Machine Learning in Poverty Measurement

May 11, 2021

According to The Sustainable Development Goals (SDG) from the United Nations, the first goal is to "end poverty in all its forms everywhere". However, a common method to measure poverty is census data or large sample research, which collects data from a large sample size. The cost for conducting these researches is even higher in low-income areas due to the scarce infrastructure (Blumenstock, 2016; Jean et al., 2016; Perez et al., 2019, McBride&Nichols, 2015).

As technology develops, scholars and researchers have begun to apply new techniques and massive machine-generated data sources to measure poverty. In this blog, I discuss three general trends in machine learning about poverty measurement and some concerns in the current application.  


  1. From single modal to multimodal representation

New techniques generate new feature variables from multiple sources that allow researchers to measure poverty beyond survey and census data. Nowadays, several alternative information sources in different modals are applied for prediction (Pokhriyal & Jacques, 2017). Satellite imagery has become a common method as well as mobile phone data for measuring poverty rates (Smith-Clarke et al., 2014, Khan&Blumenstock, 2019; Aiken et al., 2020). Besides these two sources, several other sources may also reveal features that could efficiently predict poverty. For example, Google Street View (e.g. Car brand, housing situation), OpenStreetMap (e.g. distance to the town), and e-commerce data (e.g. shopping and shipping information) were applied in various ways to studying poverty (Tingzon et al., 2019, Gebru et al., 2017; Wijaya et. al., 2020; Watmough et al., 2016). In different circumstances, these sources may or may not be helpful. Overall, it is a trend that researchers started to use several data sources in various formats.

To respond to this trend, fusing these features might be a challenge, especially with high dimension vector features from images produced by satellite data. Usually, researchers may concatenate vectors or get the average of several vectors. However, fusion technology plays a significant role in the Visual Question Answering field to weigh different modals during training (Zhang et al., 2020). This application in poverty measurement is still limited.

  1. From human-oriented to machine-oriented

As the format of data sources changes and multiple data sources become available, the modeling process also changes. For example, it was hard to apply satellite imagery decades ago due to computational power, programming language infrastructure, etc. In 2014, DataKind tried to build a relationship between satellite imagery and poverty measurement but they needed people to hand-code images of different home types based on their roof materials (Abelson et al., 2014). In this manner, these machine learning features are still generated by humans. This issue is further complicated by deep learning techniques such as Convolutional Neural Networks (CNNs) that task a computer with generating features beyond single-pixel analysis and which take into account surrounding contextual information(Blumenstock, 2016; Jean et al., 2016). Joshua Blumenstock’s team at UC Berkeley and Marshall Burke’s team at Stanford University are two famous pioneers in this research.

Semi-supervised methods are useful in developing countries that have fewer existing labels readily available (e.g. poverty rate). This method makes the small proportion of labeled images possible to result in a similar accuracy as a larger data set, which could improve generalization in smaller datasets (Perez et al, 2019). When seeing the trend of machine-oriented methods in poverty measurement, we still need to reconsider the definition of poverty itself. Usually, in prediction analysis, the dependent variables in these types of research are based on the Multidimensional Poverty Index (MPI). However, the understanding or the measurement of poverty from economists, social scientists, etc are always developing (Alkire&Foster, 2011; Alkire&Santos, 2014; Alkire et al., 2017; Bourguignon&Chakravarty, 2019). This creates a space for social science theory and computation to learn from each other to prevent bias from dependent variables. 

  1. From high cost to low cost

One of the incentives that researchers consider when applying alternative methods to measure poverty rather than direct surveys is low cost. Here, cost refers to the labor and price to acquire data. Early research applied high-resolution satellite imagery, which is not very easy to access and may not be free in some areas some time series. Thus, researchers may search other open data but with lower resolution satellite imagery like Landsat 7 for the measurement but achieve similar accuracy (Perez et al, 2017; Hersh et al., 2020). Ayush et al.(2020) combined high and low-resolution satellite imagery by Deep Reinforcement Learning, which asks the machine to decide whether it is necessary to acquire high-resolution imagery in that area. Moreover, as more convenient open source programming packages and pre-trained models are built, like VGG19, ResNet50, EfficientNetB series, the cost to learn features from images will also decrease. 

Thus, in both data sources and analysis, the cost will hopefully continue to decrease However, saving the cost in training does not mean that we save time in understanding the data as well as the theory and application behind those pre-trained packages. We still need to check the quality of the dataset and understand the data collection mechanism. For example, if daylight satellite imagery does not update frequently, for instance, Google Earth does not have a fixed update frequency in one area, we might need to rethink whether this factor impacts our inference. 

Poverty is very complicated, impacted by multiple factors, and there is no single path into or out of poverty (McKernan&Ratcliffe, 2005; Asongu&Roux, 2019; Njong&Ningaye, 2008). Similarly, Salganik et al.(2020) concluded that The Fragile Families Challenge, (160 teams),  a mass collaboration to yield insights for disadvantaged children, could not achieve desirable results in life trajectories prediction like lay off. Thus, we should always learn how to prevent simplifying the relationship between poverty and various factors and also avoid new techniques stressing the existing injustice. Furthermore, subject matter expertise is an important component in data science. Thus, if possible, social welfare scholars should be considered in data science for social good and promoting social justice, especially when we beyond measurement and move to the intervention stage. 

