Conference Agenda

Globally, a large fraction of solid waste is disposed in open dumps, generating a range of harmful externalities. Research on this topic is limited by a lack of data, however. We address this gap by developing methods to construct a globally representative time- series of open dumps, starting with a small, unrepresentative sample of crowdsourced observations. We use these observations to train a potentially biased machine learning model that can predict dumps using satellite imagery. We then employ active learning techniques to select and verify an approximately optimal representative subset of verified points to debias our machine learning predictions. This method gives unbiased estimates of dump prevalence and results in significant efficiency gains relative to standard methods, even in the presence of model miscalibration. We use our data to quantify how international trade in waste generates open dumps. Our results show a dramatic increase in open-air landfills globally after China banned imports of plastic waste in 2018. The increase is concentrated in low-income countries that saw increased imports, suggesting that trade plays a role in overwhelming local waste management systems and the subsequent leakage of waste into the environment.