UAV Approach for Detecting Plastic Marine Debris on the Beach: A Case Study in the Po River Delta (Italy)

: Anthropogenic marine debris (AMD) represent a global threat for aquatic environments. It is important to locate and monitor the distribution and presence of macroplastics along beaches to prevent degradation into microplastics (MP), which are potentially more harmful and more difﬁcult to remove. UAV imaging represents a quick method for acquiring pictures with a ground spatial resolution of a few centimeters. In this work, we investigate strategies for AMD mapping on beaches with different ground resolutions and with elevation and multispectral data in support of RGB orthomosaics. Operators with varying levels of expertise and knowledge of the coastal environment map the AMD on four to ﬁve transects manually, using a range of photogrammetric tools. The initial survey was repeated after one year; in both surveys, beach litter was collected and further analyzed in the laboratory. Operators assign three levels of conﬁdence when recognizing and describing AMD. Preliminary validation of results shows that items identiﬁed with high conﬁdence were almost always classiﬁed properly. Approaching the detected items in terms of surface instead of a simple count increased the percentage of mapped litter signiﬁcantly when compared to those collected. Multispectral data in near-infrared (NIR) wavelengths and digital surface models (DSMs) did not signiﬁcantly improve the efﬁciency of manual mapping, even if vegetation features were removed using NDVI maps. In conclusion, this research shows that a good solution for performing beach AMD mapping can be represented by using RGB imagery with a spatial resolution of about 200 pix/m for detecting macroplastics and, in particular, focusing on the largest items. From the point of view of assessing and monitoring potential sources of MP, this approach is not only feasible but also quick, practical, and sustainable.


Introduction
The presence of anthropogenic marine debris (AMD) along coastlines globally is one of most critical pollution issues to date [1][2][3][4][5][6][7][8]. In particular, plastics represent from 60 to 95% of the global marine litter in the oceans [9][10][11][12]; therefore, understanding the dynamics in which they originate, move, and accumulate is crucial [13]. The monitoring of plastic debris abundance in the marine environment [14][15][16] is important for assessing the potential impact of measures adopted to reduce the presence of plastic AMD. This work can be complicated by large spatial and temporal heterogeneity in the amount of plastic debris, as pointed out by Ryan et al. [17]. Surveys at sea can sample floating [18] and suspended debris using a manta trawl net or sample litter on the seabed [19] with submersibles or remote-operated vehicles or trawling equipment. However, most of the studies show that AMD accumulates along beaches, and hence beach surveys [20,21] are not only necessary but also represent the most common monitoring method for macro-debris. That said, beach cleanups can alter litter loads and abundance on beaches, and beach dynamics can results achieved (i) by varying the spatial resolution of images, (ii) by using information provided by an elevation model, and (iii) by using multispectral (MS) data. Finally, we used two operators with a different level of expertise in the knowledge of the coastal environment every time. The novelty of this work is exploring all of these aspects to describe the optimal method for performing UAV-based mapping of litter and macroplastics on the beaches.

Materials and Methods
In this section, we present the methodology from the in situ data collection to the mapping and classification of the marine litter based on UAV-derived data. Figure 1 summarizes the overall set of information acquired and processed to obtain (i) the mapping of AMD and (ii) a ground truth dataset.

Test Site Description
In this work, we selected a test site located in the Natura 2000 site of Sacca di Goro-Foce del Po di Volano in Northern Italy (Figure 2a). This site is one of the NET4mPLASTIC project sites used for characterizing the origin, abundance, distribution, and potential accumulation area of MPs and co-pollutants in marine systems. In fact, Po river (652 km long) is the main Italian river, and its basin covers an area of 71,000 km 2 , which flows through 3200 municipalities in 7 regions. Since the Po river collects the water of the entire basin of the Pianura Padana, it carries a notable amount of material into the Adriatic Sea. In this way, a test site within this river delta (Figure 2b) represents an area of high macroplastic and litter accumulation. The examined coast stretch is situated in the northern part of the Lidi di Comacchio, which is a well-known Italian seaside resort in the Northern Adriatic Sea. The beach that is the focus of this research is at Lido di Volano, which is located in a microtidal context, has a NNE-SSW orientation, and is dominated by low wave energy. The beach consists of fine sand and has a morphology with well-developed single swash bar systems. The beach width ranges between 20 and 60 m.
For the experiment, ten transects (i.e., sampling units) have been staked out (Figure 2c), and 4 to 5 of them have been surveyed for the research presented here. Each transect had a width of 10 m and an approximate length of 45-50 m, from the shoreline to the foredune.

