As global economic systems develop, it is important that no group of people get left behind. The United Nations has listed the reduction of poverty among it’s sustainability and development goals, and international/national resource allocation programs exist that target low resource areas. However, it can be challenging to asses which areas are most in need of aid. To understand the resource distribution, some countries perform rountine demographic and health surveys. However, other countries are unwilling or are unable to collect such information. In this paper, the authors developed a machine learning model to fill in the gaps of demographic and health surveys to predict poverty.
Methodologies Used
At its’s core, this model uses a Convolutional Neural Network (CNN). CNN’s are popular in computer vision tasks because of their ability to pick out image features like horizontal lines, circles, diagonal lines etc. For this task the authors use convolutional filters that correspond to urban areas, non-urban areas, water, and roads. In other words these filters trace across an image and are activated anytime they pick up on one of the four features listed previously. However, CNN’s cannot be directly applied to predicting household income because demographic and health surveys group hundreds of communities under a single household income. This means there isn’t enough data for a CNN to learn from. In a previous paper, the authors built a proxy for economic outcomes that relies on night time lights. This proxy predicts night time lights from the daytime imagery. Then, with the newly trained model, they use this proxy to predict either average household income or average household wealth. The CNN is trained on data from Nigeria (2012), Tanzania (2012), Uganda (2011), Malawi (2013).
Key Takeaways
For each of the four countreis, this model modestly explains 37 to 55 percent of variation in the average household consumption, and even better on average household asset weatlh, 55 to 77 percent. Using their unique proxy for economic outcomes, the author’s achieve better varaince explanation than a model that directly uses night time lights as it’s sole predictor of economic outcomes. Additionally, a model trained on country A often performs just as well or sligthly better when tested on country B. This result implies that performing an expensive demographic and health survey on a single country can produce modest predictions that generalize to other countries. Finally, despite the challenges in predicting economic outcomes, such as scarce and noisy data, the performance of the proxy model offered by the authors demonstrates the effectiveness of daytime satellite imagery machine learning.
Citation
Jean, N., Burke, M., Xie, M., Davis, W. M., Lobell, D. B., & Ermon, S. (2016). Combining satellite imagery and machine learning to predict poverty. Science, 353(6301), 790-794. https://www.science.org/doi/epdf/10.1126/science.aaf7894