This study presents a machine learning-based framework for detecting and classifying floating pollutants in surface water using object detection techniques. Five pollutant categories are targeted: plastic waste, oil spills, acid contamination, eutrophication, and dead animals. The dataset comprises 1440 annotated images generated through GAN-based augmentation and real-world sampling, with an 80-10-10 split for training, validation, and testing. Two models were evaluated: a Convolutional Neural Network (CNN) classifier and a YOLOv8m object detector. The CNN achieved a classification accuracy of 93.18%. YOLOv8m yielded a mean Average Precision (mAP@0.5) of 77.17%, making it suitable for real-time detection scenarios. The study highlights the trade-off between CNN’s higher precision in classification tasks and YOLO’s real-time efficiency. These results demonstrate the potential of deep learning models for scalable, cost-effective environmental monitoring solutions and open pathways for future integration into automated water quality surveillance systems.