Marine litter weight estimation from UAV imagery: Three potential methodologies to advance macrolitter reports

In the context of marine litter monitoring, reporting the weight of beached litter can contribute to a better understanding of pollution sources and support clean-up activities. However, the litter scaling task requires considerable effort and specific equipment. This experimental study proposes and evaluates three methods to estimate beached litter weight from aerial images, employing different levels of litter categorization. The most promising approach (accuracy of 80 %) combined the outcomes of manual image screening with a generalized litter mean weight (14 g) derived from studies in the literature. Although the other two methods returned values of the same magnitude as the ground-truth, they were found less feasible for the aim. This study represents the first attempt to assess marine litter weight using remote sensing technology. Considering the exploratory nature of this study, further research is needed to enhance the reliability and robustness of the methods.

During the last decade, significant efforts have been dedicated to identify the primary sources of litter (GESAMP, 2019;IUCN and UNEP-WCMC, 2016;OSPAR Commission, 2010).It has also been found that about 70 % of litter in the ocean come from rivers (Lebreton et al., 2017;van Emmerik et al., 2022;van Emmerik and Schwarz, 2020), and that ocean circulation plays a role in transporting floating litter (Chassignet et al., 2021).On the coast, it has been outlined how the nearshore hydrodynamics influence the dynamics of litter stranding (Andriolo and Gonçalves, 2023a), while the processes responsible for burying litter on coastal dunes (Andriolo and Gonçalves, 2022) and mangroves (Martin et al., 2019), among others, have been identified.
Nevertheless, it is necessary to increase both spatial coverage and frequency of litter observations, along with enhancing the quality of litter monitoring.These improvements are essential to support and promote optimized measures aimed at tackling litter (GESAMP, 2019).In this regard, remote sensing platforms, such as satellites and drones, have been providing new techniques for monitoring floating litter in oceans and riverine waters (Martínez-Vicente et al., 2019;Salgado-Hernanz et al., 2021;Topouzelis et al., 2021;Veettil et al., 2022), as well as assessing the abundance of beached litter on coasts (Acuña-Ruz et al., 2018;Andriolo et al., 2021a;Corbau et al., 2023;Gonçalves et al., 2022a;Gonçalves and Andriolo, 2022;Martin et al., 2021;Merlino et al., 2020).
In the context of beached litter monitoring, traditional visual census surveys mostly report the number of items, the litter categories, and the spatial litter density over the limited monitored area (Andriolo and Gonçalves, 2023a;GESAMP, 2019;OSPAR Commission, 2010).However, the weight of beached litter is also an important parameter to be considered.Firstly, monitoring litter by weight can change the perception of the importance of pollution sources.For instance, fishing-related items composed 6 % by number and 41 % of weight on ten Scottish beaches (Smith and Turrell, 2021).Secondly, the contribution of different materials can vary in term of percentage, as for instance plastic litter accounted for 50 % by number but about 30 % by weight on twenty Korean beaches (Hong et al., 2014).Thirdly, weight information can help the understanding of litter stranding, accumulation and resuspension processes.For instance, weight can influence the residence time on beaches, as lighter items may be more easily transported by wind and waves alongshore and landward, whereas heavier litter is more prone to be buried (Andriolo and Gonçalves, 2023a).Finally, litter weight has direct implications for cleaning-up and removal efforts, as knowledge of litter mass can support the planning of clean-up activities with appropriate machinery and loading vehicles.
The task of weighing litter requires significant effort in the field when portable weighing scales are adopted, and/or the work of specialized personnel in dedicated facilities for laboratory measurements.Yet, the presence of large and heavy mega-litter items can further complicate the task.For such reasons, data on litter weight are sparse (Haarr et al., 2022).For instance, among 45 surveys of beached litter reviewed by Zielinski et al. (2022), <30 % reported the litter weight.Given the widespread use of drones as aerial platforms for litter surveys in coastal and riverine environments (Andriolo et al., 2022(Andriolo et al., , 2023;;Gonçalves et al., 2022a and references therein), can litter weight be estimated from images collected by such platforms?
The main objective of this study is the proposal of potential imagebased solutions to estimate beached litter weight from a drone-based litter survey.A traditional visual census survey was conducted on a 100-m sector of a North Atlantic Portuguese beach.Litter items were categorized, recollected and successively weighed in the laboratory.Before the visual census, a drone acquired high-resolution images of the monitored area.Successively, both manual and automated techniques were employed to identify and categorize visible litter on the images.From such drone-based litter categorization, we conceptualized and tested three methods for estimating litter weight, comparing the results with ground-truth data.
Overall, this work aims at promoting and improving the use of remote sensing technologies for litter monitoring and support clean-up initiatives, thereby contributing for environmental protection and mitigation efforts.

