search

thinkingmachines / unicef-ai4d-poverty-mapping

UNICEF AI4D Relative Wealth Mapping Project - datasets, models, and scripts for building relative wealth estimation models across Southeast Asia (SEA)
https://thinkingmachines.github.io/unicef-ai4d-poverty-mapping
MIT License
21 stars 9 forks source link
geospatial-analysis machine-learning social-impact

readme

# UNICEF AI4D Relative Wealth Project

Python Code style: black



📜 Description

The UNICEF AI4D Relative Wealth Project aims to develop open datasets and machine learning (ML) models for poverty mapping estimation across nine countries in Southeast Asia (SEA).

We also aim to open source all the scripts, experiments and other artifacts used for developing these datasets and models in order to allow others to replicate our work as well as to collaborate and extend our work for their own use cases.

This project is part of Thinking Machines's overall push for open science through the AI4D (AI for Development) Research Bank which aims to accelerate the development and adoption of effective machine learning (ML) models for development across Southeast Asia.

Documentation geared towards our methodology and experiments can be found here.



💻 Replicating model training and rollout for a country

Our final trained models and their use to produce nationwide estimates can replicated through our notebooks, assuming you've followed the Data and Local Development setup below.

All the output files (models, datasets, intermediate files) can all be downloaded from here.



📚 Data Setup

DHS Data

Due to the sensitive nature of the data and the DHS program terms of use, we cannot provide the raw DHS data used in our experiments.

You will have to request for access to raw data yourself on the DHS website.

Generally, for all the experiment notebooks in this repo, they assume that the DHS Stata and Shape zip files contents are unzipped to its own folder under data/dhs/<iso-country-code>/ where the <iso-country-code> is the two-letter ISO country code.

For example, from the data for the Philippines will have this directory structure:

data/
    dhs/
        ph/
            PHGE81FL/
                DHS_README.txt
                GPS_Displacement_README.txt
                PHGE81FL.cpg
                PHGE81FL.dbf
                PHGE81FL.prj
                PHGE81FL.sbn
                PHGE81FL.sbx
                PHGE81FL.shp
                PHGE81FL.shp.xml
                PHGE81FL.shx
            PHHR82DT/
                PHHR82FL.DCT
                PHHR82FL.DO
                PHHR82FL.DTA
                PHHR82FL.FRQ
                PHHR82FL.FRW
                PHHR82FL.MAP

If you create your own notebook, of course you are free to modify these conventions for filepaths yourself. But out-of-the-box, this is what our notebooks assume.

Night Lights from EOG

The only other data access requirement is for the EOG Nightlights Data which requires registering for an account. The notebooks require the use of these credentials (user name and password) to download the nightlights data automatically.

General Dataset Notes

All the other datasets used in this project are publically available and the notebooks provide the code necessary to automatically download and cache the data.

Due to the size of the datasets, please make sure you have enough disk space (minimum 40GB-50GB) to accommodate all the data used in building the models.



⚙️ Local Setup for Development

This repo assumes the use of miniconda for simplicity in installing GDAL.

Requirements

  1. Python 3.9
  2. make
  3. miniconda

🐍 One-time Set-up

Run this the very first time you are setting-up the project on a machine to set-up a local Python environment for this project.

  1. Install miniconda for your environment if you don't have it yet.

    wget "https://repo.anaconda.com/miniconda/Miniconda3-latest-$(uname)-$(uname -m).sh"
bash Miniconda3-latest-$(uname)-$(uname -m).sh

  2. Create a local conda env and activate it. This will create a conda env folder in your project directory.

    make conda-env
conda activate ./env

  3. Run the one-time set-up make command.

    make setup

  4. To test if the setup was successful, run the tests. You should get a message that all the tests passed.

    make test

At this point, you should be ready to run all the existing notebooks on your local.

📦 Dependencies

Over the course of development, you will likely introduce new library dependencies. This repo uses pip-tools to manage the python dependencies.

There are two main files involved:

When you add new python libs, please do the ff:

  1. Add the library to the requirements.in file. You may optionally pin the version if you need a particular version of the library.

  2. Run make requirements to compile a new version of the requirements.txt file and update your python env.

  3. Commit both the requirements.in and requirements.txt files so other devs can get the updated list of project requirements.

Note: When you are the one updating your python env to follow library changes from other devs (reflected through an updated requirements.txt file), simply run pip-sync requirements.txt

📜Documentation

We are using Quarto to maintain the Unicef AI4D Relative Wealth documentation site.

Here are some quick tips to running quarto/updating the doc site, assuming you're on Linux.

For other platforms, please refer to Quarto's website.

☸️Running in Docker

We have created a docker image (ghcr.io/butchtm/povmap-jupyter) of the poverty mapping repo for those who want to view the notebooks or rollout the models for new countries and new data (e.g. new nightlights and ookla years)

To run these docker images please copy and paste the following scripts to run on your linux, mac or windows (wsl) terminals:

curl -s https://raw.githubusercontent.com/thinkingmachines/unicef-ai4d-poverty-mapping/main/localscripts/run-povmap-jupyter-notebook.sh > run-povmap-jupyter-notebook.sh && \
chmod +x run-povmap-jupyter-notebook.sh && \
./run-povmap-jupyter-notebook.sh
curl -s https://raw.githubusercontent.com/thinkingmachines/unicef-ai4d-poverty-mapping/main/localscripts/run-povmap-rollout.sh > run-povmap-rollout.sh && \
chmod +x run-povmap-rollout.sh && \
./run-povmap-rollout.sh
curl -s https://raw.githubusercontent.com/thinkingmachines/unicef-ai4d-poverty-mapping/main/localscripts/copy-rollout-to-local.sh > copy-rollout-to-local.sh && \
chmod +x copy-rollout-to-local.sh && \
./copy-rollout-to-local.sh

Note: These commands assume that curl is installed and will download the scripts, change their permissions to executable as well as run them. After the initial download, you can just rerun the scripts which would would have been downloaded to your current directory.

Note: The scripts create and use a docker volume named povmap-data which contains the outputs as well as caches the data used for generating the features from public datasets

Note: Rolling out the notebooks requires downloading EOG nightlights data so a user id and password are required as detailed in the previous section above.