Estimation of socioeconomic indicators through satellite imagery -
Analysis of urban areas overlapping
Abstract
The NEXUS area covers approximately 30% of the Brazilian territory. In
order to assist preservation and sustainable development policies in
that region, this study proposes to replicate the work done by Yeh et al
in Africa , in which a convolutional neural network estimates indicators
through satellite images, each covering a region of approximately 45
km². This work compares the size and distribution of Brazil’s census
tracts with those in Africa to define if the scale of images can be
maintained and to define the clusters that will be used. To avoid
biasing the model, special care must be taken in selecting clusters,
such as keeping a balance between urban and rural sectors and, most
importantly, making sure that there is little to no overlap of clusters.
To do so, two approaches were proposed. The first one samples tracts in
each municipality as centroids for clusters, the second merges
neighboring urban tracts into a single group and fits clusters to these
groups.