Study site and visual census
Leirosa Beach (40 • 03′16.6"N8 • 53′33.1"W) is located in a sandy coastal stretch on the North Atlantic coast of Portugal (Fig. 1).The beach has a NE-SW orientation, and it is limited northward by a 200 m-long groin (Gonçalves et al., 2022b), backward by a dune ridge with an average height of 5 m (Fig. 1).The mesotidal regime has an average amplitude of 2.10 m, reaching a maximum elevation of 4 m during U. Andriolo et al. spring tides (Andriolo and Gonçalves, 2023b;Fontán-Bouzas et al., 2022).The dominant wave regime comes from NW with average significant wave height (Hs) of 2 m, and periods from 7 s to 15 s (Fontán-Bouzas et al., 2024;López-Olmedilla et al., 2022;Santos et al., 2020).
Following the standard OSPAR monitoring protocol, a beached macro-litter visual census survey was conducted on voluntary basis in late September 2020, at the start of autumn (OSPAR Commission, 2010).The operators screened the 100-m longshore extent southern the groin, and the cross-shore width between the water edge and the dune toe (Fig. 1a), during low tide.All visible beached macro-litter items (>2.5 cm) were collected and stored, successively items were catalogued with OSPAR litter identification codes (OSPAR Commission, 2010), and finally weighed on a scale in the laboratory.The final datasheet reported the abundance of each litter type and category, along with items weight (Fig. 2).

Drone survey and image analysis
Before the visual census survey, a multirotor quadcopter drone DJI Matrix 210 RTK V2, mounting a camera RGB DJI Zenmuse X7 (24 MP, 6016 × 4008 pixels) with a 24 mm focal length, was used to take aerial images over the selected beach area (Fig. 1b).The camera gimbal acquired photos perpendicularly to the direction of the flight, from an altitude of 35 m.Images of 6016 × 4008 pixels were recorded and overlapped with 80 % front and 70 % side rates.
The orthophoto was tiled with a 2 m × 2 m grid, and visually screened by an experienced operator in GIS environment.Following the image manual screening procedure (Andriolo et al., 2021a;Merlino et al., 2021), each identified litter item was marked and labelled with the OSPAR code number (OSPAR Commission, 2010).The final outputs were the litter type and category table, and the litter spatial distribution map (Fig. 2).
Exploiting the point shapefile produced by the manual image screening in GIS environment (Fig. 2), a semi-automated object-based multi-resolution segmentation was performed with using Trimble eCognition® (version 9.4), to characterize items by their length (over the longest dimension) and area (Andriolo et al., 2021b).A second type of image segmentation, namely a semantic pixel-based segmentation (Hidaka et al., 2022(Hidaka et al., , 2023;;Sugiyama et al., 2022), classified each image pixel in order to identify litter items among other objects and the background.The algorithm was based on a High-Resolution Network (HRNet), previously trained with about 2800 beach images from monitoring records provided by the Yamagata Prefecture (Japan).The final outcome was the list of detected litter items, generally classified as anthropogenic litter (Fig. 2).The litter list provided by this semantic segmentation was superimposed on the DSM, to finally estimate litter volume by the technique described in Kako et al. (2020).The final output was the estimation of the volume of each item, which eventually contributed to determining the cubic meter mass of the entire litter bulk (Fig. 2).

