Assessing monitoring methods for plastic contamination of the Galápagos Marine Reserve.

Despite the growing threat to coastal ecosystems posed by plastic pollution, effective monitoring tools remain limited, particularly for remote archipelagos like the Galápagos Islands. This thesis assesses different methods of monitoring plastic in coastal habitats, and investigates the application of machine-learning models for efficient detection and monitoring of coastal plastic pollution, using video and drone image data. Mangrove forests are coastal habitats that trap large quantities of plastic due to their complex root structure, simultaneously making sampling difficult. GoPro and smartphone transect recordings were annotated using the software VIAME (Video and Image Analytics for Multiple Environments). Vulnerable Ecuadorian mangroves on the Galápagos Islands and within the Guayas River Basin on mainland Ecuador were targeted, investigating how location, protected area designation, and transect method influenced the spatial distribution and composition of plastic pollution. Less plastic was recorded manually in protected mangroves than in urban mangroves, highlighting how protection can mitigate plastic accumulation. Whilst no significant difference in total plastic load was observed between both locations, significantly more rope was observed in Galápagos than in Guayaquil and significantly more packaging was observed in Guayaquil than Galápagos. Therefore, despite the similar plastic counts between sites, monitoring methods still need to identify an array of plastic types to avoid bias. This dataset was used to train the object detection model YOLOv5 (You Only Look Once version 5), to develop an automated detection method for plastic in mangroves. Whilst the precision of this method was expectedly lower due to the complex background and varied transect methods, the model successfully identified 31% of bottles. Observations from the mangrove algorithm informed the development of a model for monitoring Galápagos coastline, using a 600,000 m2 drone dataset provided by the ‘Iguanas from Above’ project. Annotation of the dataset produced approximately 330,000 images, with over 12,800 and 11,400 plastic and fauna annotations respectively across five islands, prior to augmentation. This dataset was used to train 6 YOLOv11 to detect coastal plastic, fauna, and fish aggregating devices (FADs). The algorithm achieved 68% precision for bottles, as well as 63% and 64% precision for sea lions and possible FADs depending on the model, with anecdotal estimates that the model was 1100 times faster than human annotation. Significant differences in object counts were observed between the islands, with more plastic and FADs appearing to accumulate on eastern islands than western islands. Priority monitoring sites could also be identified such as the island of Genovesa, which had a high fauna count, as well as the highest plastic count by a large margin, despite its remote and uninhabited nature. Ultimately, the models developed could identify plastic pollution efficiently with reasonable precision across coastal environments, and with refinement, the use of machine-learning methods could soon be applied to monitor plastic across large regions with complex topography.<p></p>