Welfare estimations from imagery: A test of domain experts’ ability to
rate poverty from visual inspection of satellite imagery
Abstract
Background : Rapidly and yet accurately estimating
welfare levels at different spatial scales is critical to ensuring that
no region is left behind in the quest for poverty reduction and
eradication. Useful as the traditional workhorse of household surveys
are, they are expensive to implement and often have time lags between
them. Recent advances in remote sensing (mainly satellite imagery) and
Artificial Intelligence (in the fields of machine learning, deep
learning, and transfer learning) have led to increased accuracies in
poverty and welfare estimation. These systems are, however, largely
opaque in terms of explaining how these impressive results are achieved.
To achieve explainable AI, domain knowledge of poverty features become
essential and this requires collaboration between humans and machines.
Methods : The present study uses domain experts to
estimate welfare levels and indicators from high-resolution satellite
imagery. We use the wealth quintiles from the 2015 Tanzania DHS dataset
as ground truth data. We analyse the performance of the visual
estimation of relative wealth at the cluster level and compare these
with wealth rankings from the DHS survey of 2015 for that country using
correlations, ordinal regressions and multinomial logistic regressions.
Findings : Of the 608 clusters, 115 (19%)
received the same ratings from human experts and the independent DHS
rankings. For 59% of the clusters, experts’ ratings were slightly lower
(Md = 2.50, n = 358) than DHS rankings (Md = 3.00, n = 135), z = -11.32,
p = <0.001, with a moderate effect size, r = -0.32. On the one
hand, significant positive predictors of wealth are the presence of
modern roofs and wider roads. For instance, the log odds of receiving a
rating in a higher quintile on the wealth rankings is 0.917 points
higher on average for clusters with buildings with slate or tile roofing
compared to those without. On the other hand, significant negative
predictors included poor road coverage, low to medium greenery coverage,
and low to medium building density. Other key predictors from the
multinomial regression model include settlement structure and farm
sizes. Significance : These findings are
significant to the extent that these correlates of wealth and poverty
are visually readable from satellite imagery and can be used to train
machine learning models in poverty predictions. Using these features for
training will contribute to more transparent ML models and,
consequently, explainable AI.