Downscaling Forcing for Lake Garda using Neural Networks: Present & Future

Abstract

Deep mixing events (DMEs) are an important aspect of the ecosystem in Lake Garda. To predict when DMEs will happen in the future, especially when considering climate change, a hydrodynamical model is used (such as Delft3D). The input for such models has to be atmospheric forcing with a high resolution. The northern part of Lake Garda lies between high mountain 5 ranges, where high-resolution observational atmospheric data is inaccessible due to sparse observational posts. Therefore, data from the models of the National Centers for Environmental Prediction’s Global Forecasting System (NCEP GFS) for presentday, and EC Earth (future climate scenarios model) are needed. These datasets, however, have too low of a resolution compared to the lake size. NCEP has a resolution of 0.25◦ (17 km in width and 27 km in height), and EC Earth has a 25 km resolution. Both datasets need to be dynamically downscaled, normally with a hydrodynamical model such as the Weather Research and 10 Forecasting (WRF) to 3 km or 4 km resolution. This, however, takes a lot of time; here, neural networks could prove to be very useful. After a long time training the network (just under two weeks), downscaling new data takes significantly less time (a couple of minutes). For the Lake Garda region, two different neural networks (Generative Adversarial Networks or GANs) will be trained using the work of van Rijk (2022), who based his work on Stengel et al. (2020). The first neural network will use training data from 2017 until May 2018, while the second network will use training data until 2020. The test data is from 2021, 15 and all data consists of the following variables: 2 m above surface temperature (T2m) and 10 m above surface wind velocity (U10m and V10m). With the second neural network, a future test dataset will be downscaled from EC Earth 25 km resolution to a 3 km resolution and compared to WRF 4 km. The validation metrics of the generated super-resolution, when compared to the WRF data, are the root mean squared error (RMSE) and the structural similarity index measure (SSIM). Improvements to both SSIM and RMSE for all variables are observed with more training, though they do not improve at the same rate. For 20 future data, both are worse than the present-day data.