simonlazarus/Predicting-Poverty-with-Satellite-Images

We develop a model that takes the geographic coordinates of a village in Africa and predicts whether that village suffers from extreme poverty. Our model uses publicly available satellite imagery along with convolutional neural networks and ensembling to catch 79% of cases of extreme poverty while maintaining an overall accuracy of 85%. # Predicting-Poverty-with-Satellite-Images ## Authors: Stanley Azuakola, Atigon Hongchumpol, Simon Lazarus --- # Executive Summary We aim to predict whether a given village in Africa suffers from extreme poverty by using nothing other than the geographic coordinates of that village and satellite imagery that can be freely and automatically downloaded using those coordinates. We train convolutional neural networks and use ensemble modeling techniques to develop a production model that correctly identifies extreme poverty in 79% of cases in which it appears (Recall = .79) while maintaining an overall predictive accuracy of 85% across all cases. # Contents ### Jupyter notebooks 1. Data collection 2. Extracting images and other data from downloaded files, and cleaning them 3. Exploratory data analysis of non-image data 4. Baseline modeling that does not use satellite imagery 5. Neural network modeling and training 6. Final modeling and analysis of our production model's performance If using our pre-downloaded and pre-exported data, you can skip notebooks 1 and 2. --- ### Link to Google Drive for this project: drive.google.com This project uses several gigabytes' worth of satellite imagery and other data. See the Google Drive to access these data, or replicate the steps in our first and second notebooks to download and export the data for yourself (this will take at least 12 hours). --- ### Software Requirements - Python (Numpy, Pandas, Scikit-Learn) - Jupyter (Ideally Google Colab) - TensorFlow version > 2.0 # Introduction In this project, we use satellite imagery from 23 countries in Africa combined with survey data about wealth and poverty to train machine learning models to predict whether a given area ("village") in Africa suffers from extreme poverty. We use convolutional neural networks, transfer learning and ensemble modeling to develop a model that correctly identifies extreme …