UAV Image Acquisition
The acquisition of UAV image datasets has been performed using a DJI Phantom 4 RTK (DJI-P4RTK, Figure 3a) and a DJI Phantom 3 Professional (DJI-P3P, Figure 3b). Both UAVs were provided with their own RGB camera (20 Megapixel for DJI-P4RTK and 12 Megapixel for DJI-P3P). The DJI-P3P was also equipped with a Micasense RedEdge M multispectral camera (1.2 Megapixel) able to acquire data in the blue, green, red, red edge, and near infrared (NIR) wavelengths. The main differences between the aircrafts were (i) that the app used for the mission planning of DJI-P3P (Pix4D Capture) allowed us to fly at a minimum altitude of 10 m above ground level, whereas the DJI-P4RTK app integrated into the remote controller (GS RTK) had the limitation of a minimum altitude of 25 m, and (ii) that the DJI-P4RTK was used to acquire image datasets with a centimeter-level georeferencing accuracy due to an on-board multi-frequency, multi-constellation GNSS receiver in RTK mode. A base receiver (DJI Mobile Station 2, Figure 3c) was set up on a benchmark whose coordinates were surveyed using a Topcon GR-3 geodetic receiver in Network-RTK (NRTK). Alternatively, an approach based on NRTK can also be used for drone navigation. Ground control points (GCPs) were deployed using targets prior to image acquisition and surveyed in NRTK mode ( Figure 3d) in order to achieve a centimeterlevel accuracy for DJI-P3P image datasets as well. Each GCP position was collected through a 30 s stop-and-go method. In this way, data collection was framed in the official Italian reference system ETRS89-ETRF2000(2008.0). Topcon GR-3 in NRTK mode was also used for the ground survey of beach litter (Figure 3e) as reported in Section 2.2.2.
The data collection was repeated twice, on 25 October 2019 and 5 November 2020. Automatic flight missions were planned in order to map the test site extent at different altitudes. In fact, the ground sample distance (GSD) was varied to assess mapping results. Altitudes and corresponding GSDs are reported in Tables 1 and 2. Flight missions at very low altitudes were focused on a smaller amount of transects.  After flying the drone over the test site and before collecting the material, a litter survey was performed at a level of transect, as previously mentioned in Section 2.2.1. Items larger than 2.5 cm were surveyed, collecting both the position and a detailed picture. Data acquired in this way were used to generate a database of AMD spatial distribution. The ground truth dataset was created, combining the GNSS RTK position collected with a description of the item derived from the picture. The color was also assessed.

Generation of orthomosaics and DSMs
UAV image datasets were used to reconstruct a photogrammetric model of the beach. For this task, we used Agisoft Metashape to generate the orthomosaic of test site transects. Accurate georeferencing was obtained using RTK camera locations for DJI-P4RTK drone or with a set of well-distributed ground control points (GCPs) for the non-RTK drone DJI-P3P during image alignment. The workflow of data processing from acquired image datasets to orthomosaics and the digital surface model (DSMs) is summarized in Figure 5. In particular, we generated RGB and near-infrared (NIR) orthomosaics and the NDVI map. Images with the smallest GSD-thus, for the configuration when flying at a 10 m altitude with the DJI-P3P aircraft-were also used to reconstruct a DSM of the beach. In fact, the elevation variation across the beach profile can provide additional information useful for classifying the litter properly. Visualizing the DSM with hillshade rendering or extracting cross-section profiles on the fly in a geographic information system (GIS) environment may allow operators to evaluate the shape of items with more confidence. However, since the generation of a DSM with a poorer spatial resolution on the ground (higher GSDs) allowed one to identify and classify items of large dimension, we decided not to generate DSMs for flights carried out at altitudes higher than 10m. All the georeferenced orthomosaics and DSMs were exported to further proceed with image-based litter detection. Products obtained after image processing are reported in Table 3.

