UC Berkeley

Over 700 Earth Observation satellites orbit our planet, capturing terabytes of imagery that together form a continuous record of our planet's surface and atmosphere. When combined with machine learning, these data enable quantification of human impacts on the environment: from rising greenhouse gas levels to deforestation. Yet before we can answer questions like ``How much carbon dioxide (CO2) can we emit and still limit warming to 1.5°C?'', we need to understand what these tools can and cannot reliably measure. This dissertation systematically explores those limits and demonstrates how satellite data and aerial photography can advance our understanding of the carbon budget. Chapter 1 establishes the methodological foundation. Using an image-embedding pipeline built on random convolutional features---a flexible, low cost approach for large scale prediction---we evaluate how well satellite imagery and machine learning can recover a wide range of real world outcomes. Some, like tree cover, are captured reliably; others, like child mortality rates, are far more difficult to capture, particularly as distance from ground measurements increases. These findings shape how data splits and model evaluations are structured throughout this dissertation. Chapter 2 turns to the past. Historical land use changes are a major source of uncertainty for the historical record of emissions. By applying deep learning to nearly one million archival aerial photographs, I reconstruct historical land cover maps in Sub-Saharan Africa. I use the Gambia as a case study to demonstrate the high fidelity of our prediction model at scale. This opens the door to answering questions about the impact of the introduction of cash crops, the effects of independence after colonization, and many other research directions. These outputs fill an information gap where detailed land cover maps did not previously exist, and can later be used for carbon budget estimation. Chapter 3 addresses the present. With over 35,000 energy-generating facilities worldwide and energy-related CO2 emissions reaching 36.8 gigatons in 2022, tracking industrial emissions consistently and transparently is an urgent challenge. I compare empirical and deep learning approaches for estimating power plant emissions using remote sensing, showing that each method has distinct strengths and that combining them improves emission estimates regardless of whether the methods have been trained on those power plants. Together, this work expands our understanding of land cover change in Africa, quantifies the limits of satellite based emissions estimation, and contributes tools for more accurate, scalable monitoring of greenhouse gas emissions--- bringing us closer to the transparency needed to meet global climate targets.