Weight estimation from images
We conceptualized and tested three methods to estimate the weight of litter from drone images, taking into account the different levels of litter categorization provided by the image processing techniques (Fig. 2).The results of the three methods were compared with the actual weight of the litter bulk derived from the visual census campaign.
The first method (hereinafter, W1) considered the list of litter types identified through manual image screening.The scaled weights determined from the visual census data were assigned to the corresponding litter types recognized in the images.Three additional values (low, mean and maximum weight) were assigned following the range of weights proposed by Grundlehner et al. (2023).For each of the litter types catalogued by OSPAR (OSPAR Commission, 2010), Grundlehner et al. (2023) derived such range of weights through a combination of actual scaling and online research, to eventually estimate litter bulk weight when only count data were available.
The second method (W2) assigned an average litter weight value to the total number of items identified through both manual image screening and semantic segmentation.Unlike W1, W2 did not considered the specific litter types or categories, treating all items equally in terms of weight estimation.The litter weights reported by 16 visual census surveys conducted on beaches worldwide were used to derive a generalized mean litter weight value, which was assigned to the total number of items found by image techniques (Becherucci et al., 2017;  Faizal et al., 2022;Fruergaard et al., 2023;Hong et al., 2014;Kumar et al., 2016;Lessy and Nagu, 2020;Meakins et al., 2022;Mishra et al., 2023;Nakashima et al., 2011;Pervez et al., 2020;da Silva et al., 2015;Simeonova and Chuturkova, 2019;Smith and Turrell, 2021;Sulochanan et al., 2019;van Gool et al., 2021;Wessel et al., 2019).
The third method (W3) was based on the physical dimensions of litter items.Utilizing the area of items identified through object-based segmentation, a range of potential third-dimensional values was hypothesized to compute the total litter volume.Subsequently, given that plastic is usually the most prevalent material (e.g., Rangel-Buitrago et al., 2022b) and to simplify the method, we multiplied the total litter volume by a range of specific weights for plastic, typically comprised between 0.8 g/cm 3 and 1.5 g/cm 3 .The same specific weight values were used to estimate the total litter weight from the volume computed using the DSM (Kako et al., 2020).

Table 1
Report of the visual census survey.For each litter type, the table reports the OSPAR code number, the category, the number of items found and collected, the percentage by number (if >2 %), the total and the average weight, and the percentage by weight (if >2 %).

Visual census
A total of 1505 litter items were collected and weighed during the OSPAR-based visual census (Table 1).The most common types were foamed polystyrene fragments, with size comprised between 2.5 cm and 5 cm (37.5 %), followed by combs/hair brushes (17.6 %).The litter bulk consisted mostly of plastic (94 %, Table 2), which was represented by 32 different OSPAR litter types (Table 1).Clothes represented 2 % of the litter, while other categories constituted approximately 1 %.
The whole litter bulk weighed 24,720 g (Table 1 and Table 2), with plastic accounting for approximately 64 % (about 16,000 g).Among the singular types, foamed polystyrene fragments (2.5 cm × 5 cm) were 37.5 %.Plastic bottles, drums and containers represented 18 % of the total weight, while fishing nets comprised 13 % (Table 1).The six processed wood pieces constituted 17 % of the total weight, while both clothes and glass bottles were 7 %.The weight of other types and categories was negligible (Table 2).

Manual image screening and automated litter detection
A total of 1427 litter items were identified and marked on the orthophoto (Fig. 3a and Table 3).The operator categorized 79 % of these items, while the remaining 21 % (306 items) were labelled as "undefined items".Despite the lower numbers of recognized clothes, metals, and glass bottles, the composition of the litter bulk resembled that obtained from the visual census, assuming the "undefined items" to be plastic (Fig. 3b and Table 3).
From the object-based segmentation, the average and median litter lengths were 10 cm and 7 cm, respectively.The total area covered by litter was 52,276 cm 2 , with the average and median areas of items being 22 cm 2 and 16 cm 2 , respectively.Finally, the semantic segmentation identified 899 litter items (Fig. 3d), resulting in an estimated litter volume of 12,230 cm 3 .

