Predictions and reanalysis of meteorological data through AI models

In recent years traditional numerical methods for the re-analysis and forecasting of weather predictors have been increasingly challenged by deep learning methods. This thesis explores two applications of deep learning in this context. The first focuses on the reconstruction of Sea Surface Temperature (SST), a critical variable in climate systems. SST data often contains gaps due to cloud cover, and the challenge lies in accurately reconstructing this data. Several experiments were conducted to address the problem of SST reconstruction, utilizing two distinct deep learning architectures: U-Net and Vision Transformer (ViT). These models were tested under various configurations to determine the most effective approach for reconstructing SST data. Key factors such as image resolution and input parameters were explored in detail to evaluate their impact on model performance. Additionally, a comparison was made with the state-of-the-art model DINCAE, allowing for an evaluation of how the newly developed models performed in relation to established benchmarks in SST reconstruction. The second application involves precipitation nowcasting, specifically analyzing which predictors significantly impact short-term precipitation forecasts. By using an existing model, the study evaluates how integrating additional parameters into the initial dataset influences prediction accuracy.