The catchments of the studied rivers Rhine, IJssel, and Meuse (Figure 1), are heavily industrialized and densely populated (~300 inhabitants/km²) (van der Wal et al., 2013). The river Rhine (Bovenrijn) enters the Netherlands at Spijk, 161 km from the river mouth. At 147 km the Rhine bifurcates into the Waal (67% of the discharge), Nederrijn (22%), and IJssel (11%) (Schielen et al., 2007). The Waal and Nederrijn then converge at 42 km from the river mouth. The river Meuse enters the Netherlands at Eijsden, 250 km from the river mouth, and discharges 10% of the mean discharge of the Rhine-system (230 and 2,200 m³/s, respectively). Near the coast (~80 km from the sea), the branches of the Rhine and Meuse systems converge and intertwine. Ultimately, the Rhine-Meuse system drains into the North Sea, while the river IJssel drains into lake IJssel after 125 km.

Sampling locations were chosen to be at the upstream and downstream end of the Dutch section of the rivers Rhine (R), Meuse (M), and IJssel (IJ) (Figure 1). Supplementary material A provides a detailed description of the sampling areas. The sampling areas at Nijmegen (R1) and Rotterdam (R3) are located along the river Rhine, while Arnhem (R2) is located at the Nederrijn beyond the first major bifurcation of the Rhine. Arnhem (IJ1) and Kampen (IJ2) are situated on the river IJssel, while the river Meuse was sampled at locations in Maastricht (M1), Ravenstein (M2), and Moerdijk (M3). Location M3 is located beyond the point where the rivers Rhine and Meuse merge, and is therefore affected by both river systems. Location M3 and R3 are in the tidal zone, and can therefore be subject to bidirectional currents.

Riverbank macrolitter was collected once per month between January and December 2021 at eight riverbank sites. Location R2 was sampled only in January and December, and location M1 was not sampled in January due to limited sample collection and processing capacity. The width of the sampling area was defined as the distance from the waterline to the high waterline, having a maximum value of 25 m (van Emmerik et al., 2020). The waterline is defined here as the interface between the river and the riverbank. The high waterline can be identified in the field by the fact that a proportion of the organic matter floating at the river surface is deposited at this elevation along the water margin once the peak flow begins to recede. Sampling was carried out until one of the following criteria was met: (1) coverage of 100 m length, (2) collection of material equaling 80 L, or (3) a sampling time exceeding 90 min. These limits were set based upon the availability of surveyors for the sample collection, the state of the riverbank (the required sampling time can be considerably higher if there is dense vegetation), and available capacity for subsequent laboratory analysis of the sampled material. The width of the sampled locations varied between 1 and 10 m and the length between 10 and 100 m. It should be noted that riverbank sampling is biased toward larger items, since smaller items are more difficult to identify by eye (Hanke et al., 2019), hence statistics for the smaller macrolitter items (<1 cm) should be taken with caution.

Collected samples were analyzed in the Laboratory for Water and Sediment Dynamics at Wageningen University. First, the items were manually and superficially cleaned of sediment and organic debris to preserve the state in which they were sampled. Items may have fragmented during transport, which may have led to more litter items being analyzed in the laboratory than originally sampled. Second, the items were categorized using the River-OSPAR protocol (Supplementary material B), developed by the North Sea Foundation (van Emmerik et al., 2020). This protocol is based on the OSPAR guidelines for beach litter monitoring (OSPAR Commission, 2010), with adjusted categories to better account for items frequently found in (Dutch) rivers. The protocol includes 111 specific item categories, divided over nine parent categories (i.e., plastic, rubber, textile, paper, wood, metal, glass, sanitary, and medical items). The River-OSPAR categorization system gives a detailed overview of the abundance of various types of litter. To facilitate direct comparison with other categorization methods in future research efforts, we included a “conversion table” (Supplementary material F) for rapid re-categorization in one of the other published categorization methods (Nally et al., 2017; Kiessling et al., 2019; Schwarz et al., 2019; Vriend et al., 2020a; Fleet et al., 2021).

Finally, we determined the mass, length, and width of the 14,052 items sampled between January and May, and in the months of August and November. Due to limited resources, items were not analyzed in the other months. The mass was weighed on a scale (0.01 g accuracy). In case individual items did not reach the minimum detectable mass, multiple items of the same category were weighed collectively, and a mean value assigned to each. For item length and width, the two longest axes were measured with 0.1 cm accuracy.