Weight estimation from images
The total litter weight calculated by method W1 ranged from 10,400 g to 547 kg (Table 3).Despite the reasonable categorization of litter through manual image screening, computing the weight using the mean weight of litter types from the field survey led to an overestimation of approximately 40 % (34,657 g).When considering the individual litter type values suggested by Grundlehner et al. (2023), the use of the lower limit underestimated the weight by approximately 60 %, while employing the average values resulted in an overestimation of about threefold.Eventually, the utilization of the upper weight limit for each type overestimated the total weight of approximately fortyfold (Table 3).Of note that for items labelled as "undefined," the weight value of "plastic pieces" was assigned, as this material is commonly the most abundant in litter bulk.Overall, the total plastic weight determined the differences with the ground-truth, as wood weight contribution (17 %) was correctly estimated.
The assessments of method W2 are shown in Table 4.According to the literature review, the mean litter weight varied between 5.7 g and 25.1 g.The estimated weight by W2 was aligned with the ground-truth when the values were closer to the mean item weight from the visual census (16.4 g).However, the weight was underestimated by about 65 % and overestimated by 45 % when considering the lowest and highest values, respectively.Nevertheless, the total weight estimated considering the average and median values from the 16 studies had an accuracy of about 80 % (Table 4).Considering the number of items detected by the semantic segmentation, total weight values ranged between 5000 g and 22,565 g, with an average underestimation of about 50 % (Table 4).
Reminding that the object-based segmentation computed a total litter area of 52,276 cm 2 , best assessments from W3 were obtained when i) the third-dimension of items was hypothesized to be 0.4 cm, and ii) the plastic mean specific weight of 1.2 g/cm 3 was adopted (Table 5).However, an acceptable accuracy (±25 %) was achieved considering a third-dimension range between 0.3 cm and 0.6 cm, in combination with different values of plastic specific weight range.Finally, by adopting the same range of specific weights for the volume computed by the semantic segmentation (12,230 cm 3 ), the total weight varied between 9784 g and 18,345 g (Table 5).

Litter detection
From the visual census results, plastic category constituted about 64 % of total weight, while the number of plastic items represented 94 % of the total litter bulk (Table 2).Conversely, the number of pieces of wood (6) was negligible, whereas their weight accounted for 17 %.As highlighted by Smith and Turrell (2021), weight information can change the characterization of litter bulk.
The categorization of litter from drone orthophotos returned satisfactory results, comparable to those obtained from the visual census (Fig. 3a-b).The differences were due to the inability to recognize some semi-buried items, and/or litter hidden within beach wrack and natural debris in the images (Fig. 1d).Yet, distinguishing between processed and natural wood debris, as well as identifying wood items with colour similarity to sand, posed additional challenges (Fig. 1d).Consequently, the image analysis underestimated the number of wood pieces (Table 1 and Fig. 3).Finally, certain litter types, such as six-pack yokes, plastic caps, cutlery, small strings, syringes and small fragments (< 2.5 cm), were not identified on the orthophoto due to their limited size.The size of items also contributed to biases in the automated semantic segmentation technique, as the algorithm was unable to identify items smaller than 1000 cm 3 (Hidaka et al., 2022).

