Deep Learning for Daily Precipitation and Temperature Downscaling

Abstract:

Global climate models (GCMs) are important tools for assessing historical and future climate change and variability. However, their resolutions are too coarse to be implemented at local scales. Statistical downscaling of GCMs outputs is considered as a means to bridge the gap between large-scale climate modeling and local-scale planning and management. Current statistical downscaling methods rarely exploited full spatio-temporal dependencies of large- and local-scale climate data and therefore did not well capture local-, small-scale features such as extreme events during both current and future periods. These limitations can be potentially addressed through deep learning, an emerging field that has achieved notable progress in modeling data with spatial context in computer vision field. In this project, we construct and evaluate a novel method to downscale precipitation and temperature based on a deep learning architecture of generative neural networks. We design a synthetic experiment to downscale coarse-resolution daily precipitation and temperature data from 25-km, 50-km and 100-km resolutions into 4-km with varying downscaling ratios in the Mobile Bay watershed. Different metrics are used to evaluate the downscaled average and extreme precipitation and temperature. The result shows that deep learning downscaled precipitation and temperature well capture the spatial-temporal variability and extremes of high-resolution observations compared to the local constructed analog downscaling method, one of the best statistical downscaling methods in the current literature.