Satellite data can help policy-makers quickly identify areas of the world in need of aid and development, but research shows it can also contain bias against marginalised groups, potentially compromising policy goals.
Machine-learning systems that scan satellite images for indicators of poverty or disaster damage are becoming a popular tool for assessing humanitarian and development needs. But Lukas Kondmann and Xiao Xiang Zhu at the German Aerospace Center in Cologne say little attention is being paid to potential biases built into this data.
They to predict poverty levels and rates of electricity in Indian villages using satellite images of night-time lights, which are a common measure of development. It made consistent errors for villages with large populations of scheduled castes and scheduled tribes 鈥� two officially designated groups in India with a long history of experiencing discrimination. The results were presented on 15 August at a virtual conference run by the .
Advertisement
These findings suggest that features of the data reflect historical biases, says Kondmann, who led the research, which could lead to discrimination if used to guide policy. 鈥淲e鈥檙e capturing the patterns we see on the ground, and if these patterns have some sort of bias built into them, the satellite is just going to capture and replicate it,鈥� he says.
During training, the model was fed socioeconomic data, satellite images and coordinates from 386,000 Indian villages. It then used satellite data and coordinates to make predictions about unseen villages. In those with a larger than average number of scheduled tribes, poverty levels were overestimated and levels of electricity were underestimated by about 1 per cent on average. For scheduled castes in those areas, predictions of poverty and electricity levels were out by 0.3 per cent in the opposite direction.
Reasons for the bias are hard to assess, say Kondmann, but it could be due to unfair allocation of public investment or close proximity of wealthy and poor neighborhoods.
鈥淭his sort of analysis is really important,鈥� says at Stanford University in California, who studies poverty mapping. The measured bias is small, however, so he would like to see the experiment repeated with improved models.
These humanitarian approaches tend to be used where data is scarce and decisions are guided by guesswork and politics, says Burke, so it is important to consider whether they are more or less biased than the alternatives.
听
