Home

Awesome

[Deprecated]Poverty Prediction by Combination of Satellite Imagery and Machine Learning

The project is no longer maintained. Some files are not able to fetch from source due to some reaseons

Is it feasible for you to estimate the standard of living or energy consumption of a site based night-time satellite imagery? This means whether we can do the satellite imagery processing for socioeconomic analysis?

Let's find out in this project!!!

Key Steps

  1. Download satellite night lights images from NOAA
  2. Download DHS data for Rwanda
  3. Test whether night lights data can predict wealth, as observed in DHS
  4. Download daytime satellite imagery from Google Maps
  5. Test whether basic features of daytime imagery can predict wealth
  6. Extract features from daytime imagery using deep learning libraries
  7. Replicate final model and results of Jean et al (2016)
  8. Construct maps showing the predicted distribution of wealth in Rwanda

A night time satellite imagery:

alt text

Download nightlights images

Go to the DMSP-OLS website and download the satellite nighttime luminosity data (roughly 400MB). Here we are using the file F182010.v4d_web.stable_lights.avg_vis.tif.

Download Rwandan DHS and construct cluster-level aggregates

Demographic and Health Surveys (DHS) are nationally-representative household surveys that provide data for a wide range of monitoring and impact evaluation indicators in the areas of population, health, and nutrition. For this assignment, you will need to download the 2010 Rwandan DHS data. This requires registration Do not forget to request for the GPS dataset.

The immediate goal is to take the raw survey data, covering 12,540 households, and compute the average household wealth for each survey cluster (think of a cluster as a village). Refer to the file Recode6_DHS_22March2013_DHSG4.pdf for information on these data.

Saved output is rwanda_cluster_avg_asset_2010.csv. NOTES:

the cluster locations, overlaid on the nightlights data, are shown in the figure below. <img align="center" src="figure/map1.png" alt="Map" style="width: 400px;center;"/>

Merge nightlights and DHS data at cluster level

NOTES:

<img src="figure/scatter2.png" alt="Map" style="width: 400px;"/>

Download daytime satellite imagery

We will use the Google Static Maps API to download satellite images. Refer Google Static Maps introduction and Google Static Maps API Usage Limits. You must apply for an API key before downloading.

Download the images from Google at zoom level 16 (pixel resolution is about 2.5m). Set the image size to be 400 pixels X 400 pixels, so that each image you download will cover 1 square kilometer. In this way, each daytime image you download will correspond to a single pixel from the nighttime imagery from Step 1 above.

NOTES:

Construct a high-resolution map of the distribution of predicted wealth

Use this model to calculate the predicted wealth of every one of your original images. Create a heatmap showing the distribution of predicted wealth in Rwanda.

<img align="center" src="figure/pmap.png" alt="Map"/>