I develop and apply novel datasets and statistical methods to empirically estimate human impacts on climate and, in turn, on our food, economic and health systems. Though my primary training is in environmental economics, I borrow extensively from other disciplines including machine learning, climate science and agronomy.
My research has two primary streams. In the first, I study how agricultural productivity responds to changes in climate focusing on how humans alter the transfer of sunlight through the atmosphere and how this, in turn, impacts crop yields. I am also particularly interested in the role of historic and future water supply in determining global agricultural productivity. In the second, I develop and characterize satellite remote sensing algorithms for use in applied research (see more here). These streams converge as I develop and apply remotely sensed datasets to estimate how global environmental change affects agricultural and other socioeconomic systems.
Sherman*, L., Proctor*, J., Druckenmiller, H., Tapia, H., Hsiang, S. (2023). “Global High-Resolution Estimates of the United Nations Human Development Index Using Satellite Imagery and Machine-learning." NBER Working Paper (here)
The United Nations Human Development Index (HDI) is arguably the most widely used alternative to gross domestic product for measuring national development. This is in large part due to its multidimensional nature, as it incorporates not only income, but also education and health. However, the low country-level resolution of the global HDI data released by the Human Development Report Office of the United Nations Development Programme (N=191 countries) has limited its use at the local level. Recent efforts used labor-intensive survey data to produce HDI estimates for first-level administrative units (e.g., states/provinces). Here, we build on recent advances in machine learning and satellite imagery to develop the first global estimates of HDI for second-level administrative units (e.g., municipalities/counties, N = 61,591) and for a global 0.1 × 0.1 degree grid (N=806,361). To accomplish this we develop and validate a generalizable downscaling technique based on satellite imagery that allows for training and prediction with observations of arbitrary shape and size. This enables us to train a model using provincial administrative data and generate HDI estimates at the municipality and grid levels. Our results indicate that more than half of the global population was previously assigned to the incorrect HDI quintile within each country, due to aggregation bias resulting from lower resolution estimates. We also illustrate how these data can improve decision-making. We make these high resolution HDI estimates publicly available in the hope that they increase understanding of human wellbeing globally and improve the effectiveness of policies supporting sustainable development. We also make available the satellite features and software necessary to increase the spatial resolution of any other global-scale administrative data that is detectable via imagery.
Proctor*, J., Carleton*, T., Sum*, S. (2023). “Parameter Recovery Using Remotely Sensed Variables." NBER Working Paper (here)
Remotely sensed measurements and other machine learning predictions are increasingly used in place of direct observations in empirical analyses. Errors in such measures may bias parameter estimation, but it remains unclear how large such biases are or how to correct for them. We leverage a new benchmark dataset providing co-located ground truth observations and remotely sensed measurements for multiple variables across the contiguous U.S. to show that the common practice of using remotely sensed measurements without correction leads to biased parameter point estimates and standard errors across a diversity of empirical settings. More than three-quarters of the 95% confidence intervals we estimate using remotely sensed measurements do not contain the true coefficient of interest. These biases result from both classical measurement error and more structured measurement error, which we find is common in machine learning based remotely sensed measurements. We show that multiple imputation, a standard statistical imputation technique so far untested in this setting, effectively reduces bias and improves statistical coverage with only minor reductions in power in both simple linear regression and panel fixed effects frameworks. Our results demonstrate that multiple imputation is a generalizable and easily implementable method for correcting parameter estimates relying on remotely sensed variables.
Warming temperatures tend to damage crop yields, yet the influence of water supply on global yields and its relation to temperature stress remains unclear. Here, we use satellite-based measurements to provide empirical estimates of how root-zone soil moisture and surface air temperature jointly influence the global productivity of maize, soybeans, millet, and sorghum. Relative to empirical models using precipitation as a proxy for water supply, we find that models using soil moisture explain 30% to 120% more of the interannual yield variation across crops. Models using soil moisture also better separate water supply stress from correlated heat stress and show that soil moisture and temperature contribute roughly equally to historical variations in yield. Globally, our models project yield damages of -9% to -32% across crops by end-of-century under SSP5-8.5 from changes in temperature and soil moisture. By contrast, projections using temperature and precipitation overestimate damages by 28% to 320% across crops both because they confound stresses from dryness and heat and because changes in soil moisture and temperature due to climate change diverge from their historical association. Our results demonstrate the importance of accurately representing water supply for predicting changes in global agricultural productivity and for designing effective adaptation strategies.
