Finding Plastic Patches in Coastal Waters using Optical Satellite Data

Sentinel-2 data access

Sentinel-2 is an Earth observation mission developed and operated by ESA under the Copernicus Programme. The Multi-Spectral Instruments (MSI) aboard Sentinel-2A and 2B work passively, and optical data is acquired along the orbital path at high spatial resolution (10 m, 20 m and 60 m) over land and adjoining coastal waters. Thought Copernicus program, MSI data are made available at no cost to users. We downloaded Level 1C products (at-sensor radiance) via the Copernicus and ESA Open Access Hub.

Atmospheric correction

The inherent optical properties (IOPs) of floating materials can be leveraged for detection in Sentinel-2 imagery if NIR to SWIR wavelengths are conserved during the atmospheric correction process16. Ocean and atmospheric components (scattering and absorption) were subtracted from surface reflectance values using ACOLITE (Atmospheric Correction for OLI lite version 20181210.0). This marine atmospheric correction was developed for coastal waters using high resolution data from Landsat 8 and Sentinel-2, and the process is scene-based, requiring no previously defined 'dark band' like the NIR or SWIR. Instead, using the Dark Spectrum Fitting (DSF) algorithm, darkest pixels are dynamically selected based on multiple dark targets in a given image33,34,35. The DSF algorithm is described in detail in Vanhellemont and Ruddick35.

Output for surface reflectance (rhos, (ρs) was computed using ACOLITE and visualised in the Sentinel Application Platform (SNAP) for further processing.

Defining a floating debris index

At 10 m × 10 m, the highest spatial resolution of the Sentinel-2 Multi-Spectral Instrument, individual items of debris are likely to be below detectable limits until aggregated into patches. To enhance detection of patches floating on the ocean surface in Sentinel-2 imagery, we developed a floating debris index (FDI) using four of the twelve MSI bands (Table 1).

The FDI was based on the Floating Algae Index (FAI) developed for Landsat, Medium Resolution Imaging Spectrometer (MERIS), and Moderate Resolution Imaging Spectroradiometer (MODIS)18,36,37. In place of the red band, where chlorophyll is most absorptive, we instead use the MSI Red Edge (RE) band at approximately 740 nm. Our debris detection index thus leverages the difference between the NIR, and the baseline reflectance of NIR. This baseline is derived from linear interpolation between the NIR-flanking RE2 and SWIR1 bands:

$$\begin{array}{lll}FDI & = & {R}_{rs,NIR}-{{R}^{{\prime} }}_{rs,NIR}\\ {{R}^{{\prime} }}_{rs,NIR} & = & {R}_{rs,RE2}-({R}_{rs,SWIR1}-{R}_{rs,RE2})\times \frac{({\lambda }_{NIR}-{\lambda }_{RED})}{({\lambda }_{SWIR1}-{\lambda }_{RED})}\times 10\end{array}$$

(1)

The subtraction of a baseline from the NIR reflectance serves to minimise sensitivity to changes in atmosphere and observation (aerosol type and thickness, solar/viewing angle, and glint), allowing for detection of floating objects through thin cloud or haze36.

The FDI was applied for subpixel detection of plastic targets deployed off Mytilene in Greece, as well as on dense floating patches of Sargassum seaweed off Barbados, rafted tree logs in waters off British Columbia, sea foam (spume) off the east coast of Scotland, and floating volcanic rock off Tonga. All materials were floating in relatively clear waters with low to moderate turbidity.

Simultaneously, we applied a Normalised Difference Vegetation Index (NDVI) to segregate floating vegetation from other materials:

$$NDVI=\frac{({R}_{rs,NIR}-{R}_{rs,RED})}{({R}_{rs,NIR}+{R}_{rs,RED})}$$

(2)

The NDVI is based on the fact that vegetation, including algae, show an increase in reflectance spectra at around 700 nm36,38. The difference between reflectance values in the NIR and red serves as a measure of photosynthetic capacity and/or density of vegetation. High NDVI values indicate dense patches of floating vegetation and/or high photosynthetic activity, while water generates low to negative NDVI values (no units).

Study sites

To develop and test the method, Sentinel-2 scenes containing plastic were required. The Alfred Wegner Institute (AWI) LITTERBASE portal collates scientific studies that include or are focused on marine plastics, and summarises results in a mapped global format. We searched for studies that published data on floating macroplastics published after 2010. Social media sites Twitter and Instagram, were monitored for posts containing keywords pertaining to plastic pollution in riverine and marine environments. Keywords and hashtags included #marinelitter #marineplastic #plasticsoup #cleanup #plasticocean #plasticpollution #oceancleanup #cleanseas #keepouroceansclean #trashtag #plasticpollution.

