A New Hybrid Approach Applying Feature Engineering to Model High Resolution Air Temperature Across Israel

Abstract

The increasing global planetary temperature, combined with rapid urbanisation, underscores the urgent need to evaluate human exposure to air temperature. This study proposes an innovative methodology to reconstruct air temperature (Ta) from 2004 to 2022 at both 1 km and 250 m resolutions across Israel, leveraging a novel feature engineering (FE) approach. Our method enables the daily prediction of three key Ta measures: minimum (T_min), average (T_mean) and maximum (T_max) air temperatures. First, we employed FE techniques to develop geographically weighted predictors and introduced a gap-filling procedure for satellite-derived land surface temperature (LST) data, producing continuous surface temperature (Ts) fields. Next, we applied Extreme Gradient Boosting (XGBoost) models to calibrate and predict Ta using spatial and spatio-temporal predictors, such as elevation, LST data from the Moderate Resolution Imaging Spectroradiometer (MODIS), the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis (ERA5), and the Spinning Enhanced Visible and InfraRed Imager (SEVIRI). Finally, we downscaled the 1 km Ta estimations to a 250 m resolution by calibrating the residuals of the 1 km model with variables at higher spatial detail such as Landsat LST data. The resulting models demonstrate excellent cross-validated performance, with root mean square error (RMSE) ranging from 0.53°C to 1.01°C. This methodology offers significant advancements in exposure assessment methods for various environmental modelling applications.