Science of the Total Environment xxx (xxxx) xxx-xxx

Contents lists available at ScienceDirect

Science of the Total Environment

F
journal homepage: www.elsevier.com

OO
Anthropogenic Marine Debris assessment with Unmanned Aerial Vehicle imagery
and deep learning: A case study along the beaches of the Republic of Maldives
L. Fallati a, 1, A. Polidori b, 1, C. Salvatore b, L. Saponari c, A. Savini a, ⁎, P. Galli a
a

PR
Department of Earth and Environmental Sciences, University of Milan-Bicocca, Milan, Italy
b
DeepTrace Technologies S.R.L., Milan, Italy
c
Department of Biotechnologies and Biosciences, University of Milan-Bicocca, Milan, Italy

ARTICLE INFO ABSTRACT

Article history: Anthropogenic Marine Debris (AMD) is one of the major environmental issues of our planet to date, and
Received 31 May 2019

ED
plastic accounts for 80% of total AMD. Beaches represent one of the main marine compartment where AMD
Received in revised form 17 July 2019 accumulates, but few and scattered regional assessments are available from literature reporting quantitative
Accepted 23 July 2019
estimation of AMD distributed on the shorelines. However, accessing information on the AMD accumulation
Available online xxx
rate on beaches, and the associated spatiotemporal oscillations, would be crucial to refining global estimation
Editor: Damia Barcelo on the dispersal mechanisms.
In our work, we address this issue by proposing an ad-hoc methodology for monitoring and automatically
Keywords: quantifying AMD, based on the combined use of a commercial Unmanned Aerial Vehicle (UAV) (equipped
Anthropogenic Marine-Debris with an RGB high-resolution camera) and a deep-learning based software (i.e.: PlasticFinder). Remote areas
CT
Unmanned Aerial Vehicles were monitored by UAV and were inspected by operators on the ground to check and to categorise all AMD
Machine learning dispersed on the beach. The high-resolution images obtained from UAV allowed to visually detect a percent-
Deep learning algorithms age of the objects on the shores higher than 87.8%, thus providing suitable images to populate training and
Maldives testing datasets, as well as gold standards to evaluate the software performance. PlasticFinder reached a Sen-
Beach sitivity of 67%, with a Positive Predictive Value of 94%, in the automatic detection of AMD, but a limitation
was found, due to reduced sunlight conditions, thus restricting to the use of the software in its present version.
We, therefore, confirmed the efficiency of commercial UAVs as tools for AMD monitoring and demonstrated
RE

- for the first time - the potential of deep learning for the automatic detection and quantification of AMD.
© 2019.

1. Introduction land-based), buoyant plastic can be transported by surface currents

and winds (Kako et al., 2010), recaptured by shorelines (Kako et al.,
Environmental contamination generated by Anthropogenic Ma- 2014) or degraded into microplastic (Barnes et al., 2009; Cinner et al.,
R

rine-Debris (AMD) represents one of the most ubiquitous and 2018). Distribution and accumulation of plastic into the marine en-
long-lasting environmental change of our planet (Laist, 1987; Ryan, vironment are indeed controlled by circulation patterns and prevail-
2015). AMD is responsible of several ecological, ecotoxicological, ing winds, coastal and seafloor geomorphology (Barnes et al., 2009;
CO

economic and social impacts. However, the extent to which it is harm- Galgani et al., 2000; Savini et al., 2014) and anthropogenic activi-
ing wildlife and plants, endangering human health and reducing the ties (Ramirez-Llodra et al., 2013). Well known hotspots of accumu-
availability of ecosystem good and services (Laist, 1987; Rochman lation include the sea surface, where aggregations of a large amount
et al., 2013; Hengstmann et al., 2017) is still to be properly under- of persistent and light plastic take place at ocean gyres, creating gi-
stood and quantified (Eriksen et al., 2014; Thompson et al., 2009). It ant “garbage-patches” (Eriksen et al., 2014; Law et al., 2010, 2014),
has been estimated that from 5 to 13 million tonnes of litter enter the but also submarine canyons, where litter originating from land accu-
oceans each year (Jambeck et al., 2015; Geyer et al., 2017) and that mulates in large quantities (Pierdomenico et al., 2019) and the shores,
UN

