The objective of RESMALI was to perform field sampling, experiments and modelling to determine the capability to detect plastic ML from orbiting sensors. A comprehensive SWIR spectral signature library was compiled from virgin and oceanic plastics, acquired for different concentrations and kinds of plastics in varying stages of degradation, both wet and dry and some collected and frozen to preserve the biofilm that builds up on ML over time and also contributes to the spectral signature.

This library provided spectral measurements used as proxies for BOA water-leaving reflectance, which was then attenuated by applying RTMs that model the effect that the atmosphere has on light propagation depending on the gaseous composition, water vapour and aerosol properties to infer the signal at TOA.

A paper published in the Journal of Hazardous Materials presents the findings of the study:

• No significant spectral differences between raw polymers and ocean-harvested polymers were detected, which simplifies the detection paradigm for ML.
• No distinguishable signature of plastic litter with pixel fractions below 1% seems detectable.
• The materials' apparent colour, especially visible bands, is important. Most floating plastic is white or bleached, with an intense hue; thus, discrimination between plastic and natural materials, e.g., floating seaweed, and phytoplankton, may be achieved considering this property, and there is also a dependency between NIR and SWIR.
• Plastic polymers exhibited Lambertian-equivalent behaviour in the measured reflectance spectra.
• The impact of water on the refractive index (i.e., plastic/air vs plastic/water) can strongly affect the angular dependency of the light source and sensor position.
• Atmospheric correction may strongly impact ML detection capability. Most of the relevant signal from plastic polymers is in the NIR and SWIR, the same bands used in ocean colour applications to adjust atmospheric corrections.

Further information about this project can be found by following The Ocean Cleanup.

Funded by the European Space Agency

The radiometric signature of vegetation shows a characteristic 'red edge', strongly absorbing wavelengths in the visible but with reflectance increasing dramatically between the red and NIR (Sentinel-2 bands 4 and 8). In contrast, water reflects visible light but absorbs wavelengths beyond visible red. NDVI, the normalised difference between NIR and red bands, uses this divergent reflectance pattern to distinguish vegetaiotn from water. Ocean debris has high reflectance in NIR but, unlike vegetation, also in the visible red. The NDVI is neutral, or slightly positive, which enables floating debris to be distinquised from vegetation in the ocean.

By creating composite images using NDVI, combined with FAI and a band ratio index between NIR and SWIR, an algorithm was developed enabling analysis of both high reflectance environments including coastal and inland waters (e.g. rivers and lakes) and low reflectance environments including the open sea.

Based on results from case studies it was proven that current instruments could yield a certain degree of information whenever significant accumulations of debris, but with limitations. Sentinel-2 was found unsuitable for detecting stranded litter at beaches; the litter spectral signature is indistinct against the background environmental noise. Similarly, detection near the coast is limited. Nonetheless, and unexpectedly, Sentinel-2 proved particularly efficient detecting litter accumulations in rivers, which was encouraging as rivers are the main source of litter into the ocean.

The team developed a processor that made use of multiple spectral indices calculated from the discrete wavelengths available to the Sentinel-2 MSI sensor alongside a deterministic decision tree classifier to identify floating marine plastic in Copernicus Sentinel-2 images.

The processor proved capable of identifying litter items as small as 50 square metres, which allowed detection of marine litter windrows, elongated accumulations of marine debris ranging from a few tens of metres to over a kilometre in length.

This project was featured as an ESA Sentinel Success Story

Funded by the European Space Agency

It had been observed that sea foam, seaweed and plankton form features called windrows, transient hydrodynamic phenomena with lengths of the order of 10km, caused by convergent oceanic processes, particularly Langmuir circulation produced in water due to persistent winds. If sufficient concentrations of marine plastics were present these would also be entrained within windrow structures and the WASP project investigated whether this aggregation would enable currently orbiting EO sensors, specifically the Sentinel-2 MSI to use windrows as proxies for the detection of marine plastic litter.

The EOTrack Med processor was enhanced to improve cloud-masking and land-masking along with a novel machine learning filament detection algorithm to identify the distinctive shapes which are taken by windrows. This was all combined with an improved resampling methodology making use of the Super-Resolution algorithm to increase the resolution of all bands ahead of applying a reworked plastic identification index to the identified filament to determine the likelihood of it containing plastic.

120TB of Sentinel-2 products recorded over the Mediterranean Sea, a total of 287 475 images, were processed generating 104 284 outputs that identified a total of 708 355 filaments, 14 374 (2%) of which were classified as true litter windrows. The remaining 'false negatives' were reclassified by manual review. Once validated, the processor showed an accuracy of 89.14% in classifying windrow filaments.

Funded by the European Space Agency

The word 'plankton' comes from the Greek for 'drifter', it is used to collectively describe the diverse group of organisms that are carried by ocean tides and currents, being unable to swim against the advective forces. Marine litter and plankton cover the same size range and co-exist in patches entrained by ocean currents and, for larger floating litter, affected by the wind.

