Deep-learning based climate downscaling using the super-resolution method: a case study over the western US

Abstract. Demand for high-resolution climate information is growing rapidly to fulfill the needs of both scientists and stakeholders. However, deriving high-quality fine-resolution information is still challenging due to either the complexity of a dynamical climate model or the uncertainty of an empirical statistical model. In this work, a new downscaling framework is developed using the deep-learning based super-resolution method to generate very high-resolution output from coarse-resolution input. The modeling framework has been trained, tested, and validated for generating high-resolution (here, 4 km) climate data focusing on temperature and precipitation at daily scale from the year 1981 to 2010. This newly designed downscaling framework is composed of multiple convolutional layers involving batch normalization, rectification-linear unit, and skip connection strategies, with different loss functions explored. The overall logic for this modeling framework is to learn optimal parameters from the training data for later-on prediction applications. This new method and framework is found to largely reduce the time and computation cost (~ 23 milliseconds for one-day inference) for climate downscaling compared to current downscaling strategies. The strength and limitation of this deep-learning based downscaling have been investigated and evaluated using both fine-scale gridded observations and dynamical downscaling data from regional climate models. The performance of this deep-learning framework is found to be competitive in either generating the spatial details or maintaining the temporal evolutions at a very fine grid-scale. It is promising that this deep-learning based downscaling method can be a powerful and effective way to retrieve fine-scale climate information from other coarse-resolution climate data. When seeking an efficient and affordable way for intensive climate downscaling, an optimized convolution neural network framework like the one explored here could be an alternative option and applied to a broad relevant application.