San411/Predicting-Poverty-Replication

Predicting poverty using satellite images for African Countries # predicting-poverty-replication The purpose of this repository is to replicate the Jean et. al. (2016) paper (see `papers/jean_et_al.pdf` and `papers/aaai16.pdf`) using only Python3 and PyTorch. These more up-to-date tools and instructions should help anyone trying to recreate and build-off this work. The purpose of the paper was to use abundant sources of data (namely satellite imagery and nightlights data) to predict poverty levels at a local level (on the order of a single village). For some background, every few years the World Bank conducts surveys in developing countries to understand their living situations. As you might expect, this process is very time-consuming. If we can make a model that only uses abundant sources of data to predict values that otherwise have to be measured through expensive human efforts, then several possibilities arise: 1) prediction during "off-years" (when no surveys are collected) 2) real-time monitoring of poverty conditions 3) potential for early-warning systems Note 1: all scripts are put in Jupyter Notebook (.ipynb) files to encourage exploration and modification of the work. Note 2: I chose to work with Mali as it had the most usable data in 2014-2015 # Reproduction Result Actual Consumption vs Predicted Consumption # Setup I recommend creating a virtual environment for this project. I prefer using Anaconda. First run: ``` git clone github.com conda create -n python=3.7 pip gdal conda activate conda install pytorch==1.3.0 torchvision==0.4.1 -c pytorch pip install -r requirements.txt ``` The libraries that are most likely to fail are gdal and geoio. If a requirement fails to install, first make sure you follow this install procedure exactly. Using `pip` to install GDAL did not work for me, and the only way I got it to install was by including it when I first make the conda environment (hence `pip gdal`). There are also several Stack Overflow posts on these issues, …