Température, Pollution Atmosphérique et Santé Périnatale dans le Contexte du Changement Climatique : Exposition et Impacts dans les Populations Sensibles

Studies suggest that exposure to heat, cold, or particulate matter (PM) during pregnancy may decrease birth weight and shorten the duration of gestation, but we do not fully understand these associations. This is concerning because low birth weight and preterm birth increase risk for numerous morbidities and chronic diseases throughout childhood and into adulthood. The health burdens of temperature and PM are expected to increase with climate change and global urbanization. Improving our understanding requires accurate exposure assessment because error can bias associations with health outcomes. A scalable way to assess exposure over large areas and long time periods is to model the relationship between monitored temperature or PM and satellite-derived variables such as land surface temperature (LST) or aerosol optical depth (AOD). Improved calibration methods might increase accuracy, and finer spatial resolutions could be useful in urban areas.This PhD developed novel approaches to increase the accuracy and spatial resolution of satellite-based exposure models, applied them to reconstruct daily temperature and PM in France over the past two decades, and used the temperatures to identify periods during pregnancy when exposure to cold or heat may increase the risk of preterm birth.First, we found that allowing the relationship between monitored ambient temperature and satellite-derived LST to vary over space as well as time improves the accuracy of mixed models. We used this approach to model diurnal air temperature (mean [Tmean], minimum [Tmin], and maximum [Tmax]) over France at 1 km resolution from 2000 to 2018. We also showed that an ensemble of random forests and gradient boosting machines can downscale daily temperature over urban areas by incorporating thermal satellite data at a high spatial but low temporal resolution. We used this approach to model daily Tmean, Tmin, and Tmax at 200 m resolution over 103 cities.Second, we found that Gaussian Markov random fields can reconstruct PM based on gap-filled AOD more accurately than widely used mixed models or random forests, and that an ensemble of all three algorithms performs even better. We used this approach to model daily mean PM2.5 and PM10 (PM <2.5 and