Selected coverage: Harvard SEAS News
Burney, J., Persad, G., Proctor, J., Bendavid, E., Burke, M., Heft-Neal, S. (2022). “Geographically-resolved social cost of anthropogenic emissions accounting for both direct and climate-mediated effects." Science Advances (here)
The magnitude and distribution of physical and societal impacts from long-lived greenhouse gases are insensitive to the emission source location; the same is not true for major co-emitted short-lived pollutants like aerosols. Here we combine novel global climate model simulations with established response functions to show that a given aerosol emission from different regions produces divergent air quality and climate changes and associated human system impacts, both locally and globally. The marginal global damages to infant mortality, crop productivity, and economic growth from aerosol emissions and their climate effects differ by more than an order of magnitude depending on source region, with certain regions creating global external climate changes and impacts much larger than those felt locally. Importantly, the complex distributions of aerosol-driven societal impacts emerge from geographically-distinct and region-specific aerosol-climate interactions, estimation of which is enabled for the first time by the full earth system modeling framework used here.
Chan, D., Rigden, A., Proctor, J., Chan, P. W., Huybers, P. (2022). “Differences in Radiative Forcing, Not Sensitivity, Explain Differences in Summertime Land Temperature Variance Change Between CMIP5 and CMIP6." Earth's Future (here)
How summertime temperature variability will change with warming has important implications for climate adaptation and mitigation. CMIP5 simulations indicate a compound risk of extreme hot temperatures in western Europe from both warming and increasing temperature variance. CMIP6 simulations, however, indicate only a moderate increase in temperature variance that does not covary with warming. To explore this intergenerational discrepancy in CMIP results, we decompose changes in monthly temperature variance into those arising from changes in sensitivity to forcing and changes in forcing variance. Across models, sensitivity increases with local warming in both CMIP5 and CMIP6 at an average rate of 5.7 ([3.7, 7.9]; 95% c.i.) × 10−3°C per W m−2 per °C warming. We use a simple model of moist surface energetics to explain increased sensitivity as a consequence of greater atmospheric demand (∼70%) and drier soil (∼40%) that is partially offset by the Planck feedback (∼−10%). Conversely, forcing variance is stable in CMIP5 but decreases with warming in CMIP6 at an average rate of −21 ([−28, −15]; 95% c.i.) W2 m−4 per °C warming. We examine scaling relationships with mean cloud fraction and find that mean forcing variance decreases with decreasing cloud fraction at twice the rate in CMIP6 than CMIP5. The stability of CMIP6 temperature variance is, thus, a consequence of offsetting changes in sensitivity and forcing variance. Further work to determine which models and generations of CMIP simulations better represent changes in cloud radiative forcing is important for assessing risks associated with increased temperature variance.
Proctor, J. (2021). “Atmospheric opacity has a nonlinear effect on global crop yields." Nature Food (here and here)
Theory suggests that crop yields may have a concave response to changes in light from an increasingly opaque atmosphere, yet such an effect has never been empirically estimated. In this paper I leverage year-to-year variation in growing season cloud optical thickness to provide the first non-linear empirical estimates of how increased atmospheric opacity alters the quantity and quality of sunlight across the Earth’s surface and how this, in turn, affects maize and soy yields in the United States, Europe, Brazil and China. The central contribution of this paper is to show that the response of yield to changes in sunlight from cloud scattering and absorption is consistently concave across crops and regions, and that this response is driven by a concave response of yields to total insolation as well as by benefits from increased diffuse light in some regions. I also find that the benefits of low and moderate opacity cannot be explained by the diffuse fertilization effect alone. Rather, other benefits from moderated insolation must play a central role, such as reduced leaf temperature, increased water use efficiency or reduced photoinhibition. Globally, these results suggest that agricultural productivity is limited as much by light as by temperature. Applying these empirical estimates to earth system model simulations of air pollution and climate change, I find that anthropogenic changes in the quantity and quality of sunlight due to changes in clouds have had, and may increasingly have, economically substantial impacts on yields.