Acknowledgments: I am thankful to my advisors Rediet Abebe, Ricardo Sandoval, Julian Chow, and Evan Muzzall, who provided constructive comments and suggestions.


Alkire, S., & Foster, J. (2011). Counting and multidimensional poverty measurement. Journal of public economics, 95(7-8), 476-487.

Alkire, S., & Santos, M. E. (2014). Measuring Acute Poverty in the Developing World: Robustness and Scope of the Multidimensional Poverty Index. World Development, 59, 251–274.

Alkire, S., Apablaza, M., Chakravarty, S., & Yalonetzky, G. (2017). Measuring chronic multidimensional poverty. Journal of Policy Modeling, 39(6), 983–1006.

Abelson, B., Varshney, K. R., & Sun, J. (2014). Targeting direct cash transfers to the extremely poor. Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1563–1572.

Aiken, E. L., Bedoya, G., Coville, A., & Blumenstock, J. E. (2020). Targeting Development Aid with Machine Learning and Mobile Phone Data: Evidence from an Anti-Poverty Intervention in Afghanistan. Proceedings of the 3rd ACM SIGCAS Conference on Computing and Sustainable Societies, 310–311.

Asongu, S. A., & Le Roux, S. (2019). Understanding sub-Saharan Africa’s extreme poverty tragedy. International Journal of Public Administration, 42(6), 457-467.

Ayush, K., Uzkent, B., Tanmay, K., Burke, M., Lobell, D., & Ermon, S. (2020). Efficient poverty mapping using deep reinforcement learning. arXiv preprint arXiv:2006.04224.

Blumenstock, J. E. (2016). Fighting poverty with data. Science, 353(6301), 753–754.

Bourguignon, F., & Chakravarty, S. R. (2019). The measurement of multidimensional poverty. In Poverty, social exclusion and stochastic dominance (pp. 83-107). Springer, Singapore.

Gebru, T., Krause, J., Wang, Y., Chen, D., Deng, J., Aiden, E. L., & Fei-Fei, L. (2017). Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States. Proceedings of the National Academy of Sciences, 114(50), 13108-13113.

Hersh, J., Engstrom, R., & Mann, M. (2020). Open data for algorithms: Mapping poverty in Belize using open satellite derived features and machine learning. Information Technology for Development, 1–30.

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.

Khan, M. R., & Blumenstock, J. E. (2019, July). Multi-gcn: Graph convolutional networks for multi-view networks, with applications to global poverty. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 33, No. 01, pp. 606-613).

McBride, L., & Nichols, A. (2015). Improved poverty targeting through machine learning: An application to the USAID Poverty Assessment Tools. Unpublished manuscript. Available at: http://www. econthatmatters. com/wp-content/uploads/2015/01/improvedtargeting_21jan2015. pdf.

McKernan, S. M., & Ratcliffe, C. (2005). Events that trigger poverty entries and exits. Social Science Quarterly, 86, 1146-1169.

Njong, A. M., & Ningaye, P. (2008). Characterizing weights in the measurement of multidimensional poverty: An application of data-driven approaches to Cameroonian data.

Smith-Clarke, C., Mashadi, A., & Capra, L. (2014). Poverty on the Cheap: Estimating Poverty Maps Using Aggregated Mobile Communication Networks. 10.

Perez, A., Yeh, C., Azzari, G., Burke, M., Lobell, D., & Ermon, S. (2017). Poverty Prediction with Public Landsat 7 Satellite Imagery and Machine Learning. ArXiv:1711.03654 [Cs, Stat].

Perez, A., Ganguli, S., Ermon, S., Azzari, G., Burke, M., & Lobell, D. (2019). Semi-Supervised Multitask Learning on Multispectral Satellite Images Using Wasserstein Generative Adversarial Networks (GANs) for Predicting Poverty. ArXiv:1902.11110 [Cs].

Pokhriyal, N., & Jacques, D. C. (2017). Combining disparate data sources for improved poverty prediction and mapping. Proceedings of the National Academy of Sciences, 114(46), E9783–E9792.

Salganik, M. J., Lundberg, I., Kindel, A. T., Ahearn, C. E., Al-Ghoneim, K., Almaatouq, A., ... & McLanahan, S. (2020). Measuring the predictability of life outcomes with a scientific mass collaboration. Proceedings of the National Academy of Sciences, 117(15), 8398-8403.

Tingzon, I., Orden, A., Go, K. T., Sy, S., Sekara, V., Weber, I., Fatehkia, M., García-Herranz, M., & Kim, D. (2019). MAPPING POVERTY IN THE PHILIPPINES USING MACHINE LEARNING, SATELLITE IMAGERY, AND CROWD-SOURCED GEOSPATIAL INFORMATION. ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLII-4/W19, 425–431.

Watmough, G. R., Atkinson, P. M., Saikia, A., & Hutton, C. W. (2016). Understanding the evidence base for poverty–environment relationships using remotely sensed satellite data: an example from Assam, India. World Development, 78, 188-203.

Wijaya, D. R., Paramita, N. L. P. S. P., Uluwiyah, A., Rheza, M., Zahara, A., & Puspita, D. R. (2020). Estimating city-level poverty rate based on e-commerce data with machine learning. Electronic Commerce Research.

Zhang, C., Yang, Z., He, X., & Deng, L. (2020). Multimodal Intelligence: Representation Learning, Information Fusion, and Applications. IEEE Journal of Selected Topics in Signal Processing, 14(3), 478–493.