Hough, I; Just, AC; Zhou, B; Dorman, M; Lepeule, J; Kloog, I (2020). A multi-resolution air temperature model for France from MODIS and Landsat thermal data. ENVIRONMENTAL RESEARCH, 183, 109244.

Understanding and managing the health effects of ambient temperature (T-a) in a warming, urbanizing world requires spatially- and temporally-resolved T-a at high resolutions. This is challenging in a large area like France which includes highly variable topography, rural areas with few weather stations, and heterogeneous urban areas where T-a can vary at fine spatial scales. We have modeled daily T-a from 2000 to 2016 at a base resolution of 1 km(2) across continental France and at a 200 x 200 m(2) resolution over large urban areas. For each day we predict three T-a measures: minimum (T-min), mean (T-mean), and maximum (T-max). We start by using linear mixed models to calibrate daily T-a observations from weather stations with remotely sensed MODIS land surface temperature (LST) and other spatial predictors (e.g. NDVI, elevation) on a 1 km(2) grid. We fill gaps where LST is missing (e.g. due to cloud cover) with additional mixed models that capture the relationship between predicted T-a at each location and observed T-a at nearby weather stations. The resulting 1 km T-a models perform very well, with ten-fold cross-validated R-2 of 0.92, 0.97, and 0.95, mean absolute error (MAE) of 1.4 degrees C, 0.9 degrees C, and 1.4 degrees C, and root mean square error (RMSE) of 1.9 degrees C, 1.3 degrees C, and 1.8 degrees C (T-min, T-mean, and T-max, respectively) for the initial calibration stage. To increase the spatial resolution over large urban areas, we train random forest and extreme gradient boosting models to predict the residuals (R) of the 1 km T-a predictions on a 200 x 200 m(2) grid. In this stage we replace MODIS LST and NDVI with composited top-of-atmosphere brightness temperature and NDVI from the Landsat 5, 7, and 8 satellites. We use a generalized additive model to ensemble the random forest and extreme gradient boosting predictions with weights that vary spatially and by the magnitude of the predicted residual. The 200 m models also perform well, with ten-fold cross-validated R-2 of 0.79, 0.79, and 0.85, MAE of 0.4, 0.3, and 0.3, and RMSE of 0.6, 0.4, and 0.5 (R-min, R-mean, and R-max, respectively). Our model will reduce bias in epidemiological studies in France by improving T-a exposure assessment in both urban and rural areas, and our methodology demonstrates that MODIS and Landsat thermal data can be used to generate gap-free timeseries of daily minimum, maximum, and mean T-a at a 200 x 200 m(2) spatial resolution.