Comparing Deep Learning methodologies for Downscaling between meteorological models

The ability to obtain high spatial resolution meteorological data from coarse sources is a crucial skill needed to study local phenomena happening at finer scales such as severe storms or convective systems. This spatial downscaling can be achieved by reproducing the atmospheric state of a small region using numerical weather prediction (NWP) models that use low-resolution (LR) data as boundary conditions. However, running NWP models at high resolutions is computationally expensive and time consuming. A different approach is to establish statistical relationships between LR and HR data to increase the spatial resolution by interpolating intermediate points. In recent times machine learning (ML) based statistical methods have proven to be a cheap yet accurate alternative to dynamical downscaling. 

This work aims to develop a downscaling methodology from ERA5 to Weather Research and Forecasting (WRF) data based on deep learning. We study how the training dataset affects the downscaling performance and generalization capabilities of deep learning models and how it compares against traditional downscaling methods such as bilinear interpolation. Our models estimate the downscaling function for daily average 2-meter air temperature, between a LR dataset, and a high-resolution (HR) Weather Research and Forecasting (WRF) model outputs. The LR inputs come from different sources for each model. The first dataset is created by upscaling the HR WRF ground truth data to our target LR, and the second one is the ERA5 reanalysis used as boundary conditions to drive the NWP simulation. For validation purposes, we select data from regions that share similar climatology with data present in the training set that has been excluded from the training. To evaluate the performance of the model, we use Root Mean Square Error (RMSE) and metrics typically used in image super resolution problems such as Peak Signal-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM). 

With this study we have taken a first step in the ML modelling of weather downscaling and its generalization capabilities. However, further work is needed to understand the capabilities and behavior of these models when faced with challenges such as reproducing local-scale patterns, downscaling discrete variables (e.g. precipitation, hail) or the transferability of their results to similar climatic zones outside the simulation domain. Lastly, in future works we plan to study the performance of different deep learning model architectures, such as Vision Transformers or Latent Diffusion, on downscaling.