This thesis investigates the use of deep learning techniques for detecting marine plastic pollution using Sentinel-2 satellite imagery. It presents a comparative analysis of U-Net and U-Net++ architectures on the MARIDA dataset under realistic operational constraints, such as class imbalance, spectral confusion, and minimal pre-processing. The models were implemented using a consistent training framework without augmentation in the baseline experiments and evaluated with standard semantic segmentation metrics including Intersection over Union (IoU), precision, recall, and F1-score. Additionally, the generalization capacity of the best-performing model, U-Net++, was tested on Sentinel-2 imagery from the Motagua River, Guatemala one of the most plastic-polluted river outflow globally. Results demonstrate that U-Net++ significantly outperforms U-Net, particularly in detecting sparse and fragmented plastic patches. This study underscores the potential of advanced segmentation models for scalable, real-time marine plastic monitoring and highlights the importance of evaluating AI solutions under noisy, unfiltered remote sensing conditions.