Spatial Downscaling and Forecasting of Atmospheric Variables over Complex Terrain Using Machine Learning

Accurate fine-resolution weather and climate data is crucial for different applications like agriculture, energy, and insurance. This is particularly necessary in complex terrain where the variability of atmospheric variables is strongly influenced by topography. Traditional Global Circulation Models (GCMs), however, typically have a coarse spatial resolution that is unable to capture local scale variations, leading to limitations in regional climate applications. This thesis addresses these limitations by downscaling atmospheric variables over mountainous regions, utilizing machine learning (ML). The objectives of the study focused on downscaling ERA5-Land reanalysis daily mean temperature and precipitation data from a coarse resolution of 9 km to a finer 1 km resolution over the Non and Adige valleys in the Italian Alps. Our other objective includes spatial downscaling of the seasonal forecast of daily minimum temperature from 12 km to 250 m with one month lead. This study utilized ML models, Artificial Neural Networks (ANN), Random Forests (RF), and Convolutional Neural Networks (CNN), CNN-based encoder decoder to enhance the spatial resolution of atmospheric variables. The results of downscaling models were intercompared and evaluated for their effectiveness to determine their ability to capture complex spatial patterns. Our results suggest that CNN outperformed the other models, capturing both broad and fine-scale spatial variability across all seasons. These objective underscores the potential of ML methods to substantially improve predictions at high resolution, which is crucial for accurate representation in complex terrain for different applications.