With the growing climate crisis, extreme climate events are becoming more frequent and devastating [1]. Disadvantaged groups, including people of color and low-income households, bear the brunt of the costs of these climate disasters, including deaths, destroyed homes [2], and long-term economic and health consequences [3]. These unevenly distributed impacts can be attributed to a range of factors such as differences in geographical location, housing conditions, access to resources, which are largely rooted in a long history of institutional oppression. Furthermore, the people who experience the most vulnerability to climate change are often those with multiple marginalized identities [4], which can be understood by the framework of intersectionality. As such, to guide more equitable decision making and distribution of resources in response to extreme climate events, we need to consider socioeconomic factors in our understanding of climate disaster impacts. This project uses a data-driven approach and leverages machine learning models to detect patterns in how climate disasters impact different demographic groups and identify relationships between climate disasters and inequality in the U.S. 

 

5 datasets were used to support a multifaceted understanding of this topic: climate risk index by county, demographic social vulnerability factors to by county, surveys on how individuals were impacted in the aftermath of hurricanes, and climate disaster news headlines. This data ranges from being qualitative to quantitative, geospatial to personal, and represents different sources, from government agencies to hurricane survivors’ first-hand experiences to media coverage, enabling rich insights. Furthermore, all datasets (or merged datasets) included features related to both climate disasters and socioeconomic demographics. 7 machine learning models were applied to these datasets. These included 3 unsupervised learning models, which identify underlying patterns and structures through grouping or association, as well as 4 supervised models, which map relationships between inputs and a target output in order to predict output values for new, unseen data. These models applied to various datasets all identified patterns with greater climate disaster impacts being distributed to low-income populations and people of color, generating unique findings that contributed different pieces to this story of data-driven climate justice.

​

K-means clustering and regression analysis were used to discover relationships between where socially vulnerable populations live and where there is high climate risk. K-means clustering identified 3 groups from the structure of the data: 1. counties with high poverty and black populations, with medium total climate risk and high heat wave risk, 2. counties with medium poverty, high Hispanic populations, with high total climate risk, and 3. counties with low poverty, black and Hispanic populations and low climate risks. This suggests that geographical locations with large low-income, black and Hispanic populations also tend to have highest climate impacts. The regression analysis also determined that there was a positive correlation between 7 of 8 combinations of the socioeconomic variables and climate risk variables modeled. The linear regression models showing the largest increases in climate risk for percent population increase are: the total climate risk vs. percentage Asian population, heat wave risk vs. percentage Asian population (these may be attributed to the smaller overall Asian-American population), heat wave risk vs. percentage African-American population, total climate risk vs. percentage African-American population, and total climate risk vs. percentage Hispanic population. These findings could be used to guide policy decisions and allocation of resources, for example, prioritizing heat wave-related aid to counties with high black population, which were determined to be highly impacted by heat wave related risks in both models. These results support the understanding that climate disasters disproportionately affect socially vulnerable populations based on where they are geographically located.

​

Association rule mining, naive Bayes, and decision tree models were employed to find patterns in how different demographic groups are affected by hurricane impacts, as well as predict an individual’s ability to fully recover from hurricane impacts a year later. Association rule mining found that the strongest associations in the data were between an individual being aged 51-75, female, and/or experiencing high adversity due to the hurricane impacts—and being black. In this analysis, high adversity corresponds to experiencing at least 3 hurricane impacts, including not meeting all essential expenses, not paying the full rent or mortgage, being evicted, and not having adequate food. The naive Bayes model was able to correctly classify whether an individual would fully recover from hurricane impacts a year later based on reported storm impacts and the individual’s demographics 71.43% of the time, while the decision tree models reached an accuracy of 73.27%. The naive Bayes model indicated that the most important predictor of recovery was home damage due to the hurricane, but income and race predictors were the next most important predictors, even more so than experiencing reduced work hours due to the storm. Furthermore, the conditional probability of recovery is 1.5 times as high given someone is white compared to if they are African American, and the conditional probability of not recovering is 1.5 times as high given someone is living in poverty compared to not living in poverty.

​

Finally, the support vector machine model found that in the reporting of climate events and inequality, the appearance of the word “racial” appeared as a strong predictor, possibly suggesting the significance of racial inequality in hurricane impacts. These results all build on each other to paint the picture of how climate disasters disproportionately impact low income populations and people of color, due to a range of factors, and presenting a range of human consequences. Furthermore, several results support an intersectional understanding of climate impacts—that people with more than one marginalized identity often face the greatest risks of climate change, due to intersecting forces of privilege and oppression. For example, the association rule mining model reveals how multiple socioeconomic vulnerabilities such as being both female and black, and experiencing high adversity from Hurricanes Harvey, Nate, and Imra, are strongly connected. Despite the revealing results and the need for further data-driven climate justice studies, there were various challenges in this project, in particular, finding or creating datasets that included both socioeconomic and climate change variables and with sufficient data on underrepresented socioeconomic groups. This informs the importance of collecting more climate data with socioeconomic attributes and incorporating this data in our models on climate impacts, in order to understand social disparities in these impacts and guide more equitable decision making.

​

References:

1. NOAA National Centers for Environmental Information (NCEI) (2023). "U.S. Billion-Dollar Weather and Climate Disasters (2023)." https://www.ncei.noaa.gov/access/billions/

2. Cho, R. (2020). "Why Climate Change is an Environmental Justice Issue." Columbia Cliate School. https://news.climate.columbia.edu/2020/09/22/climate-change-environmental-justice/

3. Toldson, I. A., Ray, K., Hatcher, S. S., & Straughn Louis, L. (2011). Examining the Long-Term Racial Disparities in Health and Economic Conditions Among Hurricane Katrina Survivors: Policy Implications for Gulf Coast Recovery. Journal of Black Studies, 42(3), 360-378. https://doi.org/10.1177/0021934710372893

4. EPA (2023). "Climate Change and Children’s Health and Well-Being in the United States." U.S. Environmental Protection Agency. https://www.epa.gov/cira/climate-change-and-childrens-health-report

​

​