Fronts – the boundaries between water masses, in coastal and oceanic regions - are hotspots for rich and diverse marine life, and are also where floating marine debris tends to accumulate.
The main goal of this research is to develop a prototype risk index for the accumulation of marine plastic debris at fronts. In addition to mapping the risk areas for debris accumulation, we investigate their connectivity to the pathways into the ocean, through numerical dispersion models with high spatial resolution. In doing so, we aim to provide a tool to local and regional policymakers to identify areas where intervention would be more effective. As a case study, we are working in collaboration with local stakeholders in Da Nang, Vietnam.
The study presented here shows the analysis of satellite imagery for 2018 around the Da Nang area. First, we have used very high-resolution imagery from Worldview 3 to validate Sentinel-2 detections of accumulation of marine debris. To do so, we have compiled a dataset from the literature, from specific imagery in the area and from reports on social media. We have then selected Worldview 3 imagery and used it to compare with Sentinel-2 detection of accumulations from different categories. This comparison has been done using different distance metrics. Then we have compared the location of verified accumulations with different detection methods for fronts, to identify the matching spatial resolution scales of observation. We present preliminary results on detection of fronts from altimetry, from AVHRR sea surface temperature and from Sentinel-3 and Sentinel-2 optical data. Dynamics of fronts have been identified aligned with the monsoon variations in the area. In addition, a very high spatial resolution hydrodynamic model has also been implemented and preliminary results will be shown on how the particle tracking matches sources/pathways to fronts. Results will be discussed in terms of challenges and opportunities ahead.

In recent years, there has been an increasing interest in exploring capabilities of remote sensing technologies to monitor and detect marine litter, and in particular, plastic litter floating in our water bodies. Remote sensing is now considered as the solely tool able to provide regular information at global scale to inform about this problem (Maximenko et al, 2019), which is essential to constrain the Lagrangian transport models put in place to understand its dynamics (Maximenko et al, 2012; van Sebille, 2015). In 2020, the European Space Agency (ESA) launched the Discovery Campaign on Remote Sensing of Plastic Marine Litter, funding a total of 26 initiatives across a wide spectrum of areas. The “Windrows As Proxies” project (WASP), included in the initiative, aimed to prototype an operational processor capable of identifying filaments of floating marine debris using Copernicus Sentinel-2/MSI (S-2) images, with a high probability of containing marine litter, based in the existing oceanographic knowledge of their role as accumulating agents of marine debris and plastic litter (Cozar et al, 2021; Ruiz et al, 2020).

The work from Hu (2021) stressed that a direct plastic detection using S-2 data could fall out of the possible or being specially challenging, due to spectral mixing, lack of specific bands, and challenges in both SNR and expected spatial coverage. Moreover, Dr. Hu (personal communication) also raises questions about the potential separability between marine litter and other substances, like sea snot, with similar spectral signature in the range 480-900 nm. Contrary to other initiatives (Biermann et al, 2020; Kikaki et al., 2020), where several attempts have been made to directly classify plastic within the images at pixel level, WASP embraced the problem from a different perspective. Assuming that litter rarely appears as a separate end member of the spectral reflectance measured at each pixel of S-2 data, and considering that S-2 lacks of plastic litter-specific bands (Hu, 2021), the objective is to detect filaments of floating debris, which often contain significantly lager quantities of plastic litter than their surroundings. This detection by proxy also enables the performance of both spectral and contextual (object-based) detection, which leads towards a less ambiguous classification process, with no need of external sources of data.

In a first step, WASP developed a specific spectral index that explores the use of NIR and SWIR bands as main source of information for the classification. This is also a different solution than published, where most of referred algorithms employ techniques closer to Ocean Colour using VIS bands as main source, following the lead of exiting spectral indices like NDVI and FAI (Biermann et al, 2020; Kikaki et al., 2020). The advantage of using NIR/SWIR is that, such spectral region, are containing information relevant for the detection of plastic (Garaba et al, 2020) and floating organic matter, reason why no external data is needed to support the detection.

In this work, we explore the use of such bands in Sentinel-2 for the detection of these filaments of floating marine debris, as well as create a better understanding of the general challenges that detection of plastic litter from space has. Part of the work performed includes the generation of an ad hoc procedure for cloud detection and filtering, and the use of a deterministic contextual classifier for the detection of filaments within the image, using a multiscale approach.

