Data-driven improvements to climate risk assessment for agriculture: hazards, exposure, and vulnerability

Abstract

Understanding climate risk is crucial for accurately estimating the impacts of climate change on agricultural production and for increasing resilience via effective adaptation actions. Climate risk considers the interaction of physical climate hazards (i.e., sources of harm or danger), exposure to these hazards (i.e., geographical locations), and the relative vulnerability (i.e., susceptibility or adaptive capacity) of the affected systems. Agriculture, among all economic sectors, is uniquely sensitive to both climatic variability and socio-economic factors. While the impacts of physical climate hazards (e.g., heat stress) on crops and animals are driven by biophysical changes, agricultural producers can adjust the consequences of those impacts by changing their land locations (e.g., migration of farmland) or adaptive capacities (e.g., changes in agronomic practices). Yet, empirical applications of this full risk framework to agricultural systems remain scarce. This dissertation advances an integrated, data-driven foundation for climate risk assessment in US agricultural systems, encompassing both crop and milk production from 1981 to 2020. It develops a new empirical models and large-scale datasets to quantify how hazards, exposure, and vulnerability can interact to shape yield outcomes and economic losses. The US serves a critical case study due to its dual role as the world's largest milk producer and a leading grain supplier, with implications for global food security. Following a brief introduction, Chapter 2 quantifies yield sensitivities and corresponding spatio-temporal yield losses of US rainfed maize, soybeans, sorghum, and spring wheat to hydroclimatic stresses between root-zone soil moisture and atmospheric evaporative demand from 1981–2020. This chapter presents that crop yields can be reduced similarly by two major hydroclimatic hazards, which are defined as the most yield damaging conditions over time: 'Low Supply+High Demand' (i.e., drought) and 'High Supply+Low Demand' (i.e., waterlogging). However, more exposure to 'Low Supply+High Demand' hazard led to the largest annual yield losses (7–17%) across all four crops over time. Modeled yield losses due to these hazards are significantly associated with crop insurance lost costs. The extent of yield losses varies considerably by crop and location, highlighting the need for crop-specific and regionally tailored adaptation strategies. The next chapter comparatively analyzes the performance of existing thermal indices in capturing US milk yield response to both cold and heat stress at the national scale. Four commonly used thermal indices are selected: the Temperature and Humidity Index (THI), Black Globe Humidity Index (BGHI), Adjusted Temperature and Humidity Index (THIadj), and Comprehensive Climate Index (CCI). Using a statistical panel regression model with observational and reanalysis weather data from 1981–2020, this chapter systematically compares the patterns of yield sensitivities and statistical performance of the four indices. Results show that the US state-level milk yield variability is better explained by the THIadj and CCI, which combine the effects of temperature, humidity, wind, and solar radiation. This chapter also reveals a continuous and nonlinear responses of milk yields to a range of cold to heat stress across all four indices. This implies that solely relying on fixed thresholds of these indices to model milk yield changes may be insufficient to capture cumulative thermal stress. Cold extremes reduced milk yields comparably to those impacted by heat extremes on the national scale. Additionally, it shows large spatial variability in milk yield sensitivities, implying further limitations to the use of fixed thresholds across locations. Moreover, it presents decreased yield sensitivity to thermal stress in the most recent two decades, suggesting adaptive changes in management to reduce weather-related risks. Chapter 4 combine 155 million test-day milk records from 9 million US cows from 2000–2024 with a causality aware machine learning model to quantify the complex yield responses to temperature, radiation, humidity, and wind. Results show that high nighttime temperatures are the dominant stressor. Nationwide, heat stress reduced yields by 2.5% whereas cold-induced loss was 1.0%. Late-summer heat and springtime cold caused the largest losses, which were –4.3% and –1.5%, respectively. These losses lead to economic damages of $1.1 billion per year, with California and Wisconsin accounting for ~40% of national losses. This burden shifts to southern states for heat and northern states for cold when expressing the damage per state-level revenues, highlighting region-specific susceptibility. This chapter establishes the nonlinear nature of yield sensitivity to multiple weather stressors and reveals opportunities to enhance yields in US dairy systems. The final chapter presents a national-scale, multi-decadal assessment of agricultural social vulnerability and exposure across the US from 1981 to 2020. Using USDA Census data, this chapter construct an Agricultural Social Vulnerability (Ag-SVI) and link it with physical location of cropland to evaluate spatial and temporal dynamics of their interactive risk potential. The analysis reveals pronounced regional contrasts between low and high vulnerability, shaped by differences in demographic, socio-economic, farm structure, and infrastructural factors. The southern US is a hotspot of high vulnerability driven by large share of non-white producers, love education attainment, and persistent poverty. On the other hand, low vulnerability clusters in the Midwest and northern Great Plains with better infrastructural and farm structural factors. The improvements in infrastructural vulnerability contributed to increasing the national-level overall Ag-SVI from 0.52 in 1981 to 0.46 in 2020. Intersecting social vulnerability with exposure identifies new hotspots of high vulnerable areas in the lower Mississippi river basin, the High Great Plains, the Appalachian corridor, and central California. This chapter highlights that US agricultural vulnerability remains at a moderate level overall but it is geographically heterogeneous, offering new insights into where and why resilience gains remain uneven and how they can inform future adaptation strategies.