Modeling local precipitation in Southern South America using Deep Learning: A sensitivity analysis on the choice of input features and loss function in the climate change signal

Southern South America (SSA) covers the extratropical part of South America (20–60°S) and presents a wide variety of climates. To the west of the Andes mountain range, annual precipitation increases southward from very dry conditions along the Atacama Desert in northern Chile to more than 3000 mm in the south. Conversely, east of the Andes, it increases from the Argentinian Patagonia in the south towards southeastern South America (southern Brazil, northeastern Argentina and Uruguay) where severe thunderstorm environments are typical. Global climate models project that the observed negative (positive) trends in precipitation over the subtropical central Andes (Southeastern South America) are expected to be more intense causing concern about water availability, ecosystems and socio-economic activities. However, the regional-to-local information that can be obtained by downscaling over GCMs outputs and needed for adaptation and mitigation policies, is still scarce over SSA. Unlike other parts of the world, limited studies analyzing the statistical downscaling (ESD) potential to simulate daily precipitation are available over the region and deep learning-based models have not been tested for downscaling daily precipitation over the region up to now.

In this context, this work presents a comprehensive assessment of Convolutional Neural Networks (CNNs) to downscale daily precipitation at a continental-scale, building on the validation framework of the European project VALUE. To this end, we conduct a sensitivity analysis to the domain size as well as to the selection of the loss function on the modeling of precipitation in both present and future climates. Overall, the CNNs show skilful performance in modeling daily precipitation characteristics, including the extremes, over the different climatic regions of SSA. Nevertheless, we find the selection of the loss function to be a source of uncertainty over the arid regions of northern Chile and northwestern Argentina for both present and future climates by projecting different climate change signals. Regarding the domain size, the CNNs show to be effective in selecting informative predictors and their area of influence demonstrating their self-learning skill and their efficiency to be applied on a continental scale. These results encourage the construction of ensembles of deep learning models based on different loss functions in SSA to account for this type of uncertainty in the modeling of precipitation, especially in future climates.