Cloud and shadow regions in satellite images pose significant challenges for accurate image interpretation and analysis. This thesis investigates a deep learning approach using the U-Net model for synthetic generation and segmentation of cloud and shadow regions. A synthetic dataset was created to train and validate the model, reducing dependency on limited annotated data. The performance was evaluated using IoU and accuracy metrics, showing that the proposed method achieves reliable segmentation results. The study demonstrates the effectiveness of combining synthetic data generation with U-Net for robust cloud and shadow detection in remote sensing applications.