Authors:

Shivi P

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

The use of satellite imagery and machine learning for poverty mapping is a cost efficient and scalable approach to bridging the data gap in developing nations. While nightlights are widely recognized as a proxy for economic activity and poverty, they also pose significant challenges. In 2016, a novel transfer learning approach addressed some of these issues by combining daytime and nighttime imagery. Further developments such as the random forest approach have also progressed. However, an aspect of the field that lacks adequate studies is the temporal robustness of the machine learning approach. This study aims to investigate the temporal robustness of the two prevailing machine learning models for poverty mapping - transfer learning with convolutional neural networks and random forest - for five countries in Sub-Saharan Africa (Malawi, Nigeria, Uganda, Tanzania, and Ethiopia) from 2010 - 2024 with an intent to determine reliability for future predictions. This was achieved by iteratively training both models on satellite data (available every year) and survey data (available every 5 years) from 2010 - 2020, fine tuning the parameters as necessary to estimate asset wealth and consumption expenditure for the five countries. The performance of the fine tuned models were then tested for temporal robustness from 2010 - 2020, and utilized to predict a trendline from 2020 - 2030. The results are expected to provide insight on the degree to which the transfer learning and random forest models are temporally robust. Going forward, this knowledge can support developing nations in designing and tracking better targeted policies. Keywords: Machine Learning; Transfer Learning; Random Forest; Poverty Mapping; Temporal Robustness

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Shivi P