Using Street View Images and Machine Learning to Measure Household Consumption and Inequality in Developing Countries
- 주제어 (키워드) Poverty , Measurement , Convolutional Neural Networks (CNNs) , Inequality , Wealth-Based Approach , Google Street View Images
- 발행기관 서강대학교 일반대학원
- 지도교수 양현주
- 발행년도 2024
- 학위수여년월 2024. 8
- 학위명 석사
- 학과 및 전공 일반대학원 경제학과
- 실제 URI http://www.dcollection.net/handler/sogang/000000079187
- UCI I804:11029-000000079187
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권 보호를 받습니다.
초록 (요약문)
The scarcity of detailed economic status data poses significant challenges in analyzing the impact of exogenous economic shocks, particularly in developing countries. This paper, adopting a wealth-based approach, highlights the potential of image data in reliably estimating household consumption expenditure. Utilizing photographic data that capture wealth indicators from households across 66 countries, the study demonstrates that a convolutional neural network (CNN) model can explain up to 77% of the variation in household-level economic outcomes. The model’s validity is further substantiated by the high correlation between the predicted economic status of neighborhoods and the actual status, using the best-performing model with South African educational data and Google Street View (GSV) images. To explore the usefulness of the introduced granular measurements, two applications are showcased. First, a substantial negative impact of Hurricane Michael on the predicted consumption is presented by comparing estimates of pre- and post-disaster GSV images. Second, I calculate a sub-national consumption inequality measure in India, featuring a significant variation even in the capital city, Delhi. These results underscore the potential of image-based data as a powerful tool for measuring reliable and detailed economic data for targeted poverty alleviation and policy evaluation.
more목차
1 Introduction 2
2 Data 7
2.1 Image data from Dollar Street 7
2.2 School Masterlist Data (2021) 9
2.3 Google Street View images 10
2.4 Administrative boundary shapefiles of South Africa and India 11
3 Prediction Methods 12
3.1 Web scraping and image downloading 12
3.2 Pre-processing 14
3.3 Model training 14
3.4 Out-of-sample validation and application 15
4 Results 16
4.1 Indoor group 16
4.2 Outdoor group 18
4.3 Falsification test 20
5 Out-of-Sample Validation and Application 22
5.1 Out-of-sample validation: Quintile prediction in South Africa 22
5.2 Application 1: Assessing consumption level changes post-natural disaster 24
5.3 Application 2: Constructing a sub-national inequality measure in India 26
6 Conclusions 27
References 30
