Spatial Downscaling of Daily Precipitation Using Convolutional Neural Network-Based Model in Complex Mountainous Terrain: Trentino and South Tyrol, Northern Italy

Study region: Trentino and South Tyrol region in the northern part of Italy. Study focus: This study explores the use of convolutional neural networks (CNNs) for downscaling reanalysis ERA5-Land (ERA5L) daily precipitation data from a coarse resolution of 9 km to 1 km in a mountainous terrain area in the Italian Alps. The CNN-based model was trained and evaluated using high-resolution spatial observational data, focusing on capturing fine-scale precipitation patterns and extreme rainfall events. The custom loss function is utilized to capture extreme precipitation events. The CNN model shows the lowest errors in winter, with an MAE of 2.1 mm, and the highest errors in autumn, with an MAE of 3.9 mm. Additionally, two extreme rainfall events are analyzed to assess the spatial consistency and the ability of the downscaling CNN model to capture extreme precipitation events. Challenges remain in identifying zero precipitation and capturing extreme precipitation magnitudes, particularly during events characterized by strong convection and in high-elevation areas, where biases and smoothing effects of the input data were observed. New hydrological insights for the region: The findings of the study confirm that the CNN-based downscaling approach provides enhanced high-resolution precipitation fields, offering substantial benefits for hydrological modeling in complex alpine catchments. Despite some limitations in reconstructing extreme events, the deep learning approach improves the spatial detail of precipitation inputs, contributing to more reliable flood risk assessments and water resource management in complex terrain.