Based on the literature, scientific reports, news articles, and/or social media posts about marine litter posing an acute, increasing or persistent issue, several sites were identified (Table 2): the coastal waters off Accra (Ghana), the San Juan islands of British Columbia (Canada), Da Nang (Vietnam), and the east coast of Scotland (United Kingdom) were examined for floating debris. A number of Sentinel-2 scenes were identified for each. Following the method shown in in Fig. 4, we focused on near-shore waters within these scenes, where aggregating features such as river plumes, fronts and/or eddies were visible in the 'true colour' (RGB) imagery. Imagery exhibiting whitecaps were not used as these are also reflective in the solar spectral range3,39,40.

Determining spectral signatures of plastics and plants

During the Plastic Litter Projects (2018 and 2019), the Marine Remote Sensing Group from the University of the Aegean deployed floating targets of 5 × 5 m, 5 × 10 m, 10 × 10 m, and 5 × 20 m sizes. Targets were composed of plastics bags, bottles or fishing nets, and were deployed off Mytilene in Greece on the 7th of June 201821, 18th of April, and 3rd and 18th of May 2019. Target detection in Sentinel-2 imagery acquired on these days was carried out at subpixel level using the FDI, which allowed for detection of 9 useable pixels in total. A spectral 'signature' for plastic was generated from the mean of these subpixel detections (n = 9).

To generate spectral identifiers for natural materials likely to be aggregated with floating macroplastics, rafts of Sargassum detected off Barbados on the 29th of January 2019 were used to generate a mean spectral signature for seaweed. In many coastal waters, mixed aggregations are also known to include non-photosynthetic vegetation such as driftwood41,42. Thus, a mean spectral signature of timber or woody debris was generated from rafted logs floating in waters off British Columbia, Canada. Finally, foam (spume) from a river off the east coast of Scotland and aggregations of floating pumice off Tonga were used to generate mean spectral signatures of natural but non-plant floating materials (Fig. 1).

Manual work flowSupervised classification

After detecting suspected plastics using spectral signature, NDVI and the FDI, we tested if it was possible to discriminate different floating objects using a Naïve Bayes (Bayesian) classification. Although there are a number of supervised classification algorithms which could have been used, the Naïve Bayes algorithm was chosen as it requires only a small number of samples to train, and has demonstrated good performance compared to other algorithms43. This Naïve Bayes classifier is a probabilistic model, which relies on the assumption that predictors/features are independent - hence 'naive'. For given values of FDI, NDVI and remote sensing reflectance, the classifier computes the probability of a detected pixel belonging to each of the classes, and assigns it to the one with the highest probability. We used the GaussianNB implementation within the Python scikit-learn library.

Features used for the classification were FDI and NDVI and remote sensing reflectance at 740 nm (Red Edge), 833 nm (NIR), and 1610 nm (SWIR). The remote sensing reflectance was included to aid differentiation of materials, rather than just using the band indices with the subset of wavelengths chosen as the ones showing the largest differences in the spectral shape.

The Naïve Bayes classifier was trained using: the validated plastics from Durban (n = 53), seaweed detections from around Barbados (n = 48), timber detections from British Columbia (n = 60), spume detections from Scotland (n = 17), and seawater from all of the above (n = 20). The plastic targets from Mytilene were not included in the training set, and were instead reserved for validation data in the testing dataset.

The detections off Durban Harbour represented large quantities of plastic in a non-staged setting (i.e., plastics hadn't been placed specifically for detection purposes). In this case, flooding had washed substantial quantities of plastic litter into the harbour. Photos posted to social media and reported in the news showed that the waters here were filled with floating macroplastics and plant debris on the 22nd and 23rd of April 2019 (Fig. 5).

Within two days, the floating plastics in particular had been washed into the sea and back along the beaches for "kilometres on end" (J Papendorf 2019, pers. comm., 21 June). Using the FDI, bright pixels were detected along fronts or plumes and through gaps in heavy cloud in a Sentinel-2B image acquired on the 24th of April 2019. Over 50 pixels within the extensive floating debris patches matched the spectral signature of plastic. These were identified and classified as such. Using the 53 plastic detections from Durban to train the Naïve Bayes algorithm also allowed us to ensure (test) that the 9 known plastic targets were always identified as such during our classification process.

Divers find 3 tonnes of plastic floating off the Bali coast