IN010-08
Evaluation of deep-learning methods to understand the prediction of socio-economic indicators from remote sensing imagery
Tuesday, 8 December 2020: 19:28
Virtual
Jeaneth Machicao1, Robin Jarry2, Danton Ferreira Vellenich3, Jean Pierre Ometto4, Katia Maria P. M. B. Ferraz1, Solange Santos5, Nadya Deps6, Miguel Penteado7, Shelley Stall8, Alison Specht9, Laurence Mabile10, Marc Chaumont11,12, Pedro Luiz Pizzigatti Corrêa13, Romain David14 and Jeaneth Machicao, (1)University of São Paulo, São Paulo, Brazil, (2)Foundation for Research on Biodiversity, Paris, France, (3)University of São Paulo, Department of Computer Engineering and Digital Systems, São Paulo, Brazil, (4)INPE National Institute for Space Research (INPE), Earth System Science Center, Sao Jose dos Campos, Brazil, (5)Scientific Electronic Library Online, São Paulo, Brazil, (6)IBGE Brazilian Geography and Statistics Institute, Library, Rio De Janeiro, Brazil, (7)IBGE Brazilian Geography and Statistics Institute, Supervisão de Informática, Rio De Janeiro, Brazil, (8)American Geophysical Union, Data Leadership, Washington, DC, United States, (9)University of Queensland, Brisbane, Australia, (10)University Paul Sabatier Toulouse III, Toulouse Cedex 09, France, (11)Univ. Nîmes, LIRMM, Montpellier, France, (12)CNRS, Paris, France, (13)EPUSP Polytechnic School of the University of Sao Paulo, Department of Computer Engineering and Digital Systems, Sao Paulo, Brazil, (14)ERINHA, Europa Commission, Paris, France
Abstract:
Mapping poverty distribution in developing countries is one of the main challenges to measure inequalities and monitor policy impact. Currently, in Brazil, socioeconomic indicators (SEc) are collected from census surveys conducted by the government institutions every 10 years. Approaches utilizing deep learning and remote sensing have demonstrated their value as inexpensive alternative options for determining change in socioeconomic indicators. In this ongoing work, an international consortium of six countries, known as the PARSEC, funded by the Belmont Forum, aims to conduct a deep learning approach to predict SEc for selected areas in Brazil using satellite imagery.
In this paper we shall share the approach and methods we have used to date. Reviewing four case studies in this domain (Xi et al., 2016; Jean et al., 2016; Suel et al., 2019; Ayush et al., 2020) we found a common three-stage learning methodology: (1) establish a “preliminary task” (pre-trained model) consisting to train a CNN (convolutional neural network) using a large dataset of images, aiming to intensively learn the relationship between the input and their images annotations (intermediate outputs); (2) extract a feature vector from the CNN output which will be used as the transferability learning, so that each input image would correspond to a feature vector and could be annotated with a SEc; and (3) use a simpler regression model to predict poverty measures from the corresponding CNN feature vector output.
From these four sources, we observed that: (i) the input imagery can be gathered from different sources; (ii) the successful prediction will rely on the ability to find the best preliminary task, which can be fine-tuned and also combine strategies; (iii) the intermediate outputs can transfer learning from object segmentation classification, landscape area annotations, etc.; and (iv) the annotation outputs (SEc) can be collected from census data with indicators such as human development index, consumption expenditure, asset health.
By identifying inexpensive methods of predicting change in socioeconomic predictions, we hope to provide decision makers with more current and relevant information that results in effective policies with improved outcomes for citizens of Brazil and other countries.
