Large-scale, high-resolution mapping of vulnerable communities through satellite imagery and population censuses using deep learning

Description of the service

One of the challenges in the fight against poverty is the accurate location and assessment of the expansion of vulnerable communities. Characterization of vulnerability is traditionally achieved through nationwide census exercises, an elaborate process that requires field visits by trained personnel. Unfortunately, most national-level censuses are sporadic, making it difficult to monitor the short-term impact of policies to reduce poverty. This app defines vulnerability following UN-Habitat criteria. We evaluated different CNN machine learning architectures and established a mapping between satellite images and survey data. With the information of 2,178,508 residential blocks registered in the 2010 Mexican census and Landsat-7 multispectral imagery, we explored multiple deep learning architectures. The best performance is obtained with EfficientNet-B3 achieving an area under the ROC and Precision-Recall curves of 0.9421 and 0.9457, respectively. Our work allows the use of publicly available information, in the form of census data and satellite imagery, together with standard CNN architectures. Our purpose is to employ this tool to characterize vulnerability across the country at the residential block level.

The project was developed jointly by IPN and INEGI. The result can be reviewed at https://afigueroa.users.earthengine.app/view/vulnerable-settlements. The code is publicly available at https://git.inegi.org.mx/laboratorio-de-ciencia-de-datos/vulnerability. The article describing our methodology is published at https://www.mdpi.com/2072-4292/13/18/3603