A multi-resolution air temperature model for France from MODIS and Landsat thermal data

Abstract

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 km² across continental France and at a 200 × 200 m² 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 km² 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 R² of 0.92, 0.97, and 0.95, mean absolute error (MAE) of 1.4 °C, 0.9 °C, and 1.4 °C, and root mean square error (RMSE) of 1.9 °C, 1.3 °C, and 1.8 °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 × 200 m² 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 R² 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 × 200 m² spatial resolution.

Keywords: Air temperature; Exposure error; Land surface temperature; Landsat; MODIS.

Fig. 1.

Climatic regions of France according to Joly et al. (2010) and METEO-FRANCE stations used in the current study

Fig. 2.

Predicted 1 km T_a from the stage 2 model on selected days: Feb 18, 2003 (top row) and Aug 10, 2012 (bottom row).

Fig. 3.

Predicted 1 km T_min from the stage 2 model alone (top row) and with predicted 200 m T_min from the stage 4 model overlaid (bottom row) on Feb 18, 2003 over the Paris, Toulouse, and Nancy metropolitan areas.