Publications

In applications across agriculture, ecology, and human development, machine learning with satellite imagery (SatML) is limited by the sparsity of labeled training data. While satellite data cover the globe, labeled training datasets for SatML are often small, spatially clustered, and collected for other purposes (e.g., administrative surveys or field measurements). Despite the pervasiveness of this issue in practice, past SatML research has largely focused on new model architectures and training algorithms to handle scarce training data, rather than modeling data conditions directly. This leaves scientists and policymakers who wish to use SatML for large-scale monitoring uncertain about whether and how to collect additional data to maximize performance. Here, we present the first problem formulation for the optimization of spatial training data in the presence of heterogeneous data collection costs and realistic budget constraints, as well as novel methods for addressing this problem. In experiments simulating different problem settings across three continents and four tasks, our strategies reveal substantial gains from sample optimization. Further experiments delineate settings for which optimized sampling is particularly effective. The problem formulation and methods we introduce are designed to generalize across application domains for SatML; we put special emphasis on a specific problem setting where our coauthors can immediately use our findings to augment clustered agricultural surveys for SatML monitoring in Togo.

Understanding food demand patterns is critical for developing countries in Sub-Saharan Africa (SSA), which are facing increasing urbanization, intensifying climate change, and rapid transitions in food production systems. This study models food demand in Zambia and Tanzania using data from the Living Conditions Monitoring Survey (LCMS) (2015 and 2022) and the Living Standards Measurement Survey (LSMS) (2015 and 2021). This study applies the Quadratic Almost Ideal Demand System (QUAIDS) model to analyze demand behavior and calculate expenditure and price elasticities of both food and nutrition. Urbanization is linked to significant shifts in dietary preferences and greater sensitivity to the prices of non-staple, nutrient-dense foods, reflecting the diversifying food demand in urbanized regions. Households allocate more resources to calories with increases in income, but the rate of increase slows as basic caloric needs are met. Conversely, expenditure nutrition elasticity for Vitamin B12 is particularly high, as rising incomes enable households to spend more on animal-based foods such as meat, fish, eggs, and dairy. Once basic caloric needs are satisfied, additional income is more likely to be directed toward diversifying the diet with higher-value, nutrient-dense foods rather than increasing carbohydrate consumption. The positive expenditure elasticities for both food consumption and nutrient intake in both countries emphasize that increasing household income remains a crucial strategy for improving food security.

Large-scale deployment of wind and solar power is essential for climate change mitigation but requires substantial land and offshore area due to their relatively low installed capacity densities. However, wide variation in installed capacity density estimates creates major uncertainty in projecting future spatial requirements. Here, we apply multiple spatial methods to satellite imagery to estimate installed capacity densities for 60 onshore wind, 24 offshore wind, and 54 utility-scale solar PV plants (including 36 fixed-tilt and 18 single-axis tracking) across six countries: Australia, China, Germany, India, United Kingdom, and the United States. For wind power plants, mean estimates of installed capacity density using the 8D buffer method (eight times the rotor diameter), the convex hull method with a 5D buffer, and the median Voronoi polygon (bounded by an 8D buffer) method are similar (4.0, 4.4, and 4.9 MW km⁻² for onshore wind and 4.6, 5.2, and 5.4 MW km⁻² for offshore wind, respectively) but the 5D buffer method provides a relatively larger estimate because of unaccounted areas within power plant boundaries. Single-axis tracking solar PV power plants have an approximately 33% lower mean installed capacity density than fixed-tilt power plants, with estimates of 36 and 29 MW km⁻² (without and with a 50 m buffer, respectively), compared to 54 and 43 MW km⁻² for fixed-tilt systems. Estimates vary more within countries and across methods than between countries, underscoring the importance of methods and assumptions in estimating spatial requirements for large-scale wind and solar deployment.

Rigorous evidence detailing the myriad interconnections between humans and ecosystems will be critical to slow the loss of biodiversity. Effective conservation interventions will depend upon a detailed understanding of the benefits that biodiversity provides to people, the ways that human activities drive biodiversity decline, and the potential for conservation policies to stem this decline. Although hundreds of papers explore these relationships, a careful review of this literature shows that the overwhelming majority of studies fall short of documenting causal relationships with biodiversity per se. However, a combination of data and methodological advances has led to rapid growth of quasi-experimental analyses that are advancing our understanding of human interactions with biodiversity in three domains. First, economists have provided valuable insights into the causes of biodiversity loss, which complement traditional ecological experiments by better reflecting real-world conditions. Second, quasi-experimental studies have begun to identify which policy interventions, in what contexts, have slowed biodiversity loss. Finally, recent quasi-experimental studies have shown that the loss of species can impose extremely large costs on humanity but that these costs vary widely depending upon the species and opportunities for human adaptation.