EO4SDGs: Spatiotemporal poverty mapping using earth observation data and deep learning in Africa

点击翻译按钮获取中文摘要

This project will build spatiotemporal maps of poverty on a sub-national scale to support the implementation of the SDGs and enable evidence-based decision-making by combining EO data with local fine-resolution assessments. This task will involve understanding associations between EO metrics and socioeconomic conditions as well as the relationships between poverty and geospatial proxies in different countries, counties, and wards. Moreover, the high temporal resolution of the EO data will be used to track changes in SDGs metrics and identify spatial locations with unusual changes in patterns in the signal so that new surveys targeting those regions can be commissioned.Deep learning techniques, such as Convolutional Neural Networks (CNNs), are increasingly used for predictive analytics with remote sensing images and tasks such as ground object detection, population, land mapping, etc. This project will investigate deep learning techniques to fill spatial gaps in earth observation-based (EO) products. However, a drawback of using solely deep learning models to derive data-driven policy and geographic targeting across time and space is their lack of interpretability. Indeed, these models are well known to be black boxes, making the results not easily explained, justified or intuitive, therefore reducing their practicability for policy-making purposes. Statistical models, on the other hand, are designed such that the parameters reflect the relationship between different features of the data and therefore are interpretable and transferable. Although this level of interpretability is not possible in a black-box deep learning model, they are remarkably accurate for prediction purposes. To address this dichotomy, this project will develop a novel workflow that accurately reproduces SGD indicators while retaining the interpretability of statistical models. Previous studies established relationships between household poverty from household survey data and geospatial data for the surrounding area, but household data is available only partially for a specific ward. The assumption of homogeneity between wards is not valid in general, making the transferability an issue for wards with large variations in socioecological systems. Geostatistical models based on Gaussian processes will be investigated to address the problem of transferability and ultimately predict poverty even at locations where no data is available by borrowing information from neighboring regions. The approach will incorporate multiple EO satellite data and local fine-resolution assessments via spatiotemporal modeling. It is likely that the study will have a focus in East Africa.