Weight estimation from images
Fig. 4 provides an overview of the final assessments of the three methods (W) proposed for estimating litter weight from drone images.Regarding method W1, which was based on individual litter type weights, assessments were susceptible to accurate litter detection and categorization by manual image screening.The image-based categorization was performed reasonably (Fig. 3b), however the total litter weight was overestimated, even when actual mean weight of each litter type was considered (Table 3).Specifically, the plastic bulk was overestimated by about 10 kg (40 %), primarily due to the overrated weight adopted for the category "other plastic items".While litter recognition in the field is accurate, image interpretation often does not allow for a rigorous distinction of item types.Thus, during manual image screening, the classification "other plastic items" may have been used to label items much lighter than 25.7 g (Table 3).From the values proposed by Grundlehner et al. (2023), upper limits seemed inappropriate, given the higher magnitude returned, whereas lower limit and mean weight values can be considered for future works (Table 3).In this regard, Table 1 and Table 2 can be used to build a more robust litter weight catalogue.
The results from method W2, which assigned a single mean weight value to all litter types and categories, were satisfactory (Fig. 4).Despite the variability in the 16 values found in the literature, ranging from 5.7 g to 25.1 g (Table 4), the overall mean weight (about 14 g) aligned with the value obtained from the visual census at the study site (16 g).The mean values found in literature were lower and of different magnitude (< 10 g) when small items (e.g., cigarette butts), which cannot be detected on drone images, constituted a substantial percentage of the litter bulk (Table 4 and references therein).
The main limitation was the low number ( 16) of works reporting litter weight, compounded by the absence of data from our study site.The available literature from Europe was restricted to ten beaches in Scotland (Smith and Turrell, 2021) and on the coast of the Black Sea (Simeonova and Chuturkova, 2019), with reports from Brazil and Argentina comprising additional sources from the Atlantic Ocean (Becherucci et al., 2017;da Silva et al., 2015).To address such limitation, we considered integrating citizen science data from Spain (https ://ambienteeuropeo.org/datos-recogidas-basuras-marinas/),however the litter weight measurements provided by volunteers were not sufficiently homogeneous and rigorous to be considered for this scientific analysis (Soliveres et al., 2024).
Other limitations and issues encountered in the application of method W2 must be highlighted.Firstly, weight values were reported differently across studies, with some authors indicating the entire litter bulk weight, others reporting the weight of single items, and some  4 and references therein).Secondly, there was a lack of homogenization in the unit systems used to report litter weight, with authors using different units such as weight (g, kg), or weight per length (g/m, Kg/m), or weight per area (g/m 2 , kg/ m 2 ).Thirdly, the influence of different scaling methods on litter weight was not clearly addressed in the literature, as litter can be weighed using Table 3 Assessments of method W1.Each litter type identified through manual image screening was assigned four values: the mean weight derived from the actual scaling after visual census, along with three values (*) proposed by Grundlehner et al. (2023): lower limit, mean, and upper limit values.All values and results are expressed in grams (g).The acronyms in the Category column represent plastic (P), clothes (Cl), wood (W), glass (G), ceramics (Ce), and sanitary waste (S).portable scales or more sophisticated laboratory equipment.Besides, it was often unspecified whether litter weight was reported as dry and/or wet, further complicating comparisons across studies.Finally, comparisons among assessments from different standardized sampling protocols can be challenging due to variations in litter classification (Cesarano et al., 2023;Fleet et al., 2021;GESAMP, 2019;OSPAR, 2010; UNEP (United Nations Environment Programme), 2021; Veiga et al., 2016).
The method W3 was based on the physical dimensions of items.While the object-based segmentation provided valuable information on item length and area (Fig. 3e and Fig. 4b), the need to hypothesize the items third dimension (Table 5) made the weight estimations of W3 data-driven, thus rendering the method more uncertain and less practical.Additionally, the total weight assessed using the litter volume was underestimated because the semantic segmentation detection was limited to items larger than one litre.Improvements could be achieved by better estimating litter volume using the DSM, although this would require higher resolution (lower GSD) to accurately detect smaller items.
Of interest is the possibility of generating a weight-based map of beached litter (Fig. 4b).With advances in estimating litter weight from drone images, the weight map can be integrated with other thematic spatial representations, such as material-based (Andriolo et al., 2020(Andriolo et al., , 2021b;;Gonçalves et al., 2022a) or area-based (Fig. 4b) maps.These outputs can enhance the efficiency of clean-up operations, by enabling the deployment of specific loading equipment tailored to the actual litter weight distribution on the beach.