plastic accounts for over 80% of the total AMD (UNEP, 2005; Laist, particularly beaches (Corcoran et al., 2009). Although data docu-
2011; Thiel et al., 2013; Penca, 2018). menting the occurrence of plastic everywhere in the oceans (from
Plastic is persistent and for the most part (roughly 60%) less dense the surface to the deep seafloor - Thompson et al., 2004 and Van
than seawater (Andrady, 2011; Ryan et al., 2009). Once introduced Cauwenberghe et al., 2013) are quite exhaustive, a consistent quan-
into the marine environment from multiple sources (both sea- and tification of the total amount accumulated within the diverse marine
compartments, has not been accurately outlined. While reliable es-
timations have been provided for the giant surface garbage-patches
⁎ (Lebreton et al., 2018; Eriksen et al., 2014), scarce information is
Corresponding author.
available from the deep and poorly unexplored seafloor, but nonethe
Email address: [email protected] (A. Savini)
1
These two authors equally contributed to the work.

https://doi.org/10.1016/j.scitotenv.2019.133581
0048-9697/ © 2019.
2 Science of the Total Environment xxx (xxxx) xxx-xxx

less for the shorelines, where only a few and scattered regional assess- rithm, 2) provide smart gold standards to estimate the algorithm per-
ments were provided (Martin et al., 2018; Vlachogianni et al., 2018; formances, 3) train and test the deep-learning algorithm in near
Andrades et al., 2016; Ebbesmeyer et al., 2012). Plastic accumula- real-time conditions. We believe that our work could be useful to pro-
tion on beaches may represent the terminal phase of oceanic trans- pose new best-practices for applying deep learning to automate the
port or a transient stage with a successive washed to the sea follow- procedure of litter detection and quantification by UAV systems on
ing storms or tides movements (e.g. Shimizu et al., 2008). Know- beaches, which in turn could offer an instrumental tool for sustainable

F
ing the accumulation rate on beaches and associated spatiotemporal solid waste management.
oscillations would be a crucial information to refine global estima-

OO
tion on the dispersal mechanisms of plastic in the marine environ- 2. Materials and methods
ment and its amount in each compartment. Most of our knowledge
on the quantity of plastic accumulated on beaches, at different tempo- 2.1. Study area
ral scales, is based on sparse and regional monitoring activities, per-
formed following different protocols and without standardized proce- The case-study area consists of different islands of the Republic
dures, making difficult data integration and comparisons among re- of Maldives (Fig. 1a), an archipelago composed of 1192 atoll islands
gions (e.g. Galgani et al., 2015; Watts et al., 2017). Beach litter esti- stretched for 860 km, in North-South direction, located in the mid-
mation, at places performed within the framework of dedicated moni- dle of the Indian Ocean. The islands are grouped in 20 administrative

PR
toring activities (among others the Marine Strategy Framework Direc- atolls and divided under three distinct categories: inhabited, uninhab-
tive - Directive 2008/56/EC – Galgani et al., 2014), is also commonly ited and resort islands (Fallati et al., 2017). The archipelago, with its
subjective and time-labour consuming, since it relies on visual census peculiar geographical location and its 644 km of coastline, represents
where items are recorded along transects (Lavers et al., 2016; Lavers the perfect place for the deposition of plastic debris that are drifted
and Bond, 2017). Only recently the use of aerial imagery has been from the surface currents of the Indian Ocean (Barnes, 2004). In addi-
proved to be an appropriate and efficient method to monitor beach lit- tion, local sources of littering are represented by the waste production
ter (Kako et al., 2012; Kataoka et al., 2018; Sha et al., 2018; Deidun et on the inhabited islands and those discharged into the sea from the nu-

ED
al., 2018). In particular, the use of Unmanned Aerial Vehicles (UAVs) merous boats that daily cross the atolls. In the Maldives, the high dis-
equipped with RGB cameras, beside the advantage of the low-cost, al- persion of land mass and population, both of them spread over a dis-
lows the collection of high resolution imagery data (i.e.: at centime- tance of 860 km, creates a negative effect on solid waste management
tre level - Casella et al., 2016; Flynn and Chapra, 2014) over quite issue. With the exception of resort islands, which represent the 6% of
large areas (e.g. hundreds of hectares), also not easily accessible, with the total archipelago surface and where beach clean-up is a daily rou-
great flexibility in terms of time and frequency of data collection (i.e. tine operated by resort employees, most part of the coastline of each
decades of hectares per day), and under conditions where satellites Maldivian island, is indeed covered, to an undefined extent, by AMD
CT
would be of limited use (i.e.: high cloud cover, limited image reso- coming from different sources (Fig. 2).
lution). Nevertheless, estimation of beach litter from RGB imagery As testing regions for our study, we selected three different coastal
of various sources (UAVs included), over large and even remote ar- areas. Two of these islands (Adangau and Jinnathuga) are in Faafu
eas, still requires standardization of sampling techniques and data pro- Atoll, one (En'Boodhoo) is in Alif Dhaalu Atoll (Fig. 1b, c). The three
cessing. Also, objective identification of plastic items on aerial im- islands were chosen as representatives of small-size uninhabited is-
agery, based on automatic image classification is a novel field of in- lands of the archipelago, where beach-cleaning cannot be guaranteed
RE

vestigation. To the best of our knowledge, only one work has been by resort employees or government personnel. En'Boodhoo is an is-
recently published on the use of UAV Remote Sensing combined to land of 1.8 ha, located in the western lagoon of Alif Dhaalu Atoll. The
Artificial Intelligence (AI) for beach-litter monitoring by Martin et al. island is desert, and the human presence is mainly related to the sa-
(2018). They proved the ability of machine learning (ML) in perform- fari boats that stop nearby, and to touristic picnic and barbeques on
ing less time-labour consuming (40 times faster than humans) and sub- the beach. Adangau (1.1 ha) and Jinnathuga (1.9 ha) are two islands of
jective methodologies to detect AMD, but the best sensitivity reported Faafu Atoll located in the Atoll's eastern lagoon. These two are simi-
R

in Martin et al. (2018) for AMD automatic quantification was low larly used as picnic island from the inhabitants of the Atoll. The pri-
(i.e.: 44%). mary sources of litter on these islands are both the direct release of
In our study, we therefore focused on the improvement of the sen- waste and oceanic transport. The target testing area within the islands
sitivity of the AI algorithm and the associated positive predicted val- was selected as a portion of the beach with direct access to the sea, dif-
CO

ues, which account for the false positive AMD. For this purpose, we ferent exposition to the winds and currents, with the presence of psam-
provided a deep learning, rather than a random-forest, machine-learn- mophytes plants and natural debris (leaves, roots and twigs) as well as
ing approach, as previously implemented by Martin et al. (2018), be- litter.
ing deep learning more beneficial for object detection (LeCun et al.,
2015; Chollet, 2017; Guest et al., 2018). An essential output of our 2.2. Aerial surveys
work is, in addition, the formulation of a combination of protocols to
automatically detect and quantify beach litter along the shores of se- 2.2.1. UAV
UN

lected remote islands in the Republic of Maldives, defined by the 2010 In order to achieve large-scale reproducibility of a protocol to
UNPD's Assessment of Development Results a vulnerable “Small Is- collect UAV images, we propose to use a consumer-grade UAV,
land Developing State (SIDS)”. The protection of the environment equipped with a high-resolution RGB camera, to survey the study
from pollution is indeed extremely important for SIDS as, aside from area. For this purpose, we used the DJI Phantom 4 drone, a quad-
other reasons that are common to all countries, two important indus- copter with high sensing qualities, equipped with a 1/2.3″ CMOS
tries (tourism and fisheries) depend on a pristine environment (UNEP, camera sensor (12.4 MP) that can collect images with a resolution
1999). (R) of 4000 × 3000 pixels and an integrated GPS/GLONASS system.
Our study proposes an ad-hoc combination of protocols to: 1) col- Compared to fixed-wing UAVs, that can cover with a single flight
lect UAV-images suitable for the training of a deep-learning algo a larger area and can handle a higher quality camera, Phantom 4 is
 Science of the Total Environment xxx (xxxx) xxx-xxx 3

F
OO
PR
Fig. 1. Geographic location of the study area, Republic of Maldives (a.), Alif Dhaalu and Faafu Atolls (b.), and islands selected as testing sites (c.): 1 En'Boodhoo, 2 Jinnathuga, 3

ED
Adangau.
CT
R RE
CO

Fig. 2. Debris accumulation in the Maldives: (a) one of the ferries port in Malè (capital city) with evident accumulation of plastic bottles released in the ocean from the boats or from
the streets of the city; (b) Litter and plastic debris accumulated near the shore of Thilafushi, the only landfill island of the archipelago; (c, d) plastic waste deposited by the high tide
UN

on the beach of Adangau; (e, f) plastic waste deposited in the bushes and on the shores of En'Boodhoo.

lightweight, easy to carry, and can smoothly fly at low altitude to ob- 2.2.2. UAV survey protocols
tain good ground-resolution images. Moreover, easy take off and land- Three different altitudes, namely 10, 15 and 35 m were considered
ing procedures make this drone an outstanding, cost-effective solution to define the optimal protocol in terms of image quality and number of
for low altitude and short-range studies. All the metadata are recorded images required to cover the area of interest (AOI):
in an EXIF (Exchangeable Image File Format) file, which includes in-
formation on the pictures such as shutter speed, apertures, ISO and 1) ground sample distance (GSD) being defined as:
GPS coordinates (latitude, longitude and altitude). Flight time with a
single battery is roughly 25 min.
4 Science of the Total Environment xxx (xxxx) xxx-xxx

The surveys were planned using DJI GS PRO (www.dji.com/it/

ground-station-pro) a free Ipad application released by DJI. This app
allows designing all the aspects of the drone mission: generate opti-
mal flight paths, set camera parameters and directly monitor data ac-
quisition on the Ipad screen. For all the surveys we set a fix flight
where SW is the sensor width, FH is the flight high, FL is the focal altitude with a frontal and lateral overlap of 80% and 70%, respec-

F
length of the camera, and IW is the image width (Ventura et al., 2018), tively, a −90° gimbal angle (nadir orientation), a shooting interval of
and 2 s (equal time interval mode) and a constant velocity of 1.3 m/s. Be-

OO
fore starting the UAV overflight weather condition (wind speed, cloud
2) number of images of interest (IOI) being defined as:
coverage) and the presence of obstacles along the path was checked.
A metric tape was laid on the beach (e.g. for several meters), in order
to check the spatial accuracy of the orthomosaic during the postpro-
cessing. Once all the parameters were set, the UAV automatically took
off and completed the mission (e.g. trajectory in Fig. 3a, b).

2.2.3. Reconstruction of the AOI

PR
The images were processed by Agisoft PhotoScan (www.agisoft.
where D is the dimension of the area covered by a single image at a com), a commercial Structure from Motion (SfM) software, widely
specific GSD, and R is the resolution in terms of pixel of the images,
as defined in Section 2.2.1.

ED
CT
R RE
CO
UN

Fig. 3. Drone survey over Adangau (a.) sandy long beach. The path followed by the drone (b.) is overlayed on the high-resolution orthomosaic that was generated applying the SfM
workflow (c.) to RGB images.
 Science of the Total Environment xxx (xxxx) xxx-xxx 5

used by the scientific community for its user-friendly interface, spon- Then, these objects were re-arranged in the specific AOI to be moni-
taneous workflow and the excellent quality of the point cloud out- tored by UAV overflights.
put (Burns and Delparte, 2017; Cook, 2017; Bonali et al., 2019). The
process is dived in three main steps (Fig. 3c): drone photos alignment 2.3.2. Image screening
using high accuracy setting; high-quality dense 3D point cloud gen- In order to evaluate the quality of the images, a comparison was
eration; creation of a Digital Terrain Model (DTM) from the dense made between the number and type of items counted by the operators

F
could. As final outputs we obtained, from the DTM, orthomosaics during the in-situ GA on the beach, and the number and type of items
with a GSD of 4.4 mm/px, 8.2 mm/pix and 14 mm/pix respectively for counted by an operator during an image screening (IS) of the orthomo-

OO
the three flights altitude (10, 15 and 35 m). The models are geo-ref- saic on a PC. The AMD recognised via IS represents our second GS
erenced thanks to the coordinates stored into the EXIF files of each (IS-GS), and a smart best-estimate of the actual GA-GS. Indeed, prov-
image. For more extensive information on the process, see Verhoeven ing the feasibility of identifying and estimating the AMD by screening
(2011) and Ventura et al. (2016). The orthomosaics generated from images collected by UAVs, rather than by operators on the ground, is
images collected at 10 m altitude, were considered as our AOI and crucial to choose the optimal UAV survey-protocol for the collection
used for data assessment. of images suitable for deep learning, and to establish procedures that
allow avoiding the time-consuming GA of the operators.
2.3. Gold standards

PR
2.4. The deep-learning algorithm
2.3.1. In-situ ground assessment
We performed an in-situ ground assessment (GA) of the AOI, In order to allow easy access to AI non-expert users, we used a
aimed at quantifying the AMD on the studied shorelines in order to commercial software – PlasticFinder (Italian software license 012677
calculate the efficiency of the UAV survey-protocols. This quantity D011755, DeepTrace Technologies, www.deeptracetech.com/) – to
represents our first Gold Standard (GA-GS). As a first step, a recog- detect and quantify AMD. The core algorithm of the software is a
nition of the investigated shorelines was conducted to detect AMD. deep-learning convolutional neural network (CNN). CNNs are a class

ED
The items found during the inspection were counted and classified by of multilayer architecture suitable for processing RGB images for
the operators into different subtypes (Table 1). The category named classification and object detection tasks, where the stack of convolu-
“other” refers to objects and fragments that were smaller than 5 cm, tional layers allows for translation invariance - i.e. the net is trained to
which is the minimum size of the target objects that we decided when recognize an object independently of its position within the image.
defining the optimal protocol described in terms of the image quality. The adoption of a deep learning approach has one main motivation.
In order to provide a tool that could favour a scalable approach, i.e.
adaptable to different scenarios, a large image-database was needed
CT
Table 1
to provide a general training set, i.e. a set of images to let the algo-
AOI, number of minimum IOI, climate, light and weather conditions of each investi- rithm learn the classes of interest. The main advantage of deep learn-
gated Maldivian island during the UAV survey and the in-situ ground assessment (GA). ing is that it automates the most critical part of a ML workflow: the
The number of items collected during the GA on each beach is listed per each class, and feature extraction. In contrast to conventional ML methods (e.g. Ran-
the items identified via image screening (IS) via PC are also reported.
dom Forest, Support Vector Machine, Gradient Boosting Machines),
Jinnathuga Adangau En'Boodhoo that require hand-design features as input, a neural network is made of
RE

2
trainable multilayers that learn automatically the features through geo-
AOI (m ) 216 1056 225 metric transformations and gradual adjustments of learning weights
IOI 1 5 1
Climate (month) April November October
with respect to a feedback signal, thus being more suitable than con-
Light (time) 12 pm 12 pm 5 pm ventional ML for large dataset training (LeCun et al., 2015, Chollet,
Weather (conditions) Sunny Sunny Cloudy 2017). The PlasticFinder CNN has been tailored for 5 classes of im-
Use Training set Testing set Testing set ages, namely: “vegetation”, “sea”, “sand”, “AMD” and “other” (i.e.
AMD class GA IS GA IS GA IS
R

Lighter 4 1 1 1 4 4
sand with small pieces of wood, stones, algae). UAV images obtained
Bottle 21 21 50 54 47 43 from the survey of Jinnathuga island were used for the collection of
Straw 1 0 1 0 0 0 the training set (Fig. 4).
Net 11 8 3 3 2 3 The island was chosen on the basis of the fact that all the classes of
CO

Plastic bag 7 7 43 50 3 2 interest were present. Therefore, we selected training images, within
Aluminum can 8 6 11 13 21 14
Plastic containers 3 3 1 1 12 7 the AOI, representing the classes of interest. For each class, a bal-
Plastic utensils 0 0 20 8 1 0 anced number - of the order of thousands – of different samples was
Flip flop 1 1 13 8 32 29 collected, in order to tailor the algorithm on the specific experimen-
Other 26 25 4 3 13 19 tal settings. A subsample of UAV images (N = 3) collected on the
Total 82 72 147 141 135 120
Matching score (%) 87.8 95.9 88.8
UN

Fig. 4. Examples of Jinnathuga island images (a., b. and c.), used as training set for the deep-learning algorithm. Note the presence of the different classes in the images, in particular
“AMD” in (a), “sand” in (a), (b) and (c), “sea” in (b), “vegetation” and “other” in images (a) and (c).
6 Science of the Total Environment xxx (xxxx) xxx-xxx

other two islands, Adangau and En'Boodhoo, were used for the testing found via IS to the AMD found during the GA. This score accounts
set. The surveys of the testing-set islands were finalised at different for the estimated error in the use of the GS produced via IS.
experimental conditions (Table 1) which allowed investigating the in-
fluence of climate, light and shadow on the efficacy of the algorithm. 3.3. The deep-learning algorithm
When a tested image is input in the software, it returns pixel-wise
classification heatmaps, representing a pixel probability-map for each 3.3.1. Training, testing and performance

F
class, and a bounding-boxes map with the detected AMD. The training of the tailored CNN, performed on images from Jin-
The performance of the automatic detection, classification and nantuga-island (Fig. 4), achieved a validation accuracy higher than

OO
quantification were measured by comparing the results with the two 95%. Adangau and En'Boodhoo-islands images (Figs. 5a–6a) were
GS. The metric is expressed in terms of true positive (TP), false nega- used to test the algorithm. A pixel-wise probability heat-map of each
tive (FN) and the false positive (FP) items, rather than in terms of pix- input image has been obtained by the software, as well as a bound-
els, for an easier interpretation. The statistical measure of the perfor- ing-boxes map for the detected AMD (Figs. 5a–6a). In particular, for
mances is expressed through the Sensitivity = TP / (TP + FN), the Posi- each pixel, a probability is given to be classified as AMD, thus allow-
tive Predictive Value PPV = TP / (TP + FP), and the harmonized mean ing a visual understanding of the specific areas that might be subjected
of Sensitivity and PPV, given by the F-score = 2TP / (2TP + FP + FN). – with a different probability of risk exposure – to the presence of plas-
tic debris.

PR
3. Results Figs. 5c and 6c, shows bounding-boxes maps with all AMD de-
tected. Table 2 reports the numerical results obtained by comparing
3.1. Optimization of UAV survey protocols the software output and the IS-GS for each image. The results high-
light the average software-performance for En'Boodhoo drops of a
The optimal protocol for the UAV survey, as a compromise be- factor of about 3 with respect to the Adangau case.
tween image resolution and number of IOI to cover the AOI, was Therefore, these results give the evidence that the collection of
found at a UAV altitude of 10 m, corresponding to a GSD of 4.4 mm/ UAV- images suitable for the training and testing of the deep-learn-

ED
pixels. The know dimension of the objects (metric tape) in the or- ing algorithm, should rely on specific recommendations regarding the
thomosaics strongly matches the true dimensions measured on the optimization of the UAV survey, the collection of the GS, and the de-
beaches with an average accuracy of ≈1 mm. Table 1 reports the dif- velopment of the algorithm itself.
ferent AOI covered following the optimal protocol for three selected
islands, namely Jinnathuga, Adangau and En'Boodhoo islands, and the 4. Discussion
corresponding number of minimum IOI.
4.1. The UAV survey and the AMD detection
CT
3.2. Gold standards
The low cost, the high resolution and the high flexibility of UAVs
Table 1 reports the experimental results of the in-situ GA and of quickly turned out to make them extremely versatile and useful tools
the IS of the AOI. The matching scores express the ratio of the AMD for the investigation and analysis of a number of environmental is-
sues. Small UAVs are used indeed with increasing frequency, in
R RE
CO
UN

Fig. 5. Canvas on the left (a.), from top to bottom: testing-set images A1, A2 and A3 for Adangau-island. Central canvas (b.), from top to bottom: PlasticFinder pixel-wise classifica-
tion heatmaps for Adangau-island images A1, A2 and A3 representing, respectively, the classes sea (A), sand (B), vegetation (C) other/AMD (D), and AMD (E) with the probability
scale ranging from 0 to 1. Canvas on the right (c.), from top to bottom: PlasticFinder bounding-boxes maps for the Adangau-island images A1, A2 and A3. Each green bounding box
is identified as an item of AMD by the software. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 Science of the Total Environment xxx (xxxx) xxx-xxx 7

F
OO
PR
ED
Fig. 6. Canvas on the left (a.), from top to bottom: testing-set images E1, E2, E3 for En'Boodhoo island. Central canvas (b.), from top to bottom: PlasticFinder pixel-wise classification
heatmaps for En'Boodhoo-island images E1, E2, E3 representing, respectively, the classes sea (A), sand (B), vegetation (C), other/AMD (D), and AMD (E) with the probability scale
ranging from 0 to 1. Canvas on the right (c.), from top to bottom: PlasticFinder bounding-boxes maps for the En'Boodhoo-island images E1, E2, E3. Each green bounding box is
identified as an item of AMD by the software. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 2 camera (Kako et al., 2012) and aerial photographs (Kataoka et al.,
Results for the Adangau and En'Boodhoo testing-set images A1, A2, A3 and E1, E2, 2018), faced problems related to the orthorectification and to the
E3, respectively. Average scores (AVG) are also given for each set. AMD accounts for
pixel-size of the images: a GSD of 10 cm/pix allowed identifying only
CT
the total real items in each image, as identified by the gold standard. True positive (TP),
false negative (FN) and false positive (FP) items are combined to express the software groups of debris and not the single objects. The recent adoption of
performance in terms of sensitivity, PPV and F-score. UAVs for AMD monitoring overcome a number of limitations mainly
related to the flight altitude and to the GSD, to the orthorectifica-
Sensitivity PPV F-score
IMG AMD TP FN FP (%) (%) (%)
tion of the images, and to the repeatability of the surveys in a short
time. However, the data-processing procedures are not uniform, rang-
Adangau ing from visual interpretation of the images (Deidun et al., 2018) and
RE

A1 37 25 12 0 0.68 1.00 0.81

spectral profile analysis of litter (Sha et al., 2018), to the use of ma-
A2 58 40 18 4 0.69 0.91 0.78
A3 89 56 33 6 0.66 0.90 0.74 chine learning methods (Martin et al., 2018). The use of AI classifiers
AVG 61.3 40.3 21 3.3 0.67 0.94 0.78 (Martin et al., 2018) is more complex for different reasons, among
En'Boodhoo which the lack of publicly available large databases (providing ade-
E1 71 23 48 66 0.32 0.26 0.29 quate images to train algorithms) is notable. The difficulty of devel-
E2 43 5 38 6 0.12 0.45 0.19
E3 98 25 73 65 0.26 0.28 0.27
oping scalable approaches, i.e. procedures that do not depend on local
R

