cyuancheng / Predicting_Poverty

Poverty Prediction by Satellite Images using Deep Learning

Poverty Prediction by Satellite Images and Deep Learning

Capstone Project 2 (Springboard - Data Science Career Track)

Chiyuan Cheng (08/2020)

Summary

• This project uses transfer learning to predict poverty (Wealth index) of a sub-Saharan African country, Burundi, in 2010.
• Regression models are used to predict Wealth index from luminosity of nighttime satellite images, with r-squared of 0.54.
• Gaussian Mixture Model is used to classify the daytime satellite imagery into three classes, based on the luminosity.
• Transfer learning (VGG16, ResNet50, Inception C3) to capture features from daytime satellite imagery and predict poverty, with the 80% accuracy from the best model.

Report

• Presentation slide
• Final report

Jupyter Notebook

1. Data Access
2. Data Prep (Nightlight Satellite Image)
3. Machine Learning (Nightlight)
4. Data Processing (Daylight Satellite Image)
5. Data Analysis (Simularity
6. Deep Learnining