There are three methods to validate the maps of land and nature: field surveys, high-resolution satellite imagery, and machine learning. Two of these methods have been around since the 1950s: remote sensing scientists and cartographers calibrate and verify their maps with field or ground measurements and high-resolution satellite imagery. The third method, centered on machine learning techniques such as weak supervision, refashions mapping as a problem of overreliance on ground or field data, one that was distinct from, if not antithetical to, collecting in-situ observational data for prediction. In this way, potentially low-quality input data are said to create high-quality predicted outputs. Yet for over half a century, the concept of “rubbish in, rubbish out” was shorthand for talking about the politics of biased data collection and its impact on both accuracy and equity. This talk explores the shifting figure of ground truth in map-making, detailing how changing methods of data classification, measurement, and prediction speak to epistemic tensions around how and who can produce knowledge under conditions of uncertainty.