Mapping and Classification of The Litter
Orthomosaics generated using different ground resolution datasets were used to recognize the presence of litter items on the beach, together with NIR-NDVI data and DSM. Since it is important to establish levels of operator uncertainty during this process, we assumed 3 levels of confidence as follows: • CL 1: the operator is sure that both the detection and the classification of the item are correct; a description of the item is provided during the mapping process; • CL 2: the operator is sure that the item is litter, but unsure about the type (difficult to classify); a description of the item may be provided during mapping process; • CL 3: the operator identifies something that probably is litter, but that may be also a natural element that appears similar to litter on the orthomosaic; a description of the item cannot be provided during mapping process.
Additionally, the level of experience and expertise of the operator in knowing the coastal environment and recognizing the marine litter can be different. For this reason, we selected two types of operators: operator 1 has almost no experience in marine litter mapping and about geomorphological aspects of beaches, while operator 2 is skilled and an expert in these fields. Operator 1 was different for the 2019 and 2020 surveys; operator 2 was the same for both surveys.

Collection of Beach Litter and Laboratory Analysis
After image acquisition, all the marine litter on the beach was collected for each transect and divided into transect zones (e.g., dune, beach, see Figure 6). All the material was analyzed in the laboratory to identify and classify each item according to the Master List of Categories of Litter Items reported in Annex 8.1 of the Guidance on Monitoring of Marine Litter in European Seas [29] (see Table 4).  Artificial polymer materials 20 G21 + G8 + G7 Artificial polymer materials 2 G78 Artificial polymer materials 3 G45 Artificial polymer materials 2 G22 Artificial polymer materials 1 G14 Artificial polymer materials 5 G25 Artificial polymer materials 1 Total (with no natural vegetation-related elements) 50

Validation of Image-Based Litter Description
In order to evaluate the accuracy with which operators described litter items, we carried out a classification accuracy assessment. The description of the items identified by the operators was compared with the detailed picture of the object when available (see Table 5). These results shows that all of the CL 1 items were properly recognized and described by the operators. Most of the CL 2 items were not described by the operators, due to the higher level of uncertainty implicitly associated with the difference between assigning a CL 1 or a CL 2 class. However, Figure 7 shows that only a few items were not properly described by the operators, while most of them, when described, were properly recognized.

First Campaign (2019)
The identification and classification of litter through a manual mapping based on orthomosaics at different ground spatial resolution and with the additional information provided by the near-infrared wavelengths produced the results reported in Figure 8. Both operators preferred to assign CL 2 in most cases. In fact, the classification of litter type was found to be generally hard. Larger items or items with a well-defined shape (e.g., bottles) are easy to identify and hence assigned CL 1. However, a variety of objects that are quite different from the background sand texture are distinguishable, but not recognizable. In addition, there is no guarantee that those objects are litter. Operator 1, less expert than operator 2, preferred to assign CL 3 instead of CL 2 for many items, unsure of whether the object was actually litter. While identifying every piece of litter was not possible, litter collection provided the total number of objects that were present in each transect at the time of the survey. Figure 9 shows the number of total items recognized by the operators using the different levels of information available and those collected in the field. For transects T.2, T.3., and T.4, the collection was extended up to the dune zone, where the largest items were collected. The amount of AMD, when considering the number of items (i.e., count), is generally underestimated with the sole exception of transect T.1. For transect one, the results of operator 1 and operator 2 are consistent and comparable to the count of collected items.

Second Campaign (2020)
The second beach litter survey results are slightly different from the previous one. Operator 1 assigned a high number of CL 3 (more than 80% of the total), likely due to the lack of experience in distinguishing natural elements on a beach from potential litter items. Operator 2, still maintaining a considerable percentage of CL 3 items, was able to exploit their experience and knowledge of the beach environment morphology to classify the items with higher confidence. In fact, the percentage of items with CL 3 is lower for operator 2 than operator 1: it decreases from 86% to 39% for the dataset acquired with the DJI-P3P at a 10 m altitude (see Figure 10). Consistently, litter collection provided the actual number of objects that were present in each transect at the time of the survey. Figure 11 shows the total amount of items recognized by the operators. For most transects, operator 2 found a number of items largely different from operator 1. This is due to the large attribution of CL 3 made by operator 1. Even if the total identified object count is better compared to the collected ones than those identified by operator 2, in many cases, the additional items found by operator 1 were not AMD. Giving operator 1 the option of using UAV pictures in addition to the orthomosaic only, in fact, allowed him to exclude about 70 items that were noted using solely the orthomosaic. UAV pictures provided a view of items from different angles, helping the operator in deciding whether the item was in fact litter. However, under these conditions, the mapping task became more time-consuming.

