Noise, or unwanted sound, is ubiquitous in urban environments. I introduce a new approach to measuring noise at scale using seismometers and compile the largest publicly available database of ambient noise. Using a research design leveraging idiosyncratic variation in electric passenger rail noise exposure, I estimate the impact of noise pollution on infant health. In utero noise pollution exposure harms health at birth. A 2 decibel increase in average noise levels during pregnancy—a small but perceptible increase—lowers an overall index measure of infant health by 4 percent of a standard deviation, equivalent to one-third of the Black-White gap in the index. This effect is driven by nighttime noise, suggesting disruptions to maternal sleep as a main mechanism. Overnight rail services, which account for only about 5 percent of overall ridership, generate an average externality of $18 per trip. In per passenger-mile terms, overnight rail noise externalities are comparable to rush hour traffic congestion externalities from private vehicles. Using seismic data and machine learning, I produce a novel map of noise for the contiguous United States and use this map to assess noise exposure and costs nationally. Eighty percent of urban residents are exposed to potentially harmful levels of nighttime noise. I estimate that the annual cost of noise pollution due to harms to health at birth is $9.8 billion. Urban, Black, and Hispanic Americans disproportionately bear these costs.

Projecting the effects of proposed policy reforms is challenging because no outcome data exist for regulations that governments have not yet implemented. We propose an ex ante deep learning framework that can project effects of proposed reforms by mapping outcomes observed under past regulations onto the legal criteria of proposed future policies (i.e., by “relabeling”). We apply this framework to study changes in jurisdiction of the US Clean Water Act (CWA), which regulates many sites used for renewable and fossil energy, transmission lines and pipelines, transportation infrastructure, and other types of land use. We compare our ex ante deep learning projection of jurisdiction under the Supreme Court’s Sackett decision against widely used projections from domain experts. Ex ante machine learning generates exceptional performance improvements over the leading domain expert model that the US Environmental Protection Agency currently uses, with 65 times more accurate identification of jurisdictional sites. We also develop an ex post deep learning model trained with data after policy implementation. Ex post deep learning performs best.Sackett deregulates one-third of all previously regulated US waters, particularly floodplains and pristine fish habitats, totaling 700,000 deregulated stream miles and 17 million deregulated wetland acres. Deep learning can effectively project consequences of far-reaching regulatory reforms before they are implemented, when projections are both most uncertain and most useful.

Satellite imagery and machine learning (SIML) are increasingly being combined to remotely measure social and environmental outcomes, yet use of this technology has been limited by insufficient understanding of its strengths and weaknesses. Here, we undertake the most extensive effort yet to characterize the potential and limits of using a SIML technology to measure ground conditions. We conduct 115 standardized large-scale experiments using a composite high-resolution optical image of Earth and a generalizable SIML technology to evaluate what can be accurately measured and where this technology struggles. We find that SIML alone predicts roughly half the variation in ground measurements on average, and that variables describing human society (e.g. female literacy, R²=0.55) are generally as easily measured as natural variables (e.g. bird diversity, R²=0.55). Patterns of performance across measured variable type, space, income and population density indicate that SIML can likely support many new applications and decision-making use cases, although within quantifiable limits.

Developing world cities are projected to grow dramatically throughout the 21st century. Understanding the causes and consequences of 20th century urban growth can improve 21st century urban policy, but requires detailed data on historical urbanization. To address this need, we develop a machine learning method to extract building footprints from historical imagery, producing high-resolution building maps. We produce 581 building maps for African cities between 1943 and 1990. These data measure long-run urban development at an unprecedented spatial granularity and scale. Importantly, our data allow us to study not only large, well-known cities, but also slums, hinterlands, and newly urbanizing areas, which may not be represented in other data sources.