Enhanced Detection and Classification of Beach-Stranded Macrodebris: A Pilot Study Using Polarimetric Imagery Free

Beaches are demonstrated sinks for marine debris, particularly in the backshore (Andriolo et al., 2020; Brennan, Wilcox, and Hardesty, 2018; Collins and Hermes, 2019; Merlino et al., 2020; Olivelli, Hardesty, and Wilcox, 2020; Onink et al., 2021; Rech et al., 2014; Roman et al., 2020; Ryan, 2020; van Sebille et al., 2020; Willis et al., 2017). Monitoring of beaches is perhaps the most practical and efficient way to evaluate the status and trend of marine debris. In fact, beaches are the environmental compartment having the most (and consequently, the most robust) information on marine debris abundance (GESAMP, 2019) given ease of access and the relatively nontechnical nature of typical beach survey methods (visual identification and collection while walking). Several well-established marine debris beach monitoring methodologies across the globe rely on visual assessment methods performed within replicate transects perpendicular to a specified length of beach (United States: Burgess et al., 2021; Lippiatt, Opfer, and Arthur, 2013; the European Union: Hanke et al., 2013; OSPAR Commission, 2010; the Republic of Korea: Hong et al., 2014; Jang et al., 2014; UNEP, 2020; UNEP: Cheshire et al., 2009; Australia: Schuyler et al., 2020). These programs target larger debris size classes with minimum size ranges of 5 mm (mesodebris) or 25 mm (macrodebris), as these items are more easily identified and connected to production processes and source, allowing for the development of targeted prevention and policy measures (Ryan et al., 2020a,b; Uhrin et al., 2022). However, these on-the-ground methods require substantial investments in time and labor, which limit their scalability.

For this reason, uncrewed aircraft systems (UAS) have been tested in recent years as a tool for augmenting on-the-ground shoreline marine debris detection and monitoring (see review by Gonçalves et al., 2022; see also Andriolo et al., 2024; Bekova and Prodanov, 2023; Corbau et al., 2023). UAS can surveil on-the-ground survey areas more rapidly or provide greater spatial coverage than traditional visual assessment methods on foot (Martin et al., 2018). UAS can also reach areas difficult to access from the ground, expanding the extent of spatial survey designs (Martin et al., 2018; Merlino et al., 2020; Song et al., 2022). As UAS can provide continuous spatial coverage of an entire site, this precludes the need for a randomized subsampling strategy, which may be necessary when conducting on-the-ground monitoring (Martin et al., 2018).

To date, UAS for marine debris detection are most commonly equipped with conventional cameras having three spectral bands (red, green, and blue, typically abbreviated RGB) (Andriolo et al., 2024; Bekova and Prodanov, 2023; Corbau et al., 2023; Gonçalves et al., 2022). RGB imagery has proven to be useful for identifying and counting debris objects when there is strong contrast between the object and its background (e.g., dark debris stranded on a sandy beach). However, because RGB imagery is sensitive to illumination, exhibits high band correlation, and is perceptually nonuniform, there are some drawbacks in its utility for object detection (Li and Yuen, 2002; Wang, Wang, and Bu, 2011; Yang, Liu, and Zhang, 2010). Translucent objects and small objects (<10 cm in size) are more difficult to detect, and detection is often driven by the altitude at which the imagery is collected (Lo et al., 2020; Martin et al., 2018). Additionally, partially buried objects, as well as objects in shadow and objects stranded on non-homogeneous backgrounds (e.g., presence of gravel, pebbles, shells), are also more difficult to detect due to the objects’ physical features (e.g., edges) being obscured (Martin et al., 2018). To overcome these challenges, investigation and utilization of emerging image sensor technologies such as polarimetric imaging may lead to improved marine debris detection via UAS.

Polarization is a fundamental property of light waves that describes the direction of their electric field oscillations. Generally, direct sunlight is unpolarized, meaning that the direction of electric field oscillations is random in time. Sunlight that has reflected from an object’s surface can become partially or completely linearly polarized, meaning the electric field oscillations occur within a specific plane. The polarization state depends on surface characteristics, with human-made objects often having distinctive polarization characteristics relative to natural features (Bartlett et al., 2011; Gupta et al., 2001; Islam, Tahtali, and Pickering, 2019). When combined with the spectral bands obtained from conventional optical sensors (e.g., multispectral or hyperspectral cameras) or nonoptical sensors such as synthetic aperture radar, polarimetric images have shown promise in several remote-sensing applications ranging from estimation of atmospheric aerosol densities to land-cover classification, exploration of geologic landforms, red tide detection, and oil spill detection (Snik et al., 2014; Zhao et al., 2016).