The results of WASP consist on snippets of S-2 analysed data containing the detected filaments. Such snippets are manually supervised by operators, in order to discard known false positives. Validation of the detection method has been carried out with the support of information over artificial targets deployed at sea in Lesbos Island (Topouzelis et al, 2020). The results of the validation show the capability the method has to detect both plastic and floating debris with high organic load.

The work presented here has helped also to understand the roadmap and needs for improvement of the techniques, as well as to advance in the detection principles of plastic litter with its challenges, a cornerstone information for the definition of a specific future EO mission devoted to plastic litter monitoring.

Biermann, L., Clewley, D., Martinez-Vicente, V., & Topouzelis, K. “Finding plastic patches in coastal waters using optical satellite data”. Scientific reports, 10(1), p1-10, 2020.

Cózar, A., Aliani, S., Basurko, O. C., Arias, M., Isobe, A., Topouzelis, K., ... & Morales-Caselles, C. (2021). Marine litter windrows: a strategic target to understand and manage the ocean plastic pollution. Frontiers in Marine Science, 8, 98.

Garaba, S. P., Arias, M., Corradi, P., Harmel, T., de Vries, R., & Lebreton, L. “Concentration, anisotropic and apparent colour effects on optical reflectance properties of virgin and ocean-harvested plastics”. Journal of Hazardous Materials, 124290, 2020.

Hu, C. (2021). Remote detection of marine debris using satellite observations in the visible and near infrared spectral range: Challenges and potentials. Remote Sensing of Environment, 259, 112414.

Kikaki, A., Karantzalos, K., Power, C. A., & Raitsos, D. E. “Remotely Sensing the Source and Transport of Marine Plastic Debris in Bay Islands of Honduras (Caribbean Sea)”. Remote Sensing, 12(11), 1727, 2020.

Topouzelis, K., Papageorgiou, D., Karagaitanakis, A., Papakonstantinou, A., & Arias, M. “Remote Sensing of Sea Surface Artificial Floating Plastic Targets with Sentinel-2 and Unmanned Aerial Systems (Plastic Litter Project 2019)”. Remote Sensing, 12(12), p2013, 2020.
Maximenko, N., Hafner, J., & Niiler, P. “Pathways of marine debris derived from trajectories of Lagrangian drifters”. Marine pollution bulletin, 65(1-3), p51-62, 2012.

Maximenko, N., Corradi, P., Law, K. L., Van Sebille, E., Garaba, S. P., Lampitt, R. S., ... & Wilcox, C. “Toward the integrated marine debris observing system”. Frontiers in marine science, 6, 447, 2019.

Ruiz, I., Basurko, O. C., Rubio, A., Delpey, M., Granado, I., Declerck, A., ... & Cózar-Cabañas, A. “Litter windrows in the south-east coast of the Bay of Biscay: an ocean process enabling effective active fishing for litter”: Frontiers in marine science 2020.

Van Sebille, E., Wilcox, C., Lebreton, L., Maximenko, N., Hardesty, B. D., Van Franeker, J. A., ... & Law, K. L. “A global inventory of small floating plastic debris”. Environmental Research Letters, 10(12), 124006, 2015.

The remote sensing of marine litter is becoming increasingly important. Disasters such as floods cause large amounts of natural and man-made waste to enter coastal waters. Since such large-scale floating litter is an obstacle to safe operation of ships and recreation activities, it is necessary to issue an alert and remove it as quickly as possible. In addition, since man-made plastic litter adversely affects not only the aesthetics but also the marine ecosystem, interest in reducing plastic waste is higher than ever. Some of the plastic waste entered into the ocean and moves along the ocean current over a long period of time, threatening the life of marine animals. Numerous clean-up operations are being undertaken. Mapping areas of plastic litter accumulation in the ocean is certainly useful for efficient operations.
However, it is still challenging to reliably detect marine litter from space. One of the key technique to do is to detect changes in reflectance to identify small patches floating on the ocean surface. The spatial anomaly in reflectance was calculated by subtracting the spatially varying reflectance of the surrounding background water from the satellite-measured reflectance, and it was used to detect reflectance changes due to presence of drifting litter patches.
By applying this technique to high-resolution images of Worldview-3, we investigated the possibility of detecting plastic litter in Great Pacific Garbage Patch. Specifically, anomaly spectra were evaluated to detect plastic litter using the Ocean cleanup System 001 Wilson as a known target. While floating litter in the open ocean is found isolated, it is often observed accumulated along the water fronts. We applied the same technique to different-scale events in coastal waters using images from various satellite sensors including PlanetScope, MSI and GOCI. In addition to the change of reflectance due to the presence of litter, the reflectance spectrum of the target litter itself is very important information to identify the kind of litter, which will be discussed in this presentation.

Introduction

Ocean pollution is growing into a serious threat not only to the ecosystem and marine species but also to human life. In fact, significant quantities of anthropogenic trash, especially plastic, keeps getting discarded each year in the ocean mostly via rivers during extreme weather events. The pollutants could stay for long years in the ocean and cause long-term harm. For instance, plastic's degradation could take more than 400 years to happen, which is why it is very important to act rapidly to tackle the issue of marine pollution. Initiatives across the world such as the UN Sustainable Development Goal 14 and the EU Marine Strategy Framework Directive's descriptor 10, encourage improving the ocean's health.

A remarkable increase in methods to detect plastic and floating objects in general can be witnessed in the last few years and months (Topouzelis et al.,2019; Biermann et al., 2020; Mifdal et al., 2021). Spotting plastic can be very hard and expensive. In fact, most of the current monitoring methods rely on drones and/or coastal cameras which is not always affordable and convenient. Thus, an increasing amount of research is shifting the attention on the use of satellite data for floating targets monitoring. Currently, the most convenient satellite data is Sentinel-2 thanks to its global coverage and the public availability of its data, thus, it offers the possibility to monitor large-scale areas across the globe. Sentinel-2 provides images with thirteen or twelve bands, four bands have the spatial resolution of 10 m and the remaining ones have a 20 m and 60 m spatial-resolution. In this setting, only big patches of objects can be detected using Sentinel-2 data. Thanks to different natural processes such as wind, waves etc, floating debris are agglomerated and form patches that could be spotted on Sentinel-2 images. Thus, our goal is to focus on the detection of agglomerated floating targets in water bodies and provide a map of the detected objects which could be useful for other downstream tasks such as the detection of plastic, ocean clean-up operations, automating the monitoring of ocean pollutants etc. Recently, Sentinel-2 data along with deep learning methods proved effective for the detection of floating targets in water bodies (Mifdal et al., 2021). The authors hand-labeled floating patches on Sentinel-2 images based on the FDI (Biermann et al., 2020) and NDVI indices. Then, after running neural networks on the labeled datasets, the authors concluded from the results that the neural networks learnt efficiently the spatial patterns of the floating objects and were able to detect them with competitive accuracy. The labeled Sentinel-2 dataset along with the learnt weights are publicly available.

Contribution

Despite the results in (Mifdal et al., 2021), the detection of floating objects on water bodies remains a difficult problem and many issues could impact the network's performance. For example, on Sentinel-2 image of a coastal area, there are more pixels belonging to water than to the floating objects themselves, this creates an imbalanced learning problem. Also, the variability of regions across the world makes it hard sometimes for the network to spot the floating objects, and finally, the labels suffer from noise which could be misleading to the network during the learning phase. Thus, the goal of our contribution is to make the previous networks more robust to the domain adaptation problem and noisy labels by focusing on the utilization of self-supervised learning (SSL) while using attention mechanisms. More exactly, we use contrastive SSL as it is a simple and modular approach. The latter relies on a "pretext training task", defined by the user, that helps the network to learn invariances and latent patterns in the data without the need of any human annotation. To do so, the neural network forms positive pairs by considering various augmented views of a given sample and then maximizes their similarity through the noise contrastive estimation (NCE) loss. This emerging and yet promising paradigm increases model's robustness towards noise or frugality (Cio-carlan and Stoian, 2021), and representation learnt this way are known to have a better semantic and contextual dimension and, therefore, to better disentangle the main triggers of the network's decisions. Also, the conclusions of (Carmo et al., 2021) [1] emphasized the superiority of the MA-Net (U-Net with attention modules) with respect to U-Net, which encourages the use of attention mechanisms, especially transformers encoders. Indeed, despite the power of CNNs to extract meaningful features and textures, the spatial context and objects shape are less taken into account during the learning step when comparing with transformers encoders (Tuli et al., 2021). The lack of spatial information limits the global comprehension of the image and also the deep neural network's ability to identify and distinguish floating debris from lands or ships. Combining SSL with methods based on attention mechanisms altogether has shown a significant improvement in performances, especially in domain issues and generalization to challenging settings.

