High-resolution global earth system modeling (ESM) data is computationally expensive to produce, but crucial for assessing climate change effects at local and regional scales. Here, we adapt a new deep learning technique, called fast super-resolution convolutional neural network, termed FSRCNN-ESM, to inexpensively remap climate data from low resolution to high-resolution grids.
FSRCNN-ESM’s reconstruction of high-resolution spatial patterns demonstrates better skill at statistical downscaling of surface temperature, radiative fluxes, and precipitation as compared to other prevalent machine learning methods, while also being computationally less expensive to train.
We present a first application of a fast super-resolution convolutional neural network (FSRCNN) based approach for downscaling earth system model (ESM) simulations. Unlike other super-resolution approaches, FSRCNN uses the same input feature dimensions as the low-resolution input. This allows it to have smaller convolution layers, avoiding over-smoothing, and reduced computational costs. We adapt the FSRCNN to improve reconstruction on ESM data, which includes an additional patch extraction step with non-linear mapping to improve accuracy. We term this adaptation the FSRCNN-ESM. We use high-resolution (25 km) monthly averaged model output of five surface variables over a part of North America from the US Department of Energy's Energy Exascale Earth System Model's control simulation. These high-resolution and corresponding coarsened low-resolution (100 km) pairs of images are used to train the FSRCNN-ESM and evaluate its use as a downscaling approach. We find that FSRCNN-ESM outperforms FSRCNN and other super-resolution methods in reconstructing high-resolution images producing finer spatial scale features with better accuracy for surface temperature, surface radiative fluxes, and precipitation.