Polarimetric cameras typically use polarizing filters oriented at four different angles overlaid on the camera chip to capture light intensity corresponding to each angle of polarization. Using the recorded data, the polarization state of the received light can be described mathematically (Yan et al., 2020; Zappa et al., 2008), with the output of the computations stored as image bands. Polarimetric images are known to enhance physical characteristics of objects in images such as surface texture, shape, and edge detail, thus improving object discrimination (Kupinski et al., 2019; McCormick, Nascimento, and Hendricks, 2018). Although detection of floating debris was beyond the scope of this study, some previous studies have indicated that polarized light and polarimetric imaging may also be useful in the detection of surface-floating plastic pollution (Garaba and Harmel, 2022; Goddijn-Murphy et al., 2024). Toward this aim, a few laboratory studies have described inherent optical properties of microplastics with encouraging results (Koestner, Foster, and El-Habashi, 2023; Koestner et al., 2024; Valentino et al., 2022; Yu et al., 2021).

This study evaluated the suitability and practicality of using a commercially available and low-cost RGB-polarimetric camera for imaging beach-stranded macrodebris (≥2.5 cm in size) and assessed the technical feasibility. Mathematically derived polarimetric bands, when used in combination with spectral bands obtained from standard RGB images, were investigated for their potential to improve the ability to visually identify beach-stranded macrodebris, especially for items obscured in shadow or partially buried items. Polarimetric bands were also evaluated for their ability to augment semi-automatic classification of debris.

Specifically, this study asked:

1. Does the electronic display (computer screen) of polarimetric bands improve visual (naked-eye) recognition and identification of debris objects over the display of RGB spectral bands?

2. Are polarimetric bands correlated with RGB spectral bands?

3. Does the spectral separability of debris improve when polarimetric bands are layer-stacked with RGB spectral bands (over RGB alone)?

4. Does semi-automated debris object classification accuracy improve when polarimetric bands are layer-stacked with RGB spectral bands (over RGB alone)?

Answers to these questions are anticipated to assist the marine debris community in evaluating the utility of polarimetric imaging for monitoring programs. In turn, these programs play a crucial role in providing question-driven data for guiding management decisions about specific ecological conditions or populations of interest (Field et al., 2007; Jones et al., 2013; Legg and Nagy, 2006; Lindenmayer and Likens, 2010). Monitoring data can inform management considerations regarding the condition/population of interest such as determining what processes and drivers influence abundance and distribution, detecting changes that need to be addressed through management interventions, assessing whether management interventions were implemented properly and were effective in reaching the stated goals, and identifying unintended consequences of the intervention (DeLuca et al., 2010; Eyre et al., 2011; Hutto and Belote, 2013; Nichols and Williams, 2006). In the United States, marine debris shoreline monitoring data have been used to estimate marine debris abundance and temporal trends (Blickley, Currie, and Kaufman, 2016; Ribic, Sheavly, and Rugg, 2011; Ribic et al., 2010, 2012; Uhrin et al., 2020; Wessel et al., 2019), assess disaster-generated debris (Murray, Maximenko, and Lippiatt, 2018), document the effectiveness of legislation aimed at reducing marine debris, including plastic bag bans (Blickley, Currie, and Kaufman, 2016), tobacco use bans (Currie and Stack, 2021), and container deposit laws (Schuyler et al., 2018), identify and support targeted marine debris prevention initiatives in San Francisco Bay (shotgun wads) and coastal Virginia (balloons) (Bimrose et al., 2020; Trapani, O’Hara, and Register, 2018), and document rates and causes of entanglement for endangered species (Allyn and Scordino, 2020). If this emerging technology can improve the effectiveness of marine debris shoreline monitoring, it has the potential to support a range of management and policymaking needs.