Comparison in Terms of Area
Focusing on the number of items identified does not take into account the size of each object. Here, we show the litter recognized by the operators with respect to the total collected in terms of area. Moreover, we compare this with the simple count of the same objects (2019 dataset, operator 1). Figure 12 shows that accounting for the surface actually covered by each item on the orthomosaics increases the detection of the total AMD up to more than 80% of the total collected. Larger items, in fact, are easier to identify than very small ones and contribute significantly to the overall litter amount, especially if considering them as a potential source of MP.

Object Identification: Comparisons
Macroplastic identification and classification provided different accuracy levels in recognizing the objects on the basis of available information. The consequence, as shown above, was to assign a certain CL during the operation of litter description. However, it is particularly interesting to show in which way an item can be recognized or not using the different UAV-derived products. Figure 13 shows (i) a picture of detail, (ii) the orthomosaic view, (iii) the DSM, and (iv) the NDVI. Here, we present a brief example of items that can usually be found on beaches, showing how they are recognizable using UAV image-based datasets: Most of the AMD located on the beach presented situations similar to those depicted in Figure 13.

Discussion
This article presents an approach based on UAVs for mapping and classifying the plastic AMD along a beach. The survey was repeated after one year, using drone technology to fly over the same transects. AMD over 2.5 cm was collected every time and further analyzed to create a ground truth dataset. Using UAV images captured at a distance of 10 to 25 m from the ground surface, we generated orthomosaics, DSMs, and NDVI maps in order to develop a comprehensive AMD map. To better understand the way expertise might influence accuracy with identification, we used operators with varying levels of experience in recognizing AMD. Expertise in marine debris research and identification proved an important difference in determining operator accuracy. operator 1's relative inexpertise was supplemented by UAV-acquired photographs, which improved accuracy.
Assigning a level of confidence to the classification of items proved to be a crucial aspect in monitoring the degree of mapping accuracy. In fact, less confident classifications were often preferred by mapping operators due to the difficulty in distinguishing AMD from the background texture and recognizing it properly. Lo et al. [40] demonstrated that operating conditions also affect the accuracy in litter assessment in a significant way, with an accuracy ranging from 39% to 75%. The validation of the descriptions of recognized AMD items proved that the operator's attribution of higher CLs was actually matched by a proper identification of AMD. Usually, CL 1 was assigned to items with a well-distinguishable shape. Moreover, the concordance [44] between operators was maximum for CL 1. Larger items such as plastic bottles proved easy to identify on the orthomosaic and the DSM. This is because they cover a large number of pixels, and they also deviate considerably from the terrain topography. White pieces of plastic (e.g., polystyrene) can be distinguished both using the orthomosaic and multispectral information (i.e., NDVI). However, recognizing them on the DSM may be difficult due to their small size and more rounded shape. Stoppers or bottle caps can be distinguished due to their circular shape, but the actual size results in the difference for success in recognizing them. Moreover, the visible side of the stopper is important. In fact, the accumulation of sand accumulated within the stopper can negatively affect the identification process. Flat objects, such as small pieces of plastics, are difficult to distinguish; their color may help in recognizing them on the orthomosaic, though such colors often fade over time and with exposure to the natural elements. Finally, transparent items (e.g., films, bags) are really hard to identify using all of the information acquired. They are often flat, transparent, and can also be partially covered by the sand, all of which make them less distinguishable.
Conversely, recognizing very small items is a very hard task: the spatial resolution on the ground should be increased, but generally this requires flying lower, limiting the extent of any flight mission. While microplastics do not contribute significantly to the total amount of AMD in terms of weight, they are an important component of the plastic pollution problem. The value of this approach is based on the mapping and quantification of larger items, representing a solution for a quick mapping of beach AMD.
Multispectral data in the near-infrared wavelengths allowed operators to exclude vegetation-related items such as flowers or stems on the dune but did not increase the amount of AMD recognized in a significant way. On the contrary, looking at the results obtained by the more expert operator 2 in 2020, Figure 10 highlights a decrease of both the total count of AMD in each CL class (CL 3: from 29 to 25; CL 2: from 29 to 25; CL 1: from 16 to 5) and an increase of the uncertainty in classifying the litter (CL 3: from 39% to 45%; CL 2: from 39% to 46%; CL 1: from 22% to 9%). The spectral response of litter material has been investigated in controlled laboratory environments [50,51] and leads to the characterization of the plastic [52,53], but the same response in the field [54] can be different due to degradation processes that modify material characteristics. The collection of proximal hyperspectral measurements of in situ plastic debris [55] can represent a way for developing methods to address this issue. Additionally, on the basis of specific wavelengths in which the spectral response is higher than the environmental surroundings, and hence the presence of plastics is characterized by better exploiting this, new spectral indices can be formulated and investigated compared to the simple NDVI. Similarly, the use of DSM in the support of mapping operations did not increase the amount of AMD recognized in a significant way. Although the use of DSMs can allow operators to recognize item shape, even in a semi-automatic way [56], limitations are also represented by natural wood debris, algae agglomeration, and similar items that would result in beach surface anomalies, misleading marine litter detection. Conversely, the use of DSMs allows the quantification of litter volume [57].
The comparison between the mapping performed using DJI-P4RTK and DJI-P3P image datasets showed that, in most cases, using the DJI-P3P, the confidence level assigned to litter items increased. This is quite obviously due to the higher spatial resolution of DJI-P3P images at a 10 m flight altitude (i.e., GSD DJI−P3P = 4 mm < GSD DJI−P4RTK = 7 mm). However, the use of an RTK UAV has certainly the notable advantage of speeding up all the aerial survey operations in the field since there is no need to deploy GCPs and collect their coordinates with a GNSS receiver. Current app limitations may be resolved in the future, allowing one to exploit the higher camera resolution of the DJI-P4RTK (i.e., camera resolution DJI−P4RTK = 20 Mpix > camera resolution DJI−P3P = 12 Mpix).
The precise positioning of the AMD, on the beach or in wider coastal stretches, also represents important data, especially whenever combined with ground surveys (collected items) because such information is useful to better understand their transport, distribution, and accumulation under coastal processes forcing the overall dynamic system [38,58,59]. Moreover, the spatial distribution of AMD mapped using the UAV approach can be further used to calculate beach litter density or CCI to classify beaches on the basis of cleanliness. The UAV-based AMD mapping can help in cataloging and better understanding debris that might exist in hard to access places. Finally, we believe that a UAV-designed list of litter items, as the one proposed by Andriolo et al. [44], should be refined and assumed as a standard for UAV-based mapping and monitoring of AMD in the future, both to help users with poor experience in litter monitoring (as in this work) and to make UAV survey output uniform.

