"The Satellite and the Territory: Essays on Remote Sensing and the Econ" by Matthew Gordon

Date of Award

Fall 2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Economics

First Advisor

Chertow, Marian

Abstract

The increasing availability of satellite data is creating unprecedented opportunities to study sustainable development processes at both a global scale and fine temporal and spatial resolution. Economists and environmental researchers are using satellite data products to measure land-use change, economic growth, household vulnerability, air pollution, and more. When these data products are used in research or policy-making uncritically, however, the conclusions that follow can be inappropriate. This dissertation examines the potential opportunities and the shortcomings of three different uses of satellite data to inform sustainable development policy -- targeting policy, causal inference, and measurement. The first chapter studies the question of how to target aid after a natural disaster. Disaster aid programs often use property damage as a criterion for eligibility, and satellite data could be useful in collecting rapid and accurate data on property damages. On the other hand, a household's ability to insure against shocks is harder to observe, but it may be more important in determining how the disaster affects welfare. I develop a model of household demand for reconstruction aid, incorporating both the exposure to a shock and the ability to borrow for consumption smoothing, and I calibrate the model using household survey data following the 2015 earthquake in Nepal. I use a spatial discontinuity in the distribution of reconstruction aid to test the model's assumptions, and I find that aid increases consumption and housing investment, but decreases remittances, consistent with a model of incomplete insurance. I use the calibrated model to estimate the benefits of counterfactual aid allocations. Conditioning aid on household property damage barely outperforms allocating aid at random. The property damage criterion excludes many liquidity-constrained households that have high demand for aid, and it includes wealthy, well-insured households that have low demand. An untargeted approach that divides the aid budget equally between all households in the affected areas yields substantially larger welfare gains. This shows that using satellite data to assess physical damages for targeting purposes is thus unlikely to add much value. Chapter two studies the use of satellite data for causal inference with Luke Sanford, Megan Ayers, and Eliana Stone. Advances in machine learning and the increasing availability of satellite imagery have led to the proliferation of social science research that uses remotely sensed measures of human activity or environmental outcomes to infer the impact of policy. However, when machine learning models minimize a standard loss function, the predictions they generate can produce biased estimates when used for causal inference. In this paper, we show how this bias can arise, and we demonstrate the use of an adversarial debiasing algorithm in order to correct this issue when generating machine learned predictions for use in causal inference. This method is widely applicable beyond satellite data to any setting where machine learned predictions are the dependent variable in a regression. We conduct simulations and empirical exercises using data on forest cover in Western Africa, to show that our method generates predictions that recover the same regression parameters that are estimated from ground truth data, whereas standard models result in biased predictions. We then use the method to study the relationship between economic development and deforestation in Western Africa, and we find a positive relationship between increases in wealth and forest cover. Chapter three studies the use of satellite data for measurement of pollution and externalities with Anna Papp. We study mismanaged plastic pollution, a rapidly growing source of environmental and economic damages. We seek to quantify the extent to which the international trade in waste contributes to the accumulation of plastic waste in the natural environment. To overcome the lack of data on plastic waste at its many destinations, we create a novel time series of open-air landfill area by country, using crowd-sourced training data, machine learning on satellite imagery, and an efficient stratified random sampling approach to ensure unbiased estimates. Our results show suggestive increases in open-air dumps in countries that received increased waste imports after China banned imports of plastic waste in 2018. In addition to the novel methodology for measuring a rare-land use type, our results suggest that trade may play a role in overwhelming local waste management systems and the subsequent leakage of plastic into the environment. This dissertation advances the study of environment and development economics by mapping the opportunities and limitations of an increasingly important data source. Satellite data can play an important role in measuring natural capital, monitoring economic activity in near-real time, and measuring economic and environmental variables where ground-truth data is unavailable. I show that as typically used, satellite data can lead to biased policy prescriptions, however. I provide new techniques that will improve the use of this data to inform policy, and I apply the techniques to important questions in the economics of sustainable development.

Download

Share

COinS