Detecting Marine Debris on Satellite Imagery

In our first case study of the most impactful use cases of satellite data, we will dive into how coastal marine debris can be detected using satellites to drive effective waste management policies and targeted cleanups in debris hotspots.

The escalating issue of ocean plastic pollution poses a grave threat to marine life, with millions of tons of plastic waste entering the oceans annually. This pollution not only disrupts ecosystems but also endangers countless marine species. The pervasive presence of plastics in our oceans has become a critical environmental concern, demanding innovative solutions for detection and cleanup.

Efforts by organizations like The Ocean Cleanup are crucial in tackling this problem, particularly in removing large concentrations of debris from areas such as polluted rivers and ocean garbage patches. However, these efforts alone are not enough to fully address the issue. Satellite data can fill the gaps by identifying debris hotspots and tracing the sources of plastic entering the oceans. This data can significantly enhance the effectiveness of cleanup operations by enabling targeted interventions and providing insights into the dynamics of marine debris flow.

Marine debris, predominantly originating from densely populated coastal cities, often lacks proper waste management systems. The waste from these urban areas is carried into the ocean via runoff and river outflows. Once in the marine environment, these plastics are transported by ocean currents, accumulating in various parts of the world’s oceans. The debris, now intermingled with natural materials like algae and driftwood, creates complex challenges for detection and removal. In this article, marine debris refers to floating objects on the ocean surface which can belong to one or more classes namely plastics, algae, sargassum, wood, and other artificial items.

Existing Work

Some key initiatives which have pushed the boundaries of marine debris detection research include:

1. Lauren Biermann and her team developed a model to analyze Sentinel-2 imagery and identify floating plastic debris with high accuracy. Her research focuses on distinguishing macroplastics from other floating materials by analyzing light wavelengths. The system has demonstrated an accuracy rate of 86% in distinguishing plastics in coastal waters, making it a promising tool for global plastic detection (Biermann et al. 2020).
2. Researchers in Greece, particularly from the Marine Remote Sensing Group of the University of the Aegean led by Kostas Topouzelis , are also contributing to the field by creating artificial plastic debris target structures. These targets are used to test the accuracy of satellite and UAV-based systems in detecting marine debris. This research helps refine detection methods and validate satellite imagery with in-situ observations. You can find more information on the website of the Plastic Litter Project.
3. Researchers who worked at the European Space Agency - ESA such as Jamila Mifdal, PhD , Raquel Carmo , and Marc Ru?wurm developed segmentation based algorithms to detect floating debris using Sentinel-2 satellite imagery. Their research focused on using these spatial features to detect and classify marine litter, particularly in high-risk accumulation zones (Mifdal et al. 2021).
4. The MARIDA (Marine Debris Archive) dataset by Kikaki et al. is a benchmark dataset specifically designed for detecting marine debris using Sentinel-2 multispectral satellite data. Published in 2022, MARIDA provides annotated data for marine debris. It serves as a valuable resource for the development and evaluation of machine learning algorithms for marine debris detection, enabling researchers to explore the spectral characteristics of floating debris and various marine phenomena.

Dataset Labeling

In 2021-2022, a gap we identified in the marine debris detection was the absence of high performing detection models on high resolution imagery - especially commercial satellite imagery.

To address this challenge, myself from the NASA-IMPACT team at the time in collaboration with Lilly Thomas from Development Seed undertook a comprehensive literature review which detect marine plastic pollution using advanced geospatial technologies. This review helped us identify validated marine debris events across different locations, including the Bay Islands in Honduras, Accra in Ghana, and Mytilene in Greece. These locations served as reference points for sourcing corresponding satellite imagery.

Here is an example of a table of expeditions with dates and tentative locations from the excellent MARIDA benchmarking dataset which helped us identify valid locations which can be used for model training.

We utilized PlanetScope data, known for its high spatial resolution of 3 meters and high temporal frequency. This imagery is particularly suited for detecting objects like marine debris due to its availability of visible and near-infrared channels. Lilly and I manually verified the presence of marine debris in the imagery and subsequently created a labeled dataset essential for training their machine-learning model. For finding the locations for our training dataset, we co-located the Planetscope imagery tiles with Sentinel-2 tiles from several research papers which had identified valid debris locations with the respective dates.

We labeled the resulting valid marine debris on Planet tiles using NASA IMPACT's Image Labeler tool that allowed us to draw bounding boxes over the marine debris locations. The final labeled dataset, consisting of 1844 polygons of identified marine debris. It was then segmented into smaller tiles (256x256 pixels). These tiles were used to train a deep learning model specifically designed for object detection.

Model Training

The model, trained on this labeled data, learned to recognize the distinct features of marine debris in oceanic environments. Once the model was sufficiently trained, it was applied to new imagery to detect and predict the presence of marine debris.

Our architecture of choice for this project is SSD Resnet 101 Feature Pyramid Network (FPN), which we've implemented with the Tensorflow Object Detection API. We employed a weighted sigmoid focal loss and transfer learning for our baseline model from a pre-trained resnet 101 checkpoint hosted on Tensorflow model zoo. I want to give a huge shoutout to Lilly Thomas for leading the model development.

The model’s performance was evaluated using metrics like Intersection over Union (IoU), precision, recall, and F1-score, achieving an F1-score of 0.85, indicating a strong capability in identifying marine debris.

This machine-learning approach is not only a breakthrough in detecting ocean plastic pollution but also offers a replicable methodology for other Earth science applications. With PlanetScope imagery’s spatial resolution, the model can detect various small features, including marine debris, buildings, and other environmental targets. This technology has the potential to significantly enhance monitoring efforts and contribute to global cleanup initiatives, paving the way for a healthier ocean ecosystem.

Here is an extensive video on my YouTube channel featuring Lilly Thomas which explains all of the details of this project with the results. Feel free to subscribe for more environmental education content.

As we continue to develop and refine these models, the hope is that they will not only advance scientific understanding but also drive meaningful action toward cleaner oceans and a more sustainable future.

You can find our open source Planetscope marine debris dataset here - https://beta.source.coop/repositories/nasa/marine-debris/description/

The technical model details and open source code is here on GitHub - https://github.com/NASA-IMPACT/marine_debris_ML

Follow the new company page for PlanetSapling to keep up to date!

Sponsorship

Do you want to share your company updates or work with our engaged audience of 8500+ geospatial data scientists, enthusiasts, founders, and investors? Please contact me at [email protected] to find out more about our newsletter sponsorship packages.

Collaborate

We would love to feature your work if it uses satellite and/or sensor data for climate adaptation, climate mitigation, or biodiversity conservation. Please contact me at [email protected] to discuss further.

Book a Call

If you want to connect with me for working on geospatial data science projects, please DM me or book a call on my Calendly link with the stated purpose.