Conclusions
The mapping of AMD from UAVs is strongly influenced by the relationship between the ground sample distance and the size of litter. In this work, we found that a resolution of about 200 pix/m is a good solution to map plastic debris. That said, the characteristics and the position of items on the ground affect the detection of even large and theoretically easy-to-identify objects. Moreover, the level of experience of the mapping operator is a further variable that can impact the results of litter identification. The use of multispectral information can help to distinguish vegetation from AMD but was found to be substantially unhelpful in distinguishing AMD from the sandy background. There are inherent tradeoffs in the choices made using this methodology. Focusing on large items can yield accurate results while flying over wide areas. Focusing on smaller areas at closer range can allow better quantifying debris of various sizes. Due to the operator's uncertainty in recognizing the AMD properly on the orthomosaics, operator training should be required in order to improve their confidence with UAV-based AMD mapping.
Based on this analysis, we recommend (i) to use RGB cameras with the highest geometric resolution as possible in order to have smaller ground sample distances, (ii) to use RTK drones if a centimeter-level accuracy of debris position has to be achieved, and (iii) not to use operators with no experience in matters of beach litter mapping when using drone technology to identify anthropogenic marine debris.
Future research should explore the use of hyperspectral sensors [60,61] instead of multispectral ones in order to exploit ways that some plastics respond to specific wavelengths as well as the use of algorithms of object detection, machine learning [11,[62][63][64][65][66][67], or a deep-feature-based approach [68] coupled with this type of imaging. However, these tasks can be complicated by the degradation processes which modify the characteristics of materials, plastics included. The spectral response in situ can be different from what is obtained in the laboratory.
Funding: This study was supported by the NET4mPLASTIC project, co-financed by the European Regional Development Fund within the framework of European cross-border territorial cooperation Interreg. IT-HR. CUP: F76C19000000007.

Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.

Data Availability Statement: Not applicable.
Acknowledgments: The authors would like to thank students and technical staff involved in support of operations in the field.

Conflicts of Interest:
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations
The following abbreviations are used in this manuscript: