This thesis investigates the transferability of the Swin-BathyUNet deep learning architecture from coastal marine environments to a small alpine lake, using UAV-derived data as the primary input. The study is framed within the ACLIMO project and focuses on Lake Vej del Bouc in the Maritime Alps of Italy. The processing pipeline encompasses UAV photogrammetric reconstruction, refraction correction analysis, manual point cloud cleaning, and deep learning-based depth prediction from co-registered RGB orthoimagery and SfM-MVS depth rasters. Three successive dense point cloud reconstructions were performed in Agisoft Metashape to address a persistent phantom surface artifact, and a manual cross-sectional cleaning procedure in CloudCompare was required to produce reliable training labels.