In total 16,488 items (184 kg) were collected and categorized from eight riverbanks over 12 months, of which 14,052 (85%) were measured and weighed. For a detailed description of the length distribution of the items, see Supplementary material E. The majority of items were plastics (70% of item count, 33% of total mass) and mainly composed of unidentifiable plastic fragments (50% of all items) (Table 1). This result is in line with the findings of van Emmerik et al. (2020), who found 55.8% of riverbank litter items to be fragments along the Dutch Rhine-Meuse system. Although plastic dominates the collected item count (Table 1), local spatial variations exist (Figure 2). This can mainly be contributed to the type and use of riverbank (Supplementary material A), which play a role in which items are trapped and retained (Liro et al., 2022). For example, recreational areas, such as R1, show a lower percentage of plastic items (for example only 15% of item counts for R1) and are dominated by consumer items such as cigarette filters, metal bottle caps, and glass bottles.

The average item mass was 11.1 g (6.1 g for plastics), and the median mass was 0.55 g (0.53 g for plastics) (Table 1). The summarizing statistics per item category can be found in Supplementary material C. The difference between the mean and median mass indicated a highly positively skewed distribution with many light items and relatively few heavy outliers. The large number of fragments (for example soft fragments, hard fragments, foam fragments) are responsible for this skewedness (Figure 3A). Heavy outliers include items of scrap metal such as bikes, and metal pipes (Figure 3B). The skewed distribution may have far reaching consequences for setting up a mass-balance using only summarizing statistics. For example, estimates of floating plastic flux, based upon items per hour (which is subsequently converted to mass per year), can differ by an order of magnitude when using either the mean or the median mass for this conversion (van Emmerik et al., 2022).

The 10 most frequently found items (Figure 3) represent 56% of the total amount of items and 65% of the total mass. The 20 most abundant items represent 66% of the total item count and 87% of the total mass, respectively. The top 10 items vary strongly when considering the item count or mass as demonstrated in Figure 3. In terms of frequency, plastic fragments, food packaging, and items related to consumables and cigarette filters are the most abundant categories (Figure 3A). In terms of mass, the top 10 items mainly consist of higher-density items such as metal (mean mass 41 g), wood (mean mass 176 g), and glass (mean mass 27 g) (Figure 3B). This discrepancy between abundance in count and mass emphasizes the importance of mass statistics for reliable estimates of litter mass balances. Although accumulated material on riverbanks is often expressed in item count per surface area, item mass per surface area is more relevant for closing the mass balance. Considering that items will likely increase over time due to fragmentation, we consider item mass per surface area a more appropriate indicator for riverbank litter accumulation.

Item characteristics in the dataset can vary significantly within and between litter categories. To be able to give an accurate measure of mean, median, and standard deviation of litter item categories (Supplementary material C), the sample size must be large enough to capture the mass and length variability within a category. The number of items needed to accurately represent category statistics (within a certain uncertainty level), depends on the heterogeneity of the category. Aggregated categories in the River-OSPAR system (e.g., soft fragments larger than 2.5 cm), may have large variability in item mass and size. For categories consisting of relatively uniform items (e.g., cigarette filters) this may be the opposite. The variability within a category can be characterized by a category heterogeneity Ψ (Equation 1) and is presented as histograms of length and mass (Figure 4). Wider distributions, such as that of soft and hard fragments, belong to more heterogeneous item categories, which is reflected in Ψ (1.03 and 0.92 for item length, respectively). Note the axis scale break in the x-axes of Figures 4F–J, which indicate a wider histogram than inferred from the visible histogram. Narrower distributions, such as cigarette filters and metal bottle caps are described by a lower category heterogeneity (Ψ = 0.08 and Ψ = 0.14 for item length, respectively). Item heterogeneity is one of the most important factors that determines how many items should be sampled to obtain representative item statistics and these SSRs are discussed below.

By collecting more litter items, the item statistics (such as median and mean mass or length for example) become less uncertain, and this is especially relevant for heterogeneous litter categories. The amount of statistical uncertainty decreases with increasing sample size, meaning that the possible range of outcomes of the mean or median from the subset, differs increasingly less from the total population. However, uncertainty shows an inverse exponential decrease with sample size. Larger sample sizes only reduce statical uncertainty to a minor extent after a certain threshold. This threshold represents the minimum number of item samples that is required in order to obtain a representative number (within certain confidence bounds) of mass and length statistics.

To describe the mean mass of all litter at the sample locations with a maximum deviation of 10% of the mean based upon the total population with 90% confidence, at least 8,900 items need to be sampled and measured (63% of the total amount of weighed items). To capture the representative mean length 1,200 items (9%) need to be collected, while only 173 items (1%) are needed to describe the median mass (Figures 5A–D). The more heterogeneous an item category, the more samples need to be collected to obtain representative mass and length statistics. An example for the SSR of a homogeneous and a heterogeneous subclass is presented for the heterogeneous category “soft fragments larger than 2.5 cm,” 990 items (42% of full sample) are needed to find a mean mass (within 10% of the mean mass based on the full population) with 90% confidence (Figures 5E–H). When determining the mean mass of homogeneous item categories such as “metal bottle caps” (Figures 5I–L), only 38 (6% of full sample) items suffice.