Conclusions
Three methods for estimating beached litter weight from drone aerial images were proposed and assessed against ground-truth values collected in the field.These methods can be applied to different levels of litter categorization.
The most promising results were achieved by adopting the outcomes of manual image screening and using method W2, which considered a generalized litter mean weight (14 g) derived from studies in the literature (80 % of accuracy).Results from the method that considered a detailed litter categorization (W1) were of the same magnitude as ground-truth values, however it under-or over-estimated the total weight by 50 %.Finally, although the method relying on the volume of items (W3) showed the best performance, its utility was constrained by the results-driven (thus random) selection of the third dimension of the items.This arbitrary aspect makes the method less feasible for an estimation of litter bulk weight, unless a more precise volume assessment can be achieved.
The litter weight estimation from aerial images collected in coastal and riverine environments has the potential of significantly advancing litter reports, to support mitigation measures and promote optimized clean-up actions.However, the task is challenging, and given the exploratory nature of this study, the proposed methodologies have considerable room for improvements.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.U. Andriolo et al.

Fig. 1 .
Fig. 1.Study site and surveys.a) Study site location on the Portuguese map (left, red square), and aerial picture of Leirosa beach (right).The red rectangle outlines the surveyed area; b) a picture of the drone Matrix 210 (upper) mounting the camera Zenmuse X7 (lower); c) a picture taken in the field during the visual census campaign; d) examples of cropped images acquired by the drone, showing different types of litter among natural and machined wood debris.(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 2 .
Fig. 2. Workflow.The diagram illustrates the steps involved in the litter weight estimation process based on two primary data sources: visual census and drone images.It outlines the Methods used for litter detection and categorization, the level of litter classification (Litter assessment), and the key parameters used for litter Weight Estimation.The colour-coded sections distinguish among the frameworks for litter and weight assessments based on detection methodologies used on drone images.The numbered Ws inside boxes indicate the parameters and techniques employed for weight estimation.The drone image source is divided into orthophoto (RGB map) and Digital Surface Model (DSM), which represents the 3D model of beach elevation.

Fig. 3 .
Fig. 3. Results for manual image screening and automated detection techniques.a) Map of marked items based on categories; b) percentage of categories from manual image screening (external pie), in comparison with ground-truth percentage by number (internal pie, italic font); c) examples of object-based segmentation assessments, with red contours outlining litter items; d) examples of outcomes from semantic segmentation, with detected items masked in red; e) histograms of length (left) and area (right) of litter items obtained by the object-based segmentation.Dashed and straight black lines indicate median and mean values, respectively.(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4 .
Fig. 4. Summary of weight estimation assessments.a) Scheme illustrating weight estimations from the three method.For W1, green squares indicate the final total weight values, while text labels report the source value used for the estimation.For W2 and W3, red and blue boxplots represent values obtained using object-based segmentation and semantic segmentation, respectively.Median values are indicated by filled dots.b) Map of area occupied by litter computed from the object-based segmentation (upper), and map of weight generated from method W2 (lower).Grids are composed of 2 m × 2 m tiles.(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 2
Quantity and weight of all categories reported from visual census.

Table 4
Assessments of method W2.Each reviewed publication includes the study site location (country and continent), related sea/ocean, and mean litter weight.These values were applied to the number of items returned by manual image screening and semantic segmentation.The acronyms in the Continent column represent Asia (A), Africa (AF), Europe (E), Oceania (O), and South America (SA).The last two lines, in italic, report the results for both detection techniques, considering the overall mean and median weight values.

Table 5
Assessments of method W3.The total litter area estimated by object-based segmentation was multiplied by various hypothesized third-dimensions of items to compute the volume.Specific weights include a range of plastic values.Bold font highlights the best results.