The methodology employed in this research included collection of images of beach-stranded macrodebris using an RGB-polarimetric camera secured to an extendable pole to simulate low-level UAS flight. Five polarimetric bands were mathematically derived for each image. Eight-band composite images comprised of RGB bands and the five calculated polarimetric bands were generated: S₀, S₁, S₂, angle of linear polarization (AoLP), and degree of linear polarization (DoLP). Band correlation analysis was conducted to evaluate the correlation between the RGB and polarimetric bands. Stranded macrodebris items in each image were evaluated, and it was determined if item contrast and hence identification and classification were augmented in the eight-band composite images vs. images with the RGB bands alone. Finally, a supervised classification was performed using both the RGB imagery alone and the RGB imagery enhanced with the additional parametric bands to investigate whether the addition of the polarimetric bands improved classification accuracy. The study workflow is depicted in Figure 1, and the individual steps are described below.

The camera investigated in this work is the FLIR Blackfly S USB3 RGB-polarimetric camera (Figure 2). This 5.0 megapixel (MP) polarimetric camera is based on the Sony IMX250MYR camera chip and uses four linear polarizers oriented at four different angles (0°, 45°, 90°, and 135°), allowing for simultaneous capture of five images at each exposure station: four polarization images of the scene, denoted I₀, I₄₅, I₉₀, and I₁₃₅, corresponding to the angles noted above, and a standard, three-band RGB image. The camera’s size (<100 g, without the lens) and power requirements are suitable for installation on a UAS. The FLIR Blackfly S also included a software development kit that allowed for developing custom software for working with the collected imagery. A Fujinon C-mount lens with 12.5 mm focal length was used on the FLIR Blackfly S camera.

Although the FLIR Blackfly S camera used in this study is commercially available and lightweight, it lacks fully autonomous operation capabilities and requires a constant hardwired connection to a laptop to manually perform exposure setting adjustments (ISO and shutter speed) and save images. For this reason, the camera was secured to an extendable pole and connected to a Dell Latitude 5414 Rugged laptop running Windows 10 for live parameter adjustments using the FLIR SpinView v2.0.0.146 graphical user interface (Teledyne Inc.) software. The connection from the camera to the laptop consisted of a 5 m USB3.0 cable. Data collection was a two-person operation, with one person holding the pole-mounted camera and the other person monitoring the camera settings on the laptop (Figure 3).

Preliminary testing to select the optimal FLIR Blackfly S camera and software settings as well as appropriate altitude above ground level (AGL) for image collection was performed on a simulated beach (∼140 m²) constructed at the O.H. Hinsdale Wave Research Laboratory at Oregon State University (OSU) using sandy substrate sourced from the Oregon coast and having direct exposure to sunlight in 2021 (Figure 4). Due to limited availability of actual shoreline debris at this location, the simulated beach was seeded with household debris items known to be present in shoreline marine debris distributions, representing different material types, shapes, sizes, and colors. Debris items were manually scattered by the project team throughout the site in a pseudo-random arrangement. The pole-mounted camera was tested at a 3 m altitude AGL, with the imagery nominally vertical (optical axis within ±15° of nadir) using a weighted stand. The camera exposure settings were adjusted empirically by inspecting the imagery. The lens f-stop setting was held fixed at f/5.6, and the exposure was adjusted through the ISO and shutter speed settings via slider bars provided in the graphical user interface.

Further refinement of the operational parameters was performed at Neptune State Scenic Viewpoint on the coast of Oregon (Lane County) in 2021. The site consists of an extensive basalt rock beach interspersed with sand (Lund, 1971). The site also contains an ∼6-m-wide strip of cobble, backed by a steep cliff. The site was found to be lacking in macrodebris, which is consistent with the general lack of debris for this area indicated in the National Oceanic and Atmospheric Administration (NOAA) Marine Debris Monitoring and Assessment Project (MDMAP) database (Burgess et al., 2021), as well as the relatively low population density of this portion of the Oregon coast and the level of monitoring and debris removal activities in this region (NOAA, 2019). Thus, the Oregon Parks and Recreation Department generously provided actual macrodebris objects that had accumulated over time both at Neptune and other sites along the Oregon coast, including a crab trap, a wood pallet, various ropes, polystyrene buoys, glass jars, pieces of vehicle tires, miscellaneous plastic objects (e.g., laundry detergent containers, water bottles, milk jugs) and plastic fragments, and rubber fragments. This debris was transported by the OSU research team and hand-seeded on the beach at Neptune for image acquisition. Debris items were manually arranged to achieve approximately uniform spatial density over the site and to create a range of contrast between the debris items and underlying substrates (e.g., sand, cobble, or rock). The hand-carried pole-mounted camera was tested at 3.5 m AGL, acquiring near-vertical imagery (optical axis within ±15° of nadir). The camera’s exposure settings (shutter speed and ISO) were initially set to match those determined to be optimal from the simulated beach trials and were then fine-tuned in response to periodic examinations of the collected imagery approximately every 15 min.

In December 2022, the pole-mount system was used to acquire imagery at two experimental debris fields located on San Jose Island, Texas. San Jose is a sandy barrier island oriented generally NE-SW, with the Gulf on the east and south, Aransas Bay on the west, Mesquite Bay on the north, and Aransas Pass to the south (Figure 5). The site consisted of sandy, low-slope beaches of ∼40-m width, backed by vegetated dunes. A variety of existing macrodebris (≥2.5 cm) was observed to be distributed across the width of the beach from the vegetation line to the shoreline (Figure 6). The Port Aransas north jetty was located ∼1 km south of the site. Most of the debris items encountered had been significantly weathered, making them excellent candidates for evaluating operational capabilities under the challenging types of conditions encountered in marine debris surveys.

Existing debris was collected from a 1 km² area on the beach beginning at Fisherman’s Wharf Pier (27°50′21″ N, 97°02′43″ W) and used to generate two debris fields, each containing a variety of material types and density distributions (i.e. dense and sparse debris fields). The first site, San Jose 1 (SJ1), was a 6 × 2.5 m scene that was densely seeded with seven very common debris classes that were present throughout the surrounding areas, although in lower spatial densities. This produced a debris field with a robust sample size for rigorous analysis. This scene was used to help determine whether the inclusion of polarimetric bands improved classification of these common materials. The second site, San Jose 2 (SJ2), was a larger 38 × 8 m scene selected for its many large and naturally occurring debris items, which were again consolidated from the local area into the rectangular site. This scene was selected to collect high-quality polarimetric data of the native debris items at lower densities than SJ1. The types and quantities of debris items consolidated from within the vicinity varied between SJ1 and SJ2; the debris classes and item counts within each are shown in Table 1.

The polarimetric imagery was acquired through the two-person operation shown in Figure 3, with one person holding the camera pole and the other person monitoring the SpinView software and adjusting the exposure settings, when necessary. To maintain constant viewing geometry (altitude and azimuth), the pole was kept at the same height, and the camera was kept pointing in the same direction during the imagery acquisition. In total, 80 single-frame images were collected using the pole-mounted polarimetric imaging (PI) camera system: 8 from SJ1 and 72 from SJ2.

In addition to the imagery collected with the PI camera, RGB imagery was collected from a Skydio 2 (Skydio, Inc.) UAS over the debris fields to construct accurately georeferenced orthomosaics to serve as GIS base layers and provide spatial context for the evaluation of the PI imagery (Figure 7). The Skydio 2 is a commercially available quadcopter equipped with a 12.3 MP (4056 × 3040 pixel) RGB camera with a 3.7 mm focal length, f/2.8 camera lens. The imagery was acquired from an AGL flying height of 60 m, yielding a 2 cm ground sample distance. An orthomosaic for SJ2 generated from the Skydio 2 imagery is shown in Figure 8.

Nylon photo panels (black and white checkerboard pattern) were also deployed within the debris fields that were visible in both the UAS and PI imagery. These panels, which can be seen in the corners and center of the SJ2 debris field in Figure 8, were used for georeferencing the polarimetric camera imagery and the UAS imagery, such that the dimensions of the objects could be accurately measured, and the polarimetric camera imagery could be overlaid on the UAS imagery within a GIS. The photo panels were surveyed with a Leica Viva GS14 global navigation satellite system (GNSS) receiver with occupation times of at least 15 min. The observation files were subsequently uploaded to the National Geodetic Survey (NGS) Online Positioning User Service (OPUS) to obtain coordinates tied to the National Spatial Reference System (NSRS).

Although humans have been shown to have some limited perception of polarized light under controlled conditions, true polarization vision is lacking (Misson, Timmerman, and Bryanston-Cross, 2015; Temple et al., 2015). As such, several related color mapping strategies for polarization imaging have been developed, with the idea that a single strategy is unlikely to be appropriate for every use of polarimetric imaging (Kruse et al., 2022). The most common method, HSV, maps AoLP into color hue (H), DoLP into color saturation (S), and intensity of unpolarized light, S₀, into value (V) (Wolff et al., 1992). Greater saturation (x axis) corresponds to increasing DoLP, while hue (y axis) represents the AoLP at any given pixel. Because S₀ was stored as a separate band, the modified version of HSV of Wilkie and Weidlich (2010) was followed, wherein S₀ is ignored, and DoLP and AoLP are plotted as a bivariate image describing the polarization state in terms of hue and saturation. The widely used spectrum approximation scheme was applied in which the following color sequence is linearly interpolated: red + blue, blue, blue + green, green, green + red, red. Bivariate plots were created for each of the 80 AoLP and DoLP layers. Each of the 80 single image frames was then visually evaluated by a single author (F.K.H.). The original RGB image with its corresponding bivariate color map was displayed side-by-side on a computer monitor, and observations were made with respect to how the appearance of objects in the bivariate color map differed from the RGB image, focusing on elements of contrast (Figure 9). Specifically, improvements in object edge detail were noted, paying close attention to transparent objects, objects in shadow, or partially buried objects. Enhancements in object surface texture and dimension were also documented, as were improvements in object identification against heterogeneous backgrounds (e.g., rocky substrate).

The calculated bands from polarimetric images were evaluated to determine whether they provided unique information when compared to RGB spectral bands using correlation matrices. Here, a correlation matrix takes each pair of image bands and computes the difference in digital number (pixel value) at each pixel location, from which the correlation coefficient between the bands can be computed. The correlation coefficient represents the relatedness of two bands and how well the value of one can be used to predict the value of the other. Low correlation between bands is desired, as it indicates that the bands contain different information, such that their combination may improve classification over either band individually. MATLAB version R2021a (Mathworks, Inc.) was used to compute the correlation coefficients between every band in the SJ1 and SJ2 orthomosaics.

To assess classification accuracy, the simple and effective k-nearest neighbors (KNN) supervised classification algorithm (Silverman and Jones, 1989) was chosen. This classifier was run on both the three-band RGB orthomosaics and the eight-band RGB + PI orthomosaics to determine if the PI-derived bands contributed additional information that improved classification accuracy over RGB alone. The SJ1 analysis focused on dense detection of common debris classes such as rubber, glass, wood, buoy/foam, aluminum, and substrate. Objects in each class were first manually labeled by the analyst (Figure 10), such that each debris class contained, at minimum, 12–30 objects. The labeled objects were then split into training or testing subsets to ensure each subset contained a proportional number of objects based on multiple criteria such as type, weathering, color, and translucence. The supervised classification was then performed in MultiSpec with the parameter k, which represents the number of nearest neighbors to consider (Biehl and Landgrebe, 2002), set to the default value of 5. The run time averaged 50 hours per scene on a Windows 10 Enterprise workstation.

Thematic accuracy of the debris classes was evaluated using class-by-class error matrices (Story and Congalton, 1986). Training and test accuracies (per class), training and test accuracies (overall), as well as the kappa statistic were computed. Overall thematic accuracy (proportion of pixels that were correctly classified) was calculated by dividing the total number of correctly classified pixels (i.e. sum of the diagonal) by the total number of reference pixels. Individual class accuracy was assessed using producer’s and user’s accuracies. Producer’s accuracy, also called omission error, refers to the probability of a reference pixel being correctly classified and is calculated by dividing the number of correctly classified pixels in each class (main diagonal of the error matrix) by the total number of reference pixels “known” to be of that class (column total). User’s accuracy, also called commission error, indicates the probability that a pixel classified into a given category represents that category on the ground and is calculated by dividing the number of correctly classified pixels in each class by the total number of pixels that were assigned to that category by the classifier (row total) (Congalton, 1991).

The κ statistic is a measure of how well a classification agrees with reference data and represents the proportion between actual agreement and the agreement expected by chance (Cohen, 1960). Values for κ range between −1.0 and 1.0, but positive values are expected because the classification and the reference data are positively correlated. Higher κ coefficients indicate a more accurate classification. As a (somewhat application-specific) rule of thumb, strong agreement between the classifier and the reference data is indicated by κ >0.80, while moderate agreement is indicated by κ in the range 0.40 ≤ κ ≤ 0.80 (Landis and Koch, 1977).

The results of the subjective visual assessment of polarimetric images when compared to RGB images alone described the observed differences in object contrast and edge detail. The results also compared the mathematically derived polarimetric bands to RGB bands using correlation analysis, separability analysis, and accuracy assessment from the KNN classification to document relationships among paired image bands and assess differences in object class separability and classification accuracy among the bands. These analyses provided the basis for a detailed examination of the utility of mathematically derived polarimetric bands for augmenting stranded macrodebris detection, identification, and classification on sandy beaches (“Discussion” section).

Aspects of the physical characteristics of debris objects as viewed in bivariate plots of calculated polarimetric bands (AoLP × DoLP) compared to RGB images of the same scene were found to be markedly enhanced when visually inspected on a computer monitor, allowing for greater contrast against the shoreline and more confident identification (Figure 9). Most notably, edge detail was enhanced for objects found partially buried or in shadow (e.g., plastic crate, glass bottles, wire trap) (Figure 9A,C,D). Elements of surface texture (e.g., ridging, mesh) were also more nuanced in the bivariate plots (Figure 9E). Some objects that appeared to be circular and flat in the RGB images were in fact identified as being three dimensional (e.g., entire bucket vs. just a lid; funnel vs. generic circular object; Figure 9E,F). Lastly, objects against a heterogeneous background (e.g., rocky substrate) were more easily identified (e.g., funnel, vehicle tire; Figure 9F).

The RGB and S₀ bands were highly correlated with one another in both orthomosaics, with correlation coefficients between band pairs ranging from 0.85 to 0.97 (Tables 2 and 3). From this, it is inferred that these bands contained much of the same information as one another and that the use of all of them together may not substantially improve debris classification. Given the fact that these four bands all track the intensity of reflected light, it is to be expected that their values would tend to increase and decrease as a group. The remaining polarimetric bands (S₁, S₂, DoLP, and AoLP) had low correlation with the RGB bands and relatively low correlation with one another (Tables 2 and 3). While this correlation analysis alone is insufficient to show that the PI-derived bands improved debris classification (hence the need for the two following steps, separability analysis and classification accuracy assessment), it is an informative first step: Specifically, it suggests that the inclusion of these bands may improve debris classification because they are providing unique information.

Images composed of RGB-only bands consistently had the lowest transformed divergence, with mean values of 1081 and 1213 for SJ1 and SJ2, respectively (Tables 3 and 4). TD values substantially increased when polarimetric bands were added to images: The eight-band (three RGB bands + S₀ + S₁ + S₂ + AoLP + DoLP) SJ1 and SJ2 images had mean TD values of 1531 and 1748, respectively, corresponding to increases of 41.6% and 44.1% (Tables 4 and 5).

The inclusion of the polarimetric bands stacked with the RGB bands led to improvements in overall accuracies as well as kappa statistics for both test and training data in the two orthomosaics (Tables 4 and 5). However, improvements varied across the debris classes and between the two orthomosaics. Classification of wood, buoys/foam, aluminum, and substrate improved when polarimetric bands were included in the SJ1 orthomosaic, while the classification of rubber and glass did not improve (Table 4). In some instances, producer’s accuracy was higher than user’s accuracy (e.g., rubber, glass, wood) in the SJ1 orthomosaic (Table 4). For example, in the eight-band image for SJ1, even though 89.6% of the reference rubber items were correctly identified as rubber (producer’s accuracy), only 86.6% of the pixels identified as rubber in the classification were actually rubber (user’s accuracy; Table 4). Conversely, user’s accuracy was higher for glass, buoys/foam, and aluminum (Table 4). For example, 85.2% of buoys/foam identified in the classification were buoys/foam despite only 76.2% of the reference buoys/foam items being correctly identified as buoys/foam in the classification (Table 4). Fewer debris classes were assessed for the SJ2 orthomosaic (Table 5). Here, the plastic class performed similarly between the RGB-only and the eight-band images, while wood, rope, and substrate class performance improved (Table 5). Producer’s and user’s accuracies were similar for most debris classes in the eight-band image, except for rope, where user’s accuracy was higher (Table 5).

