Data from: Attributing Marginal Contribution of Biophysical Differences to Neighborhood-Scale Heat Island Disparities with Explainable Machine Learning

Summary Numerous local to national scale studies have demonstrated that different population groups in the U.S. have distinct burdens of urban heat. However, these studies have primarily used linear models to map the outcome (generally the urban heat island effect) to its biophysical predictors. To capture non-linear effects of biophysical factors and further attribute them to these disparities, we implement an explainable machine learning approach to study surface urban heat island (SUHI) intensity at the neighborhood scale for all major U.S. urban areas. Similar to prior research, we find that Black and Hispanic people live in neighborhoods with higher SUHI than white people on average, even when accounting for economic factors, which is largely explained by biophysical differences. However, finely disaggregating SHAP (SHapley Additive exPlanations) values across race, socioeconomic indicators, and climate regions, we find that these biophysical factors vary dramatically by demographics nationally. Some patterns still hold from previous studies: the oft-studied vegetation index, is a major driving factor of SUHI disparities for Black and Hispanic residents, especially those who are low income. However, while built-up extent is the second biggest warming factor for Black residents nationally, lower normalized difference built-up index indicators seems to serve a cooling effect for Hispanic residents in urban arid regions. In addition, lower elevation for Hispanic people in arid regions seems to result in a warming effect. Our findings have implications for data-driven, equity-oriented urban and regional policy concerning heat mitigation strategies. Key features The data represents almost all cities in the U.S. (493 out of 497) at the census tract level (2010 version). Data quality assurance resulted in the removal of 4 cities with excessive amounts of missing data. The physical covariates are 1km resolution, which is appropriate for capturing intra-urban heterogeneity. Files data.csv contains the covariates for data exploration and modeling for the SUE method (2015-2019). Centroid_USUHI.csv contains the centroids for each city Metadata data.csv UHI: Land Surface Temperature Anomalies (in degrees Celsius) NDVI: Normalized Difference Vegetation Index NDBI: Normalized Difference Built-up Index BSA: Black Sky Albedo DEM: Digital Elevation Model (in meters) Climate Zone: Köppen-Geiger climate zone (arid, snow, temperate, or tropical) Coastal?: Whether a tract or urban area is coastal or inland Area (in square meters). We use the following naming conventions: _rur: Rural reference (Spatial mean of the non-urban, non-water pixels within the region of interest) _CT_act: Spatial mean of all non-water pixels intersecting the Census Tract (one value per census tract) For the anomalies: Take the difference between the urban tract value and rural reference value. See the GitHub link (particularly 1-AggregateData-SUE.ipynb) for a complete breakdown. Methodology and Data Sources For further details on methodology and data source, please consult the publication.