The number of samples to be collected and measured depends on the acceptable confidence boundary and a maximum level of deviation from the mean of the total population. In the aforementioned examples, a maximum deviation of 10% was allowed and estimated with 90% confidence. With these conditions, an accurate representation of the mean mass of food packaging is reached when 150 items are measured. However, if a deviation of ±20% is permitted, only 110 items are needed to reach the uncertainty required. Similarly, if a confidence boundary of 50% is permitted, only 95 items are required to represent the mean mass (±10%). The level of confidence and maximum level of deviation allowed therefore impact the SSR.

We show the SSR of 59 item categories with more than 10 items in Table 2, which may be used in to find a balance between statistical uncertainty and sampling effort in future monitoring efforts. These 59 item categories make up 89% of total amount of collected items. The mean SSR equals 158 items, while the median equals 40 items. Our dataset does not include sufficient samples for all categories to provide an estimate of the mean mass within the selected confidence boundaries and deviations of the mean in this study. When the number of items needed to represent the mean mass is equal to the total number of items collected (indicated by the red shade in Table 2), or when a level of uncertainty (confidence boundary and deviation from the mean) is never reached (represented by N/A in Table 2), it is not possible to provide a SSR. For the highest confidence boundary (95%) and lowest deviation from mean (5%), this is the case for 37 items categories. Table 2 also shows the category heterogeneity for each item category, calculated based upon the available dataset, even if it was not sufficiently large enough to determine SSRs. As demonstrated in the aforementioned examples, to obtain the same uncertainty levels in the mass-size statistics of riverbank litter, the SSRs of heterogeneous item categories are higher than of homogeneous item categories. This is underlined by the correlation (R-squared) between SSR and category heterogeneity for these 59 item categories, which is on average 0.45, but varies between 0.12 and 0.60.

The SSRs can be the baseline for monitoring protocol design and serve as a rule of thumb or indication when making an initial design. If required, the SSR analysis can be expanded to calculate SSR based on median mass, mean, or median length and mean or median width, based on this dataset. Since the SSR analysis depends on the used item categorization method, we included a “conversion table” (Supplementary material F) for rapid re-categorization in one of the other published litter categorization methods (Nally et al., 2017; Kiessling et al., 2019; Schwarz et al., 2019; Vriend et al., 2020a; Fleet et al., 2021).

The SSRs of the litter items can be used to assess the heterogeneity of specific locations or entire rivers. This application is shown in Figure 6, which displays the litter heterogeneity based upon item count in the Rhine (R1, R2, R3), Meuse (M1, M2, M3), and IJssel (IJ1, IJ2) rivers, assuming a 90% confidence interval with maximum deviation of 10%. The litter on the riverbanks of the river Meuse and IJssel belong mainly to heterogeneous categories such as the large amount of hard and soft plastic fragments >2.5 cm (SSR 1,300 and 1,000, respectively). Contrastingly the river Rhine riverbanks encompass mostly homogeneous categories. When zooming to location-level heterogeneity (Table 3), it is clear that location R1 accounts for this. Location R1 can largely be described as a homogeneous sampling location, which contributes to the large number of homogeneous items in location R1 (Table 3), such as cigarette filters (SSR 11) and metal bottle caps (SSR 38) (Supplementary material D). The heterogeneity of each sampling location (assuming a 90% confidence interval with maximum deviation of 10%) as shown in Table 3 strongly corresponds to the heterogeneity of its top 10 items (Supplementary material D).

Heterogeneity and SSRs vary considerably within and between rivers, which emphasizes the need for river and site-specific data collection. For example, more data should be collected for heterogeneous systems. Therefore, identifying litter heterogeneity per system can give an indication as to the resource investment required to accurately capture the systems' riverbank litter. When performing a Monte Carlo bootstrap analysis on all items found within a river system, with a 90% confidence boundary and a deviation of 10%, the river Rhine can be sampled by measuring 3,000 items (78% of all items found along the river Rhine). Similarly, 6,900 items (71%) are needed for the river Meuse, and 2,000 (96%) for the river IJssel. These items would give enough data to derive representative mean mass statistics, but it does not provide any spatiotemporal information. The SSR of river IJssel comprise of almost all items in our database, and more items should be collected to confirm the calculated SSR. The smaller SSR for river Rhine indicates its homogeneous character, while the larger SSR for river Meuse again confirms its more heterogeneous character. Furthermore, due to the intrinsic uncertainty within heterogeneous items, the uncertainty in litter statistics will always be larger for heterogeneous systems than for more homogeneous systems.