This study provided strong indication that polarimetric imaging is a useful asset in detection and classification of shoreline stranded marine debris objects across seven material classes. Overall, the polarimetric information displayed visually rich object characteristics (i.e. edge detail, surface texture, dimensionality) for several macrodebris items, which improved their visual identification. This capability could enhance the effectiveness of UAS flights for which the mission is to observe stranded debris in hard-to-reach areas and areas where debris may be stranded against complex backgrounds (e.g., presence of gravel, pebbles, shells, vegetation). There is evidence that backshore areas accumulate greater amounts of larger marine debris items vs. the beach face, particularly when vegetation is present and can trap debris (Andriolo et al., 2020; Brennan, Wilcox, and Hardesty, 2018; Merlino et al., 2020; Olivelli, Hardesty, and Wilcox, 2020). However, Burgess et al. (2024) reported lower detection rates (visual observation) for macrodebris within backshore vegetation vs. nonvegetated shore zones, which they attributed to a lack of contrast between debris items and vegetation and visual obstruction (debris items were “hidden” behind other features within the vegetation). Thus, the integration of UAS-collected polarimetric imagery with established shoreline monitoring programs based on visual observation may improve shoreline debris estimates and help inform coastal management objectives for addressing marine debris. Monitoring data on shoreline marine debris loads can support the prioritization and evaluation of preventative actions and policies. When data are collected over time, they allow for “before and after” intervention comparisons (Blickley, Currie, and Kaufman, 2016; Harris et al., 2020; Uhrin et al., 2020), and when collected over space, they can allow for comparisons under different policy or management regimes (Schuyler et al., 2018).

These improved classification results using RGB images stacked with bands derived from polarimetric images are in line with other studies that employed techniques such as alternative color spaces (e.g., histogram of gradients) to augment RGB images (o, b, and c; Martin et al., 2018; Pinto, Andriolo, and Gonçalves, 2021). Here, the overall accuracy of the KNN classifier was improved by 6.6 percentage points and 25.4 percentage points in the two scenes when RGB images were stacked with polarimetric bands. However, improvements in accuracy varied across debris classes, similar to other works where semi-automated classifiers were used (Martin et al., 2018; Pfeiffer et al., 2023; Winans et al., 2023). This could be attributed to differing surface characteristics (e.g., roughness) and the variable extent to which light reflected from those surfaces is linearly polarized.

Three technological limitations were noted during this study, which may be addressed through the ongoing maturation of polarimetric cameras and processing software. The first is that substantial preprocessing was required, including the development of custom MATLAB scripts, to generate the Stokes parameters and AoLP and DoLP bands. The second is that the polarimetric camera used in this study had lower spatial resolution than similarly sized, commercially available RGB cameras. In practice, this requires a lower acquisition altitude to achieve a desired ground sample distance, which then impacts the number of images required to complete a survey. The third technical limitation is that operation of the polarimetric camera tested in this study was insufficiently automated to facilitate image acquisition from a UAS. As the commercial market for polarimetric imaging cameras continues to grow and the sensors and processing software become more operationally mature, it is anticipated that all three of these limitations will be mitigated.

Of the technological limitations noted above, one that merits further discussion is the insufficient automation of the image acquisition process to enable the polarimetric camera to be operated from a UAS. Because UAS can provide more spatial coverage, cover the same area more quickly, and reach areas difficult to access compared to on-the-ground surveys (Martin et al., 2018), integrating polarimetric cameras with UAS would be highly advantageous. The primary reason the hand-carried, pole-mounted camera was used in this study was that the image acquisition required constant human interaction to adjust exposure settings. To test whether the polarimetric camera could be operated from a UAS, in a follow-on study, the FLIR Blackfly S USB3 RGB camera was integrated onto a commercially available UAS and used to acquire imagery over a portion of the Oregon State University campus (Parrish et al., 2023). However, integration with the UAS required considerable custom development of hardware and software, and a relatively rudimentary form of auto-exposure was utilized (Parrish et al., 2023). Hence, automated image acquisition, including robust auto-exposure, is a highly recommended topic for future work.

In related work reported in Herrera (2022), the project team performed qualitative, visual assessment of polarimetric images collected from a helicopter over vegetated dunes containing debris and observed encouraging results. However, no corresponding field data were collected that would enable the types of quantitative assessment performed in this study. Hence, another recommendation for future work is to repeat the experiments conducted in this study at sites containing vegetation.

