5. Acksen et al. 
2.3. Mesoplastic surveys 
Mesoplastic particles, in the size of 1-10 mm (Hartmann et al., 
2019), were extracted from the sandy sediments of the six surveyed 
veaches following the frame method described by Haseler et al. (2017). 
This method was applied with small adaptations, i.e., we used a smaller 
sample volume and had a wider spatial distribution of the samples across 
-he beaches. For identifying the latter, the width of each beach was 
measured and the total distance was divided into three equally sized 
sections: back, middle, and waterline. In each zone, three replicates 
were taken while keeping a 20 m distance between them. The replicates 
were always located within the 100 m beach debris survey transect that 
can parallel to the waterline and each of them covered an area of 0.5 m* 
which was identified by a wooden frame (Supplement Fig. A4). After 
»lacing the frame on the sand, the upper 30 mm of sediment was 
removed with a steel shovel and this volume was then sieved with a 
1000 um stainless steel analytical sieve. Depending on the sediment 
characteristics, the sand was either sieved dryly or was washed through 
:he sieve using non-filtered seawater that was previously inspected for 
che presence of plastic particles with the naked eye. After sieving, po- 
:ential mesoplastic particles were then picked with tweezers from the 
naterial that remained on the sieve. In total, 9 replicates were collected 
;er beach, with three located in each beach section, resulting in 54 
samples overall. 
All potential plastic particles were analyzed with a hyperspectral 
imaging (HSI) camera to identify their polymer type. For this, the 
camera was used as described in Beck et al. (2023): Particles were placed 
onto black slides without touching each other and photographed using 
che HSI camera. The model for identifying the polymer types is 
embedded in a custom analysis tool created at GEOMAR, which has been 
;ed with the spectra of reference particles made of Polyethylene (PE), 
Polypropylene (PP), Polystyrol (PS), Polyvinylchloride (PVC), Poly- 
ethylene teraphthalate (PET), Polycarbonate (PC), Poly(methyl meth- 
acrylate) (PMMA) and Aliphatic polyamides (PA66). These were 
compared to the spectrum obtained from each analyzed particle, 
‚esulting in a probability for each item representing a certain polymer 
'ype. Particles were assigned to a certain polymer when they showed a 
similarity to the corresponding reference spectrum of at least 70%. 
Darticles with a lower probability were not regarded as plastics and thus 
ıot included in the statistical analyses. This methods limitation lies with 
‘dentifying transparent and black particles, as they do not clearly stand 
out from the black background. In total, 9157 particles were collected 
rom the sieve and analyzed using the HSI camera, of which 29.5% were 
.dentified as plastic particles. Other particles were assumed to be of 
natural origin and may have consisted of sand or stone. 
2.4. Statistical analyses of macrodebris and mesoplastic data 
For the statistical analyses and for the visualization of the data the 
software R version 4.3.1 (R Core Team, 2018) and Matlab R2024a 
(MathWorks, 2022) were used. 
The mean accumulation rate of beach debris was calculated as the 
mean mass/number of debris recorded during all surveys per beach, 
excluding the first survey, which represented an initial beach cleaning 
and therefore did not reflect post-cleaning accumulation. Average 
accumulation rates were transformed into the unit items/m*/d or g/m”/ 
d, respectively, while the original sampling data were kept in the unit 
items/m?/4 d or g/m?/4 d, respectively. Dissimilarity in macrodebris 
composition was visualized with non-metric multidimensional scaling 
ordination (MDS) plots that assign relative distances between replicates 
>ased on a Bray-Curtis dissimilarity matrix and plot them in a two- 
dimensional, non-metric graph. The Bray-Curtis index is robust against 
zero inflation and therefore appropriate for analyzing the data we 
collected (Hajbane and Pattiaratchi, 2017). Furthermore, Permutational 
Multivariate Analysis of Variance (PERMANOVA) with 999 permuta- 
tions was used to test whether beach debris composition is influenced by 
Marine Pollution Bulletin 228 (2026) 119525 
{he factor “Beach” (six levels, i.e., the six locations). The analyses were 
z:onducted without prior data transformation; however, replicates in 
which no macrodebris was detected were excluded. The combination of 
MDS and PERMANOVA allows both visual assessment and formal testing 
>f compositional differences among sites. 
With Pearson's r we evaluated the correlations between the mean 
abundances of plastic debris per beach that was found during the initial 
:leaning and during the nine following surveys to the mesoplastic 
abundances as well as correlations between debris density and debris 
mass. 
A GLM (Generalized Linear Model) with family negative binomial 
and type III sums of squares was used to determine the influence of the 
factors “Beach” and “Beach section” (three levels, i.e., back, middle and 
waterline) on mesoplastic densities. The negative binomial family was 
°hosen to account for overdispersion in count data. For all statistical 
tests, results were considered statistically significant at p < 0.05. Sig- 
nificance levels and test statistics are reported in the Results section and 
indicated in the corresponding figures and figure captions. Assumptions 
of the models, including dispersion and variance homogeneity, were 
verified where appropriate. 
2.5. Ocean current modeling 
To identify potential origins of the positively buoyant plastic debris 
that we found on the beaches of Säo Vicente, backward ocean modeling 
techniques were employed using the PARCELS framework (Delandmeter 
and van Sebille, 2019). This is based on the Lagrangian equations that 
zalculate particle movements from ocean circulation fields (van Sebille 
at al., 2018). PARCELS v2.3.1 was applied to daily averaged velocities 
from output of the VIKING20X model, which has a 1/20° resolution nest 
for the Atlantic Ocean and realistically simulates the ocean circulation at 
mesoscale resolution (Biastoch et al., 2021). 
To identify possible origins of the plastic particles found on the 
beaches of Säo Vicente, virtual particles were released in the model in 
three defined release zones around the island: NW (16.9-16.98°N and 
25.0-25.1°W), NE (16.9-16.98°N and 24.8-24.0°W) and S 
(16.71-16.82°N and 24.925-25.025°W) (Supplement Fig. A5). The 
virtual particles have no weight or volume and therefore best mimic 
small and light plastic items collected in this study, rather than larger, 
heavier macrodebris items. They were positioned half a degree away 
from the coastline and were traced backward in time to track the paths 
they travelled before arriving at the shores of the island. The simulations 
were carried out in two dimensions, assuming that the majority of 
plastic items have a lower density than water (Guo and Wang, 2019) and 
therefore do not traverse through the water column. Particles were 
randomly released in the first four depth layers (down to 27.9 m). Wave 
processes like Stokes Drift or the Langmuir circulation as well as wind- 
age processes, which would affect the upper few meters are not 
modelled by VIKING20X and are thus not considered here. Simulations 
were started in five consecutive years (2011 to 2015). In every year, 
20.000 particles were released in the weeks 20, 22, 24 and 26, which 
zorresponds to the months May and June in which the sampling was 
:onducted at Säo Vicente. In total, 400,000 particles per release zone 
were simulated to obtain robust results. Each simulation was run for 
5000 days; however, only the first 1000 days were considered for the 
ınalysis of typical pathways, as at longer timescales no clear patterns 
were discernible and particles became dispersed across the entire 
Atlantic Ocean (Fig. 6D). The particle tracking was performed backward 
in time: the simulation started near the shores of Säo Vicente and ran 
‚000 days into the past to reconstruct the particles' pathways prior to 
their arrival at the island. The simulations with the ocean model were 
avaluated using Python libraries (i.e., cartopy (Met Office, 2010-2015) 
and xarray (Hoyer and Hamman, 2017)) for plotting georeferenced data. 
The single trajectories of the simulated particles were averaged to create 
probability heatmaps that show the most likely pathways.