poverty-predictor

Night Lights Data 2016 -> https://earthobservatory.nasa.gov/Features/NightLights/page3.php

Overview

Machine learning, data science, and satellite imagery used to map areas displaying indicators of poverty. The purpose of this excercise is to apply a remote sensing research case in order to gain experience in this particular python development environment. This project applies techniques established in a 2016 study from Stanford University.

Description of methods

The foundational study uses World Bank Data and computer vision to complete the poverty prediction research. View the original code here. My course of analysis does not include some of these techniques, and relies on a two step method of classification and comparison. Here is an overview of the steps involved:

I. Classification

• Daytime image - Label pixels 1-urban or 0-non-urban
  • Segment the image using the scikit-learn quickshift and felzenszwalb modules
  • Train the scikit-learn RandomForestClassifier model using training data to classify segments
  • Bin the data into urban and non-urban classes
• Nighttime image - Label pixels 1-light and 0-dark
  • Train the scikit-learn RandomForestClassifier model using training data to classify pixels
  • Bin the data into light and dark classes

II. Comparison

• Subtract the Daytime image from the Nighttime image

Outcome	Score	Meaning
1-urban - 1-light	0	expected
0-non-urban - 0-dark	0	expected
1-urban - 0-dark	1	poverty is predicted
0-non-urban - 1-light	-1	false positve (ex. fire)

The quality of the classification is not the primary focus of this excercise, though an accuracy assessment of both classified images is included. I am limited to producing the training datasets and wanted to prioritize building a pipeline, including accuracy assessments, for the process using this tech stack. Parameters of training the model were optimized, however the training-data is a limiting factor.

Data Sets

• Night-lights VIIRS Imagery
• Day-time Landsat Imagery

Dependencies

• Descartes Labs library
• Jupyter Notebook
• Python 2
• NumPy
• gdal 2.0
• scikit-learn
• scikit-image

REFERENCES

• Python for geospatial data processing by Carlos De La Torre
• Python for Object Based Image Analysis (OBIA) by Carlos De La Torre
• COMBINING SATELLITE IMAGERY AND MACHINE LEARNING TO PREDICT POVERTY, Neal Jean, Marshall Burke, Michael Xie, W. Matthew Davis, David B. Lobell, Stefano Ermon (Stanford University, August 2016)
• Open-geo Tutorial by Chris Holden