Testing the use of deep learning techniques for emulating regional reanalysis

Reanalysis datasets serve as essential components for contemporary climate monitoring, integrating historical weather observations with predictive models to create extensive climate data records for the last decades. The fifth generation ECMWF atmospheric global climate reanalysis (ERA5) dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF) represents the latest update, providing a broad temporal scope and improved spatial granularity. However, its resolution may fall short for detailed local-scale analysis required in critical sectors such as agriculture, energy, and disaster response, among others. Even though more detailed regional information for Europe like the Copernicus European Regional ReAnalysis (CERRA) do exist, its high computational costs and the lack of very near real-time data updates create limitations to conducting analyses close to real time.

To solve some of these limitations, a deep learning model has been developed to mirror CERRA's 2m temperature field utilising ERA5 as input. This approach aims to replicate the details of CERRA, ensuring rapid and efficient emulation without surpassing its original quality, i.e. treating CERRA as the ground truth. Central to this model is the Swin2SRModel component (Swin v2), which has effectively demonstrated the ability to downscale the resolution of inputs by a factor of 8. This capability aligns well with the intended task of downscaling the grid from 0.25º (ERA5) to 0.05º (CERRA). To achieve this, a Convolutional Neural Network (CNN) pre-processes the data, reshaping it to the necessary feature map size. The model training is focused on the specific region of interest of the Iberian Peninsula, instead of the entire European CERRA domain. The training, lasting 100 epochs, takes approximately 3.6 days using small batch processing. It employs the Adam optimizer, starting with a learning rate of 0.0001 that decreases following a cosine curve, integrating a warm-up phase to mitigate training instability. It utilises 32 years of data, spanning from 1985 to 2016, and its performance is validated against the independent dataset of 2017 to 2021.

A comprehensive post-training evaluation of the model shows a marked improvement – 35% reduction in Mean Absolute Error (MAE) and a nearly 30% enhancement in Root Mean Square Error (RMSE) – compared to the bicubic interpolation method. This leap in accuracy is especially notable in complex landscapes. Validation on specific locations, such as the Aneto mountain, have demonstrated a dramatic refinement in the mean error, dropping from -6.3°C to 0.06°C – 99% improvement. Similar improvements have been observed in Cantabrian Mountains such as Peña Vieja (94%) and Peña Labra (88%), illustrating the model's superior performance in areas where previous errors were substantial, highlighting its ability in areas that most require it.

In conclusion, the project shows promising results in enhancing reanalysis data with AI, demonstrating potential in both computational efficiency and near real-time application. While initial results are encouraging, indicating reduced errors compared to the bicubic interpolation, comprehensive validation against CERRA using independent observations and expansion to broader domains and variables remain crucial for confirming the method's effectiveness and reliability.