The motion of water parcels, and hence plankton and marine litter, can be estimated by running computational models that consider the NW Atlantic Shelf seas as a grid in which mixing between boxes is determined by the prevailing currents, or velocity field. These Eulerian models are parametrised using datasets of ocean currents, wind and waves with a 1.5km resolution as provided by the Copernicus Marine Service.

In addition, these inputs, along with auxiliary data, are used to run models that track individual litter parcels through time. These Langangian models are parametrised to include the effect of wind and stochastic effects of turbulence, along with parametric models to account for sinking, standing and re-floatation.

The service is hosted on the high-performance facility provided by the ACRI Group datacentre, AdwaisEO, on the ACRI-ST C-TEP platform, enabling daily maps with 5 day forecast to be automatically generated and delivered to registered users through an interactive web portal. A hindcast service, available to public authorities, provides estimates of the potential origin of litter stranded within a selected area of interest.

The performance of the EOTrack Med marine litter detector was enhanced by means of conditional probabilities using a Bayesian estimation combined with a univariate Receiving Operating Characteristic (ROC) curves, refined within the classic framework 'Multidimensional Signal Detection Theory' (MSDT). Classification sensitivity was improved by analysing band indices from study locations predicted to be litter hotspots.

A litter drift model of the Black Sea, deployed on Litter-TEP, provided information about the routes taken by floating debris from source, assuming riverine daily inputs, and the locations where high densities of marine litter are likely to be found.

The detection of plastic litter dumps on land is more complex than marine detection due to the array of features with similar spectral response that need filtering to isolate exclusively the location of the dump. The innovative solution was to train a CNN to identify litter dumps in a range of possible locations to consider the wider context of a pixel, in addition to its spectral characteristics, before assigning it a classification. This reduced false positives dramatically.

Funded by the European Space Agency

Sending an EO instrument into orbit is an expensive operation. The development of Sentinel-2 started in 1998, the first satellite was launched in 2015 and the second in 2016, with an expected operational life of 7 years, which has already been surpassed. The combined cost to build the two satellites was 350 million euro. That does not include the launch or subsequent operational cost. Mistakes or oversights in instrument design are costly and impact mission objectives post-launch. During operation, instruments degrade and require calibration and retrieval algorithms evolve.

A Mission Performance Simulator models the characteristics of the instrument, the orbiting platform and, for passive optical sensors, the propagation of light from target to acquisition, considering the effects of atmosphere and interaction with the optical and electronic components of the instrument, to simulate the expected signal and subsequent data package, which can be processed using retrieval algorithms as if from the real sensor.

The baseline architecture ML-OPSI conforms to that described ARCHEO-E2E: A Reference Architecture for Earth Observation end-to-end Mission Performance Simulators (2012). The algorithm for generating BOA water leaving reflectance was based on work published by Goddjin-Murphy et al. (2018) and laboratory measurements of spectral signatures from various litter types gathered from Garaba, S. et al. (2021) and Knaeps, E. et al. (2020).

By providing a modelling & simulation framework enabling researchers to build virtual experiments by combining pre-built, re-useable, building blocks within a low-code environment and thus, by democratization, reduce the barriers to entry for using the various sophisticated computational techniques available for investigating and understanding complex systems, like monitoring marine plastic from satellite.

Funded by the European Space Agency

In addition to developing these processors, ARGANS has implemented sophisticated software to predict the dispersion and fate of floating debris within the ocean using freely available current, wave and wind data. Finally, making strides towards a component-based framework that could be used to aid future instrument specification to optimise the detection of marine debris.

Going forward the aim is to continue to refine the techniques for litter detection, both marine and terrestrial, although there is a limit to how much informatoin can be extracted from Sentinel-2, and VHR is a costly alternative. There are eopportunities for improvement:

• Additional pre-processing to remove features leading to false positive, such as vessel tracks and oil slicks.
• Alternative atmospheric correction routines that do not impact the NIR and SWIR bands used for plastic detection
• Advancing the machine learning classification with novel techniques and more extensive and diverse training sets.

A remaining challenge is to develop methods that estimate quantitatively marine debris, not only that it is present. Another challenge is how to detect microplastics, that form the bulk of plastic in the marine environment. Both of these challenges are possibly unsurmountable with current orbiting instruments, but future missions could provide the required spectral and spatial characteristics needed by litter detection algoithms.

Whether to inform specification and design, provide a test environment for algorithm development, or act as a 'digital twin' to an operational instrument the availability of a generic, plugable, extensible workflow manager and model builder tailored for Earth Observaton processing is a long-term goal.

As one of the foremost experts in marine litter detection and monitoring a primary objective is to provide end user products, as an on-demand service, combining detection from space with model predictions of likely source and destination.