References

[1] https://210507-004.oceansvirtual.com/view/content/skdwP611e3583eba2b/ecf65c2aaf278557ad05c213247d67a54196c9376a0aed8f1875681f182daeed

L. Biermann, D. Clewley, V. Martinez-Vicente, andK. Topouzelis. Finding plastic patches in coastalwaters using optical satellite data.Scientific re-ports, 10(1):1–10, 2020.

A. Ciocarlan and A. Stoian. Ship Detection in Sen-tinel 2 Multi-Spectral Images with Self-SupervisedLearning.Remote Sensing, 13(21):4255, Oct. 2021.ISSN 2072-4292. doi: 10.3390/rs13214255. URLhttps://www.mdpi.com/2072-4292/13/21/4255.

J. Mifdal, N. Longepe, and M. Rußwurm. Towardsdetecting floating objects on a global scale withlearned spatial features using sentinel 2.ISPRSAnnals of the Photogrammetry, Remote Sensingand Spatial Information Sciences, 2021.

K. Topouzelis, A. Papakonstantinou, and S. P.Garaba. Detection of floating plastics from satelliteand unmanned aerial systems (plastic litter project2018).International Journal of Applied Earth Ob-servation and Geoinformation, 79:175–183, 2019.

S. Tuli, I. Dasgupta, E. Grant, and T. L. Griffiths.Are Convolutional Neural Networks or Transform-ers more like human vision?arXiv:2105.07197 [cs],July 2021. URLhttp://arxiv.org/abs/2105.07197. arXiv: 2105.07197.

Each year, several million metric tons of mismanaged plastic waste eventually enter the ocean from coastal environments. Once in the ocean, positively buoyant plastic debris are subjected to a wide range of physical transport processes such as coastal currents, large-scale and submesoscale open ocean processes, stokes drift, direct wind transport, vertical mixing, and beaching. These processes result in the dispersion of floating plastic objects over large distances globally, leading to particularly high concentrations in the surface ocean of remote subtropical oceanic gyres. With amounts of plastic exceeding one million pieces per km2 and hundreds of kilograms per km2, these waters are often referred to as ocean garbage patches. At present, the density and spatial variability of floating macroplastic (> 50 cm) in the subtropical gyres is still poorly understood, primarily due to limited observational tools. However, such information is crucial to more accurately quantify the global inventory of plastic debris in the world’s oceans and monitor its evolution, which in turn represents critical data for ecotoxicological risk assessments and for optimizing mitigation strategies.
This study uses fixed-wing Unmanned Aerial Vehicles (UAVs) to obtain photo grid surveys of the ocean surface in the North Pacific subtropical gyre during a 6-week offshore expedition in July-August 2021. We scanned almost 100 km2 of ocean surface in 22 flights. Plastic density maps, created by applying our previously developed object detection model (de Vries et al., 2021), reveal large daily fluctuations of background density, in addition to strongly clustered floating macroplastic (> 50 cm) within the area scanned by each flight. The UAV campaign results provide a dataset that sparks new insights into the plastic density in the North Pacific subtropical gyre, and that can aid in refining advection models specifically for floating macroplastic (> 50 cm).

Plastic litter and debris are now found all over the globe, from remote plains and mountains to estuarine systems and ocean waters. In the aquatic environment, plastic litter is fractionated into smaller sizes (nano or micro-plastics, diameter < 5 mm) and undergoes biogeochemical modifications through biofouling or incorporation into organic polymers. The diversity of plastic compounds and highly variable modifications occurring in the water column make it very difficult to have a systematic approach for documenting and classifying the optical properties of plastic litter from the field or laboratory. Plastic compounds are commonly birefringent and/or semi-transparent. As a result, knowledge of their polarization characteristics might be an asset for plastic detection and monitoring from spectro-polarimetric sensors.
In this study, full single scattering and radiative transfer computations were conducted to model the polarized water-leaving radiation of submerged plastic particles based on empirical size distributions and refractive indices. Simulations were also performed for coated plastic particles to mimic biofouling effects. The radiance and other Stokes vector terms (i.e., Q and U) were simulated at several levels in the atmosphere- water column system including top-of-atmosphere. The impacts of plastic compound, size and concentration were analyzed in terms of the linear degree of polarization and angle of polarization. This analysis was performed for several wavelengths in the visible-near-infrared part of the spectrum and for a comprehensive set of viewing geometries and Sun elevations. These results enabled the characterization of polarization signatures as measured from polarimetric sensors located directly in the water column, just above the water surface, or at the satellite level. Finally, the detectability of plastics from space with past, present, and future polarimetric satellite missions, including the PACE mission, was quantified and methods for field validation were delineated.