UC Berkeley

How does human manipulation of the quantity, directionality and spectral distribution of sunlight affect global agricultural productivity? And how might we build a global observation system to provide measurements to answer this and other key questions in environmental science, economics and policy? This thesis quantifies the impact that atmospheric scattering from volcanic sulfate aerosols and clouds has on global crop yields. In turn, this work informs how anthropogenic influences on the global optical environment -- from geoengineering to air pollution to climate change -- impact global food production and food security. This thesis also develops a system that leverages satellite imagery and machine learning to measure many social and environmental variables with high skill, low cost and no alteration of method. The hope is that the generalizablity, low cost, and simplicity of this system will democratize remote sensing, and accelerate the pace of research into our Earth's socio-environmental systems. Broadly, I hope this thesis contributes to environmental policy and improves the wellbeing of life on Earth. Chapter 1 provides the broad-scale motivation for my work.

Chapter 2 studies the agricultural impacts of Solar radiation management (SRM), which is increasingly considered as an option for managing global temperatures. Yet the economic impacts of ameliorating climatic changes by scattering sunlight back to space remain largely unknown. Though SRM may increase crop yields by reducing heat stress, its impacts from concomitant changes in available sunlight have never been empirically estimated. Here we use the volcanic eruptions that inspired modern SRM proposals as natural experiments to provide the first estimates of how the stratospheric sulfate aerosols (SS) created by the eruptions of El Chichon and Mt. Pinatubo altered the quantity and quality of global sunlight, how those changes in sunlight impacted global crop yields, and the total effect that SS may have on yields in an SRM scenario when the climatic and sunlight effects are jointly considered. We find that the sunlight-mediated impact of SS on yields is negative for both C4 (maize) and C3 (soy, rice, wheat) crops. Applying our yield model to a geoengineering scenario using SS-based SRM from 2050-2069, we find that SRM damages due to scattering sunlight are roughly equal in magnitude to SRM benefits from cooling. This suggests that SRM -- if deployed using SS similar to those emitted by the volcanic eruptions it seeks to mimic -- would attenuate little of the global agricultural damages from climate change on net. Our approach could be extended to study SRM impacts on other global systems, such as human health or ecosystem function.

Chapter 3 explores how anthropogenic emissions of air pollutants and greenhouse gases alter the amount, distribution and properties of cloud cover and, in turn, agricultural productivity. Changing cloudiness may impact crop productivity by altering temperature, precipitation and sunlight. While the impacts of temperature and precipitation on crop productivity are relatively well understood, the impacts of cloud optical scattering have never been empirically estimated and remain poorly constrained because of the potentially offsetting effects of changes in total and scattered sunlight. Here, I leverage remotely-sensed cloud observations and subnational crop yield data to provide the first empirical estimates of the sunlight-mediated efffect of cloud optical scattering on maize and soy yields in the United States, Europe, Brazil, and China. I find a consistent concave response of yields to cloud optical thickness across crops and regions. Changing ten days in the growing season from clear to the optimal cloud thickness increases maize and soy yields by 4.0% and 4.4%, respectively; further increasing cloud thickness to the 95th growing season percentile decreases maize and soy yields by 3.4% and 3.5%. Mechanistically, I find that the concavity in the cloud response is driven by concavity in the response to total sunlight as well as -- in some regions -- benefits from increased diffuse light. Applying these empirical estimates to earth system model simulations, I find that changes in sunlight, due to anthropogenic air pollution-induced changes in clouds, are suppressing maize and soy yields by as much as 5% in heavily polluted areas of India and China by increasing the frequency of days with extremely high cloud optical depths. This costs Chinese maize farmers roughly US$1 billion a year. Changes in sunlight due to changes in clouds from a quadrupling of CO2 relative to pre-industrial tend to decrease global maize yields and redistribute soy yields. The methodology developed in this chapter could be extended study the impact of changes to the global optical environment on other global-scale economic outcomes.

Chapter 4 develops a system combining satellite imagery with machine learning (SIML) to observe many variables simultaneously. Current case-by-case solutions require custom systems, extensive expert knowledge, access to imagery, and major computational resources in order to estimate a single variable (a task) using regional or global imagery. Here, we develop a general solution to constructing global observations via SIML, where a single method for transforming satellite imagery is sufficiently descriptive that it should be able to predict nearly any ground-level variables that are recoverable through inspection of a satellite image, including previously unstudied tasks. Our approach is task-independent, allowing centralized computation of features to be executed only once ever per image, then distributed and applied to potentially unlimited future tasks by users who require neither domain expertise nor access to underlying imagery. We demonstrate this generalizability across tasks by constructing high resolution (~1kmx1km) estimates for forest cover, population density, elevation, nighttime lights, household income, total road length, and housing prices across the entire US using exclusively daytime images that are processed only once and in advance. Our system outperforms spatial extrapolation of ground-truth data, especially over large distances, and matches or exceeds performance of a state-of-the-art deep convolutional neural network that is much more costly to implement. Our approach requires only that users download a tabular data set, merge it to geolocated labels, and implement a single regression on a personal computer. We demonstrate that our design scales globally with no alterations and naturally achieves super-resolution, where estimates are more spatially granular than the original labels used for training. Generalization enables democratization of SIML, potentially increasing the pace of planet-scale observation and research, accelerating our understanding of global processes and enabling progress towards tackling planetary challenges.