Deep learning for climate downscaling: generating high-resolution gridded temperature projections over India from low-resolution CMIP6 data

Global climate change leads to distinct regional impacts. Future climate projections are available through global model outputs, but they lack the spatial detail needed to inform local decision-making for adaptation. Regional- and local-scale climate information can be generated by ‘downscaling’ these global model outputs through regional climate model simulations, which require enormous supercomputing resources, thereby limiting the availability of local-scale projections, especially in underdeveloped regions. Here, we build and test four different computationally efficient machine learning models as an alternative to regional climate models for downscaling daily temperature data over India. We utilise publicly available CMIP6 coarse (1.87° × 1.87°) resolution climate data and relate them to the finer (0.1° × 0.1°) resolution ERA5-Land reanalyses 2 m daily temperatures as a target to train our models over the 2018–2022 period. We progressively improve our model design by utilising spatial learning (through a convolutional neural network) and then including temporal learning (through ConvLSTM) and optimising the target region for downscaling. We systematically evaluate all four models for a temporally near year 2017 and a temporally distant year 2007 and select the best model variant to project daily temperatures over India for the medium-term future, 2030. We present seasonal maps, frequency distributions, as well as city-specific time-series of daily temperatures for the downscaled data vis-à-vis the original global model data. We find a slight overall increase in the annual averaged downscaled temperatures for 2030 (0.3°C) and an even larger increase in the most frequent (i.e., modal) temperature (1°C), but also a decrease over certain regions, particularly in the Indo-Gangetic plain, in summer and monsoon as compared to the coarse-scale temperatures in the global model data for 2030. Our deep learning model compares favourably with reference data from the ERA5-Land reanalyses and runs much faster, with much fewer computational resources than a typical physics-based regional climate model and therefore opens up the possibility of democratising the field of climate downscaling.