Detection of plastic waste in various water bodies using remote sensing data.

This thesis presents a processing pipeline for detecting floating plastic debris in aquatic environments using optical satellite imagery and machine learning. The pipeline combines two complementary models: an instance segmentation model applied to optical features (color, shape, texture) to detect debris-like objects, and a pixel-wise classification model using spectral indices to distinguish plastics from natural materials such as algae, pumice, and timber. Outputs include trained models, evaluation metrics, and a curated dataset for future monitoring and research. Growing interest in remote sensing for marine litter has led to promising AI-based methods. A key reference is [Biermann 2020], which introduced the Floating Debris Index (FDI) and demonstrated detection of plastics in Sentinel-2 imagery using spectral classification; this study forms the basis of the present work. Other studies applied XGBoost [Duarte 2023] and deep learning segmentation [Rußwurm 2023], emphasizing both model design and the critical role of data quality. The study relies on Sentinel-2 MSI imagery, particularly its 13 spectral bands useful for debris detection. Data comes in Level-1C and Level-2A formats, with atmospheric correction applied via Sen2Cor (default) or ACOLITE DSF (compared during preprocessing). Spectral indices like NDVI and FDI form input feature space for the classification task. For modeling, the Ultralytics YOLOv11 instance segmentation architecture was used for object detection, and a Support Vector Machine (SVM) was applied for spectral classification. Datasets were created using Roboflow (segmentation), and QGIS (mostly for classification). Study sites included Visegrad Dam, Durban, Omoa, Accra, Japan, Tonga, and Barbados, chosen from scientific/media reports. Atmospheric correction methods were evaluated, with Sen2Cor selected for better plastic separation. Classification samples were labeled in QGIS, cleaned with Isolation Forest, and balanced. The SVM was trained in an overfitted fashion to reproduce class-specific signatures rather than generalize broadly. Segmentation training data was created from RGB tiles annotated in Roboflow. Three YOLOv11 model sizes (nano, small, medium) were tested in a cascade setup with hyperparameters tuned per configuration. Classification performance was evaluated on both test splits and independent sites, with and without segmentation-based filtering. The segmentation model (YOLOv11m) achieved moderate performance (validation mAP50–95: 0.214 for boxes, 0.074 for masks), with generalization limited by data scarcity and imagery complexity. The SVM classifier performed strongly on test data (balanced accuracy: 0.972, macro-F1: 0.973), though likely overfitted. Segmentation masks significantly improved classification precision in cluttered scenes; e.g., F1 increased from 0.281 to 0.827 (Visegrad) and from 0.043 to 0.529 (Omoa). Precision gains were clear in complex cases, while simpler scenes showed minor improvement. Overall, the two-stage approach proved effective but highlighted the need for more training data and robustness to edge cases.