The use of a U-Net neural network for high-resolutionc urban climate modelling

Urban areas cover 3 to 5% of total land area, however they contain nearly 2/3 of global population. It has been shown that cities also experience increased effects of climate change. Therefore for a sustainable future it is vital to be able to model not only the large-scale climate change, but also the small-scale changes of weather patterns on an intra-urban scale.

Modern methods of climate modelling do not provide sufficient resolutions to reflect on the 100 meter scale of the city processes. Downscaling methods are used to increase the spatial resolution of climate data. Dynamical downscaling is the use of mesoscale weather models. Statistical downscaling includes machine learning and deep learning methods.

This work employs both methods to create high-resolution air temperature, wind speed and thermal comfort fields. Mesoscale models are used to model the target vartiables with high resolution. Then the data is used to train a U-Net neural network. The network takes ERA5 low-resolution fields and high resolution fields of land surface data. 

The network has shown a decrease in errors in comparison wint ERA5 which does not reproduce urban microclimate. For key urban stations the mean error was reduced by several times. Moreover, there is a significant gain of speed. Computing 1 year with a weather model takes 17 days, while it takes the NN 5 minutes to do the same. This creates opportunities for high-res climate modelling