and aggregated to administrative regions. In particular, we check the status for those with status “Displaced” or “Unknown”.
- For Unknown people, once we have seen them for 3 days in a row either close to home, or away from home, we classify them as Never Displaced or Displaced.
- For Displaced people, once they are observed for 3 days in a row less than 2 kilometers away from their home we classify them as Returned.
- For the population that has been labeled as Never Displaced, we do not check their subsequent movement patterns in order to avoid confounding displacement estimates by including movements associated with unaffected people traveling for unrelated reasons.
These estimates are aggregated at the city level to protect users’ privacy and provide a clear signal. Additionally, if a city to city movement count is less than 10 people, we do not report the destination city, and instead aggregate to a higher level, such as country or overall.