Selected coverage: Harvard SEAS News
Carleton*, T., Cornetet*, J., Huybers*, P., Meng*, K. & Proctor*, J. (2021) “Evidence for Ultraviolet Radiation Decreasing COVID-19 Growth Rates: Global Estimates and Seasonal Implications.” PNAS (here) (* indicates equal contribution)
With boreal winter approaching, there is a need to understand how local environmental conditions may modify COVID-19 transmission. In this paper we combine a spatially-resolved dataset of confirmed COVID-19 cases, composed of 3,235 regions across 173 countries, with local environmental conditions and a distributed lag panel regression model. The central scientific contribution of this work is to quantify the individual influences of UV, temperature, humidity and precipitation on the growth rate of COVID-19 accounting for lagged effects and potential confounders. We find that a standard deviation increase in UV lowers the daily growth rate of COVID-19 cases by ∼1 percentage point over the subsequent 2.5 weeks, relative to an average in-sample growth rate of 13 percent. The time pattern of lagged effects peaks 9-11 days after UV exposure, consistent with the combined timescale of incubation, testing, and reporting. Simulations illustrate how seasonal changes in UV have influenced regional patterns of COVID-19 growth rates from January to June, indicating that UV has a substantially smaller effect on the spread of the disease than social distancing policies.
Rolf*, E., Proctor*, J., Carleton*, T., Bolliger*, I., Shankar*, V., Ishihara, M., Recht, B. & Hsiang, S. (2021) “A Generalizable and Accessible Approach to Machine Learning with Global Satellite Imagery.” Nature Communications (here)
Combining satellite imagery with machine learning (SIML) has the potential to address global challenges by remotely estimating socioeconomic and environmental conditions in data-poor regions, yet the resource requirements of SIML limit its accessibility and use. The central contribution of this paper is to demonstrate for the first time that a relatively simple unsupervised featurization can achieve performance across many tasks (e.g., forest cover, house price, road length…etc.) and at global scale competitive with more complex neural network-based algorithms. Our method achieves accuracy competitive with deep neural networks at orders of magnitude lower computational cost, scales globally, delivers super-resolution predictions, and facilitates characterizations of uncertainty. Since image features are shared across tasks, they can be centrally computed and distributed to unlimited researchers, who need only fit a linear regression to their own ground truth data in order to achieve state-of-the-art SIML performance. This reduces the challenge of extracting information from unstructured satellite imagery to a linear regression problem using pre-computed features. We are currently producing and publicly distributing a dense featurization of the globe. We seek to enable a greater number and diversity of researchers to make state-of-the-art SIML predictions by lowering the computational and intellectual costs required to do so.
Proctor*, J., Hsiang*, S., Burney, J., Burke, M., & Schlenker, W. (2018). “Estimating global agricultural effects of geoengineering using volcanic eruptions.” Nature, 560 (7719), 480. (here)
In this paper we leverage the major volcanic eruptions of El Chichón (1982) and Pinatubo (1991) as natural experiments to estimate the impact that solar geoengineering might have on global agricultural production. We show that changes sunlight from increased concentrations of stratospheric sulfate aerosols decrease both C4 (maize) and C3 (soy, rice, and wheat) yields. By applying our empirical yield model to earth system model simulations, we demonstrate that solar geoengineering damages from reduced sunlight wash out the technology’s benefits from cooling. A key scientific contribution of the paper is to provide the first global empirical test of the “diffuse fertilization effect” hypothesis for crop yield. We find that this theorized benefit of aerosol scattering to crop yields – namely increased photosynthetic efficiency due to a more even distribution of light throughout the canopy – was more than offset by the reduction in total sunlight. This informs not only geoengineering policy, but also regulation of other anthropogenic influences on the global optical environment such as greenhouse gas or particulate matter emissions.