Fine-scale surface climate data with deep learning

High-resolution spatial surface climate information is crucial to foster advance in several areas as hydrology, ecology, agriculture and urban studies, and to bridge the gap between these interdisciplinary sciences. Coarse resolution data of prevalent gridded climate datasets and atmospheric numerical models only partly explain the observed variability at the landscape scale, especially in complex terrain where complex physical processes are not well described. We propose a deep learning based framework that estimates high-resolution fields (30 m) of near surface air temperature, humidity and wind. An Artifical Neural Network model is used to predict the subgrid spatial variability dependent on the surrounding landscape characteristics and large scale weather estimates. Model inputs are coarse scale climate predictors from ERA5 reanalysis outputs, and non-temporal multiscale high resolution spatial features as reflectance-derived vegetation indices (using Landsat-8) and Digital Elevation Model data. The model was trained and cross-validated with a multi-year dataset of hourly weather station data distributed over contiguous Brazil. We show a pilot investigation in a mountainous watershed of South-East Brazil where a dense network of weather stations was helpful to discuss model performance and limitations.