Multimodal Satellite and Environmental Exposure Dataset for US County Life Expectancy (CONUS 2003–2019): Predictive Multiplicity in Environmental Health

Abstract

Multimodal Satellite and Environmental Exposure Dataset for US County Life Expectancy (CONUS 2000–2023): Predictive Multiplicity in Environmental Health This repository provides the harmonized multimodal environmental dataset supporting the manuscript "Predictive Multiplicity in Environmental Health: Information Theory Exposes How Summary Statistics Shape Machine Learning Attribution" (Ahmad et al., submitted to Sensors, MDPI, 2026). Dataset Overview: Primary File: Final_ML_Ready_Dataset.csv.zip (contains Final_ML_Ready_Dataset.csv). Dimensions: 74,863 county-year records x 1,318 columns spanning 2000 to 2023 across the contiguous United States (CONUS). Study Window: 52,309 county-years across 3,077 counties (2003–2019) used for the core gradient-boosted modeling and explainable AI (xAI) attribution benchmarks. Outcome: County-level life expectancy at birth derived from Institute for Health Metrics and Evaluation (IHME) small-area estimations. Sensing & Environmental Modalities: The panel synthesizes 18 multimodal environmental observation streams reduced to county polygons without census or clinical covariates: Thermal & Optical: MODIS Terra land-surface temperature (MOD11A2 day/night clear-sky composite at 1 km) and Landsat 5/7 surface reflectance. Atmospheric Chemistry: Copernicus Atmosphere Monitoring Service (CAMS) global reanalysis (EAC4) trace gases (formaldehyde, NO2, O3, CO, SO2, isoprene, propane, PAN, VOCs). Thermodynamics & Meteorology: ERA5 reanalysis and PRISM/gridMET 4-km interpolated surface climate (temperature, dew point, vapor pressure deficit, precipitation). Surface & Terrestrial: SoilGrids soil physical/chemical properties, JRC Global Surface Water, Global Human Modification index (gHM), harmonized nighttime lights, and USDA Cropland Data Layer. Engineered Physics Metrics: Night-time cooling efficiency (NCE; tail and mean constructions), nocturnal wet-bulb temperature, heat-stress indices, and photochemical ratios. Code & Reproducibility: All extraction, training, ablation refits, and figure generation scripts are publicly available on GitHub: https://github.com/albertfaiz/how-long-not-how-much