How innovation in machine learning can overcome limited data
Poverty in Ghana ↓ conflict in Nigeria ↓
Improving the sophistication of poverty estimates
All data sources have limitations. When it comes to measuring poverty, surveys are expensive, limited to small sample sizes, and available only for certain geographic points in certain years. Proxies for poverty such as nighttime lights are imperfect and often struggle to accurately measure populations most in need. AidData is expanding upon methods pioneered by Stanford researchers to overcome these challenges and produce next-gen measures of poverty by leveraging machine learning.
To better estimate the economic impact of improved market access from an MCC program in Ghana, AidData utilized DHS surveys and Landsat 7 satellite data along with convolutional neural networks to provides more accurate estimates of changes in average household wealth. The resulting estimates were used to compare changes in poverty over time between regions around road improvement projects and control locations that were not near improved roads.
Download the data (beta)
Ghana Predicted Surface
Providing a low-cost solution to predict the likelihood of deadly conflict
Uncertainty limits the humanitarian response in areas with known conflicts. AidData is working to improve how existing data can be used to train convolutional neural networks that make more accurate predictions of the likelihood of a death due to conflict at a given location. To make these estimates, the algorithm analyzes landscape features detected from moderate-resolution satellite imagery the previous year. Using Nigeria as a case study, this algorithm achieves approximately 80% classification accuracy when predicting whether a location with a known conflict in 2015, 2017, 2019 was fatal based on previous year's satellite imagery, at either yearly or six month intervals.
Read the research
A convolutional neural network approach to predict non‐permissive environments from moderate‐resolution imagery