This study is the first to examine and provide evidence that polarimetric imagery, when band-stacked with RGB imagery, can augment detection of beach-stranded macrodebris. While beyond the scope of this study, the application of polarimetry to enhanced detection and identification of microplastics is also of interest and has begun to be investigated in other work (e.g., Koestner, Foster, and El-Habashi, 2023; Koestner et al., 2024; Valentino et al., 2022; Yu et al., 2021). Unlike microplastics, macrodebris is more easily identified on the beach and can be connected to production processes and source, allowing for the development of targeted prevention and policy measures. Thus, continued development of technology that supports capabilities for rapid and accurate detection and classification of these larger debris size classes found stranded on beaches remains a worthwhile endeavor.

Several studies have examined UAS-collected imagery for use in supplementing visual surveys of shoreline marine debris, focusing on conventional intensity cameras (i.e. RGB images). Because the RGB color space is sensitive to illumination, exhibits high band correlation, and is perceptually nonuniform, there are drawbacks to its utility for object detection, particularly when there is low contrast between the object and its background. Polarimetric imagery is very effective in low-light conditions on land, when objects on land are located within a noisy/cluttered background, or when there is low contrast, which may be useful for marine pollution detection applications, particularly on shorelines with heterogeneous substrates.

This study found that polarimetric bands when stacked with RGB image bands improved debris identification and classification. With reference to the four questions that this study sought to address (listed in the introduction), the results showed that:

1. Electronic (computer screen) displays of polarimetric bands improved visual recognition and identification of debris objects over the display of RGB-only spectral bands. The polarimetric bands were found to enhance visual cues related to object surface texture, shape, edges, and shading.

2. With the exception of the S₀ Stokes parameter band, the polarimetric bands exhibited substantially lower correlation with the RGB bands than the RGB bands did with one another, indicating that the polarimetric bands may provide unique information that can benefit debris classification.

3. The spectral separability of debris classes improved when polarimetric bands were layer-stacked with RGB spectral bands (over RGB alone). The mean transformed divergence (TD) increased by 41.6% and 44.1% for sites SJ1 and SJ2, respectively.

4. Semi-automated debris object classification accuracy improved when polarimetric bands were layer-stacked with RGB spectral bands (over RGB alone). The overall classification accuracy improved by 6.3 percentage points for SJ1 and by 25.7 percentage points for SJ2, while the kappa statistics improved by 14.0% and 85.7% for SJ1 and SJ2, respectively.

This study provides strong indication that polarimetric imaging is a useful asset in detection and classification of marine debris found on sand shorelines. As the commercial market for polarimetric imaging cameras continues to grow and sensors become more affordable and operationally mature, it is anticipated that polarimetric imagery will become increasingly useful for shoreline marine debris detection and mapping. As this technological maturation continues, it is recommended that economic analyses be conducted in follow-on work to assess the costs and benefits of implementing polarimetric imaging and UAS technologies in large debris monitoring programs. Another recommended topic for follow-on studies is investigation of how polarimetric imaging can be optimally combined with existing, sampling-based methods and other remote-sensing technologies, possibly including polarimetric synthetic aperture radar (PolSAR). It is of particular interest to investigate whether fusion-based methods can aid in long-term debris monitoring and enhance understanding of debris accumulation and coastal dynamics. Additionally, it is recommended that the methods developed in this study be tested in coastal sites with different biogeophysical characteristics, including vegetated dunes.

This work, which stems from the M.S. thesis research of the first author (F.K.H.), was supported by the National Oceanic and Atmospheric Administration/Oceanic and Atmospheric Research (NOAA/OAR) Uncrewed Systems Research Transition Office (UxSRTO) through an award to T.B. and A.V.U. and a subaward to Oregon State University. Indirect support was received from the NOAA National Centers for Coastal Ocean Science (to T.B.), NOAA Marine Debris Program (to A.V.U.), Genwest Systems, Inc. (to P.M.), Oregon State University (to F.K.H., C.E.P.), and ORBTL AI (to W.R.W.). We are grateful for the preliminary work by C. Simpson and R. Slocum, which generated trial polarimetric imagery from the FLIR Blackfly S camera, and for their persistence in integrating the FLIR Blackfly S camera onto a UAS. We appreciated the field assistance of T. Weatherall. Operational logistics benefited from conversations with J. Berryhill, C. Cisneros, M. Starek, K. Swanson, J. Tunnel, and C. Wessel, and the editing assistance of G. Shoup. The scientific results and conclusions, as well as any views or opinions expressed herein, are those of the authors and do not necessarily reflect the views of NOAA or the Department of Commerce.

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.

The raw/processed data required to reproduce the above findings cannot be shared at this time due to technical/time limitations.