FlowAlpha/csec491-kenya-unet

Senior Project # Senior Project for Yale Computer Science and Economics Bachelor of Science *CSEC 491, a one-term independent-project course that explicitly combines both techniques and subject matter from computer science and economics.* ## Satellite Remote Sensing and U-Net for Predicting Population Density in Kenya One of the difficult roadblocks in empirical economics is gaining sufficient and relevant data. Natural limitations lie in methods and opportunity of collection and measurement. Data that researchers want is often expensive, time-consuming, elaborate, or often outright unfeasible to acquire. Data availability is extremely valuable for estimating clausal inference related to some of the biggest, most complex to some of the simplest, most approachable questions in economics research. Availability of satellite imagery and robust machine learning techniques and technologies have allowed powerful tools into the hands of many researchers. Increasing in the last decade, researchers have expanded testing the use of remote sensing via using these tools to try to produce survey-like data. This has been especially useful in environmental impact research, but recent projects have begun to increasingly delve into general income and population estimation. Examples of this include a paper by Khachiyan, estimating income and population data in the US and the working paper by Huang et al. doing similar to estimate anti-poverty programs in the US. In this project, I do some testing of my own. I use a streamlined Python notebook workflow, utilizing Google Earth Engine API, Tensorflow, and Google Cloud to make predictions directly from separate image collection assets, targeting the response of a given population density raster with data from the 2009 Kenyan census. I use a basic U-net architecture to regress on this raster and different remote sensing bands to run predictions on a set region of Kenya. I show that even with some slight tweaks, great improvement can be made to reduce lo …