Predicting spatiotemporally-resolved air temperature over Sweden from satellite data using an ensemble model

Abstract

BACKGROUND AND AIM: Mapping of air temperature (Ta) at high spatiotemporal resolution is critical to reduce exposure assessment errors in epidemiological studies on the health effects of air temperature. We aimed to predict daily Ta at a spatial resolution of 1×1 km2 from 2001 to 2019 using an ensemble model based on the satellite-based daily land surface temperature. The results will provide insights in modeling Ta with an ensemble model to reduce exposure measurement error and serve as a foundation for epidemiological studies on the short-term and long-terms effects of air temperature on public health at a large geographical scale. METHODS: The generalized additive model based ensemble model incorporated four base models, including a generalized additive model, a generalized additive mixed model, and two machine learning models (random forest and extreme gradient boosting), and allowed the weights for each model to vary over space, with the best-performing model for each grid cell assigned the highest weight. RESULTS:The ensemble model showed high performance with an overall R2 of 0.98 and a root mean square error of 1.38 °C in the ten-fold cross-validation, and outperformed each of the four base models. Among base models, the two machine learning models outperformed the two regression models. In the machine learning models, Ts was the dominant predictor of Ta, followed by day of year, Normalized Difference Vegetation Index, latitude, elevation, and longitude. CONCLUSIONS:The GAM-based ensemble model exploited the predictive ability of base models by allowing the weights for each model to vary over space, and provided an improved estimation of air temperature over Sweden that outperformed estimations from each base model. Among the base models, two machine learning methods (RF and XGBoost) exhibited higher predictive power than two linear regression models. The estimations from ensemble models can be applied in epidemiological studies to minimize bias caused by exposure misclassification. KEYWORDS: random forest, extreme gradient boosting, generalized additive model, ensemble model, air temperature, health exposure