Downscaling using Deep Convolutional Autoencoders, a case study for South East Asia

Abstract. Inspired by recent advancements in the field of computer vision, specifically models for generating higher-resolution images from low-resolution images, we investigate the utility of a deep convolutional autoencoder for downscaling and bias correcting climate projections for South East Asia (SEA). Downscaled projections of 2 m surface temperature are generated, using autoencoders trained with data from the Coupled Model Intercomparison Project Phase 5 (CMIP5) and data from the fifth generation ECMWF atmospheric reanalysis (ERA5) project. Using CMIP5 projections as an input, three sets of downscaled data are generated using three methods of autoencoder training, which allow us to determine how autoencoder downscaling and bias correction modify temperature values. Where possible, the downscaled outputs are compared against the Southeast Asia Regional Climate Downscaling/Coordinated Regional Climate Downscaling Experiment–Southeast Asia (SEACLID/CORDEX–SEA) project and outputs from available CMIP6 experiments, to evaluate performance. The autoencoders are found to excel at the rapid generation of highly spatially-resolved climate projections for surface temperature. Realistic spatial features due to coastal and topographic variation are generated by the autoencoder, which are not present in the CMIP5 projections. Additionally, the autoencoders are capable of generating forecast data with regional temperature profiles exceeding that of those appearing in the training set (out-of-sample extrapolation). Seasonal temperature cycles are retained after downscaling throughout the region, despite the absence of temporal information provided to the model. However, autoencoders trained to carry out bias correction display a tendency to smooth daily average temperatures and reduce daily highs and lows beyond that which can be expected to be realistic. Without bias correction, downscaled outputs have a reduced improvement in spatial resolution but the daily temperature profiles of the CMIP5 input forecasts are maintained. Autoencoders rely on the presence of structural features in the datasets to carry out downscaling, and so performance over the oceans is reduced as strong temperature gradients are absent. For this reason, ocean warming is not well represented, an artefact which is not immediately clear in the downscaled outputs. This study demonstrates the importance of rigorous analysis of 'black-box' methods, which can generate non-obvious artefacts that could potentially create misleading results. Despite these limitations, Autoencoders are clearly capable of generating much needed high-resolution climate projections, and strategies to improve upon shortcomings are numerous and well established.