Poverty Mapping in Developing Countries using High Resolution Satellite Imagery, OpenStreetMap and Household Survey Data
Abstract
For efficient targeting of policies, poverty maps should be made available at the finest administrative unit of planning i.e., at a local scale. Currently, the most reliable way to estimate poverty is through household surveys, but they are very costly and conducted every 10-15 years. This is creating a data gap, thereby delaying the actions to be taken for poorer economies within years, making the alleviation of poverty challenging. There is still no accurate, inexpensive, and scalable approach to map poverty at a local scale.
Our main research goal is to develop low scale poverty maps at the building level by combining high resolution satellite image, openstreet map and household survey data. Although a lot of research has already been done on estimating poor areas, estimation at a building level is much more detailed than them, which is made possible by integrating household survey data with machine learning. We use two kinds of deep learning models: first one is a segmentation model to extract building footprints from satellite image using an open street map as training data and the second one is a classification model to estimate the economic levels of buildings. Here, we use the household income to define economic levels and set three economic classes: poor, middle, and rich. Then, we combine the extracted building footprints with the household survey data to make training data for the classification model. At the time of training, we used the building height, area of the building and elevation data to improve the results of the model. Firstly, we develop the two proposed models with the data of Bago, Myanmar and later calibrate them with the data of Managua, Nicaragua. Our results demonstrate that we can estimate large-scale economic levels of buildings easily. Policy makers can use our results, particularly in developing nations and identify where exactly the poor people are residing and effectively design the actions to alleviate poverty. Furthermore, these results can be used in various fields such as disaster prevention, land use plan or urban planning as these fields are closely related to poverty issues. Therefore, our method will contribute to the issue of poverty reduction in many aspects.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMIN007..12C
- Keywords:
-
- 1912 Data management;
- preservation;
- rescue;
- INFORMATICS;
- 1916 Data and information discovery;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS;
- 1960 Portals and user interfaces;
- INFORMATICS