Detecting the effect of dust and other climate variables on crop yields using diagnostic statistical crop models

Open Access
- Author:
- Hoffman, Alexis Lee
- Graduate Program:
- Meteorology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- February 12, 2018
- Committee Members:
- Chris Eliot Forest, Dissertation Advisor/Co-Advisor
Chris Eliot Forest, Committee Chair/Co-Chair
Marcelo Chamecki, Committee Member
Gregory S Jenkins, Committee Member
Armen R. Kemanian, Outside Member
Natalie Mahowald, Special Member - Keywords:
- climate
dust
crop yields
statistical crop modeling
food security
agrometeorology
random forest - Abstract:
- Food security and agriculture productivity assessments require a strong understanding of how climate and other drivers influence regional crop yields. While the effects of temperature, precipitation, and carbon dioxide are relatively well-understood, the effect of dust on crop yields has yet to be thoroughly investigated. This line of inquiry is warranted because many areas of the world with frequent dust storms and high dust loadings are often food insecure, and because wind erosion is prevalent in the High Plains of the United States, a major crop-producing area. Existing research suggests that the effect of dust on yields should be largely negative, but until now this has not been investigated on a regional scale. A major hindrance to understanding the effect of dust on crop yields is insufficient data and inadequate methods of analysis. In this dissertation, we developed data and analysis methods for three distinct projects to determine whether dust affects regional crop yields. In the first project, we validated the use of random forest, a machine learning technique, as a diagnostic crop model that can be used to assess the impact of individual climate predictors on yields. Because we motivated this research with food security concerns, we analyzed climate signals in the crop yield record of sub-Saharan Africa from 1962-2014. From this work, we determined that random forest could function as a statistical crop model, but the data quality and resolution inhibited the ability to detect the effect of dust on yields in this area of the world. In our second line of inquiry, we shifted the focus to the central region of the United States for its high quality and high resolution data, as well as its importance as a major crop-producing region of the world. Because these data had higher temporal resolution, we could explore individual phases of the growing season. We developed crop-specific algorithms to compute the planting date, establishment phase, critical window, and grain filling phase to investigate yield responses to phase-specific climate predictors. Using these data, the random forest identified distinct phase-specific responses for important climate predictors. Finally, we computed dust metrics from three different data sources and merged them with climate and yield data in the central region of the United States to estimate the impact of dust on yields. Over the entire central US region, we found that including dust as a predictor in each crop model did not improve yield predictions for the region as a whole. However, when crop models were applied to individual states, we found several instances in which dust weakly reduced yields. Although these state-specific results were encouraging, we presented them cautiously because the yield responses could be an artifact of either partitioning the data or a true yield response that is obscured when data was spatially aggregated. While the results were largely inconclusive, we have advanced the capabilities of statistical crop modeling, developed data sets that can be used to move the science forward, and revealed new questions that merit further research.