Skip to content

Latest commit

 

History

History
48 lines (29 loc) · 3.88 KB

README.md

File metadata and controls

48 lines (29 loc) · 3.88 KB

summary | usage | walk through notebooks | license

Lumin Human

License

Context

Night-time lights (NTL) data from earth observation satellites is an alternative data source for research on socio-economic indicators in cases when national or large-N survey data is unavailable, uncertain, or cost-prohbitive. Unsettled questions, new satellites, and novel machine learning techniques make this a fruitful area for continuing research on a range of questions with significant social, economic and socio-environmental implications.

Scientific objective

Explore recent techniques and tools that utilize night-time lights (NTL) satellite data in conjunction with socio-economic indicators.

Scientific question

Can NTL satellite data help predict poverty indicators in rural regions of five countries in the Sub-Saharan Africa region?

Approach

A recent paper proposing novel approaches involving NTL data serves as a framework and guide for our approach: Jean et al, 2016. Combining satellite imagery and machine learning to predict poverty. Science, Vol. 353, Issue 6301, pp. 790-794, DOI: 10.1126/science.aaf7894

Data understanding

The prediction target is a geo-located indicator of mean annual economic well-being at the unit of a village (rural) or census tract (urban). The underlying data source is household consumption expenditures (spending) - a determinant of employment, poverty and health outcomes - from the World Bank Living Standards Measurement Study (LSMS), which is available at this link.

We initially focus on two potential sources of NTL data. First, the NOAA Defense Meteorological Satellite Program - Operational Linescan System (DMSP -OLS) is the default data source for existing NTL data in socio-economic research, and is available at this link. Second, we explore the Day/Night Band (DNB) of the Visible Infrared Imaging Radiometer Suite (VIIRS) of NASA/NOAA's Suomi National Polar-orbiting Partnership satellite, which offers potentially higher-quality data, in expectation of the release in 2019 of pre-processed data in a light-noise adjusted form usable for the present research task. The VIIRS dataset is described at this link.

Data understanding - Results

Both NTL data sources (DMSP and VIIRS) are extracted from Google Earth Engine. The script in Google Earth Engine can be found at this link. Replace with your own local drive pathways and file names.

Daytime satellite image data will subsequently be acquired for use in conjunction with the NTL data.

Data processing

We aggregate the household-level consumption expenditure values to the community level and plot these values on the NTL image.

Data processing - Results:

Notebook in nbviewer

Modeling

We envision modeling tasks that proceed from the more simple baseline comparison models to the more elaborate three-stage modeling worfklow (object detection - CNN - ridge regression) with enriched and expanded data that is developed in the above paper.

Results - slides

Evaluation

We envision the evaluation metric to be rooot mean squared error (RMSE), as used in the above paper.

Deployment

We envision the deployment of the model for user tasking through a basic interactive web application.