Slack Channel: #p-predicting_poverty
Project Description: Poverty prediction based on research out of Stanford University. The key idea is to predict poverty at a fine grained level using machhine learning techniques(deep learning) and more specifically Transfer Learning. The current focus is on replicating results from the paper. I have begun work on predicting poverty in the country of Rwanda and the extremely general steps for doing so are as follows:
- Downloading the survey data from the Demographic and Health Surveys website. (There is an element of requesting the data from the source, and as a result you will have to sign up to the website and then request demographic data for Rwanda.
- Downloading night light satellite imagery from NOAA.
- Downloading satellite imagery via google maps static API.
- Moving downloaded satellite imagery to corresponding night light values folders.
- Extracting features of day time satellite imagery to predict night time light intensity using deep learning.
- Replicating the results of the Stanford research (Performing ridge regression to estimate wealth).
- Constructing high resolution maps of predicted data.
The immediate aim of the project is to replicate the results and create a scalable pipeline that can be applied to various other coutries, because extracting image features in different countries will likely yield a very different feature spaces which may not be applicable globally.
** Maintainers **
- chrikeli
- [YOU]
Deep Learning Stuff:
- chrikeli
- [YOU]
- Join the Slack Channel
- Read the paper to get a proper understanding of the problem.
- Go through the [Refined-Solution.ipynb] file to get a better understanding of the approach.
- If you have ideas, or suggestions feel free to send @chrikeli a message on Slack.
- If you are new to GitHub, here are some baby steps to get comfortable with contributing to a D4D project.
- Work on your code and submit a pull request to add it to the project! Reach out for help anytime!
- Beginners are welcome! To do a lot of the analysis bits, you don't need prior data science or deep learning experience and the code for it is fairly well documented.
- Deep Learning You will need to have access to a GPU for a lot of the image processing we will be doing, you may get away with doing it on your CPU, but it may take a few days (compared to a couple of hours via GPU).