Climate and weather physical processes that cannot by resolved in current climate models due to the coarse grid spatial resolution are approximated using parameterization schemes. Such parameterizations are used as simplifications to represent those unresolved processes (McFarlane, 2011). Current parameterizations are designed by humans based on physical understanding, observations and statistical methods, and involve a large portion of the total computational time when running physics-driven dynamical models. An alternative approach to creating such parameterizations is by deriving them from large data volumes and using Machine Learning (ML) techniques (Krasnopolsky et al. 2012, Berner 2017, Brenowitz et al. 2018, Dueben and Bauer 2018, Bolton and Zana 2019, Rozas et al. 2019). Precipitation is a meteorological variable heavily affected by parameterizations in weather and climate models.
Recently, Rozas et al. (2019) implemented encoder-decoder deep Convolutional Neural Networks (CNNs) from the field of Computer Vision (CV) to learn the spatial information contained in the geopotential field to infer total precipitation with a high degree of accuracy. This study was carried out using ERA-interim reanalysis data. Such data are produced with a dynamical model and a continuous assimilation of observations resulting in the best possible picture of the weather state at a given time (Dueben and Bauer 2018). Unfortunately, training and validating an ML model using exclusively reanalysis data allows to only produce an emulator that can eventually predict the (imperfect) precipitation simulated by the atmospheric model used for the reanalysis.
Therefore, we extend previous studies by learning a neural network-based transform function between predictors, such as the geopotential height from ERA5 reanalysis and other observed boundary conditions like sea surface temperature, sea ice extent and soil moisture, and precipitation based on real observations (E-OBS). At the same time, we employ newer and more performant CV architectures such as DeepLabv3+ (Chen et al. 2018) and consider the exploitation of temporal dynamics by using spatio-temporal networks (Shi et al. 2017, Wang et al. 2019).