AVG 70.6 17.6 53 45.6 0.23 0.25 0.33 environmental constrains, is also a major issue for the use of AI clas-
sifiers.
The advantages in using UAVs, in terms both of resolution and
many research activities with applications in different fields: structural monitoring repeatability, match perfectly with the need of understand-
CO

geology (Bonali et al., 2019), forestry sciences (Baron et al., 2018; ing the pattern of aggregation in a remote area such as the Republic
Mlambo et al., 2017), mapping of sensitive marine habitats (Ventura of Maldives. Here, a considerable amount of marine litter has been
et al., 2018), marine megafauna surveys (Colefax et al., 2018; Kiszka reported, despite the remoteness of the location (Imhof et al., 2017).
et al., 2016), coral bleaching detection (Levy et al., 2018). These plat- However, Imhof et al. (2017) highlighted the need for a robust pro-
forms, especially the commercial drones, are proving to be useful tools tocol, allowing extensive sampling in space and time to collect scien-
for high-resolution remote sensing data collection, especially because tifically sound data (Imhof et al., 2017). Remote-sensing studies, re-
of their small size, the increased lifetime of the batteries and the pos- lated to the accumulation and transportation of AMD, were not con-
UN

sibility to plan autonomous flights with user-friendly ground station ducted before this work in the Republic of Maldives (Fig. 7). The
software. Moreover, SfM algorithms allow obtaining accurate Digital lack of such monitoring studies for this area is significant, consid-
Terrain Models (DTMs) and orthomosaics over large areas. ering that plastic debris from the rivers of South Asia contributes to
AMD was monitored worldwide through aerial surveys, along the 67% of the global annual input (Lebreton et al., 2017) and that coun-
beaches, since 2012 (Kako et al., 2012; Deidun et al., 2018; Kataoka tries on the Indian Ocean are among the principal producer of mis-
et al., 2018; Martin et al., 2018; Sha et al., 2018) but explored loca- managed plastic waste (Jambeck et al., 2015). Besides, it is not clear
tions are still limited (Fig. 7). Besides, more significant, it is the ab- where all this plastic, that should accumulate in the Indian Pacific,
sence of a standardized protocol for data acquisition and elaboration. gyre is going (Mheen et al., 2019). Thus, the proposed methodology
Previous studies, performed using a balloon equipped with a digital will improve and standardise the data collection of marine-litter accu
8 Science of the Total Environment xxx (xxxx) xxx-xxx

F
OO
PR
Fig. 7. Geographic distribution of studies that used remote sensing techniques to monitor and detect beach debris: 1 Vancouver Island, Canada (Kataoka et al., 2018); 2 Malta (Deidun
et al., 2018); 3 Saudi Arabia (Martin et al., 2018); 4 Fuzhou, Fujian, China (Sha et al., 2018); 5 Seto Inland Sea, Japan (Kako et al., 2012); 6 Republic of Maldives, present study area.

ED
mulation on beaches and shorelines, gathering valuable and compara- the closest inhabited island or discharged from boats that passed
ble data, even in remote and isolated areas. nearby; others may have float in the ocean for thousands of kilome-
The results of our study confirm most of the advantages of using a tres before reaching the shore. Instead, the higher presence of plastic
consumer-grade drone to carry out environmental monitoring. In par- bags (foods wraps and plastic bags) on Adangau island is most proba-
ticular, the use of a DJI Phantom 4 drone allowed speeding up consid- bly due to the use of the island as a picnic and barbeque location from
erably the standard walking beach survey and to access remote areas the inhabitants of the atoll.
CT
such as Maldives. The graphical outputs and the numerical results for Adangau Island
On the surveyed islands, anthropogenic debris were found every- show good performances (Fig. 5, Table 2) and, in particular, in the
where: on the water's edge, just left there from waves and tides; face of an average sensitivity of 67%, the average PPV reaches 94%.
on the upper part of the beaches and in the bushy coastal vegeta- This means that the deep-learning algorithm performance is affected
tion, likely carried there from storm tides and winds, or left by lo- by a non-negligible number of FN items – impacting on the sensitiv-
cal tourists. The selected GSD allowed the identification and catego- ity – but also that, on the other hand, the software is highly specific
rization of debris for each single detected item, making our remote
RE

in the ability to recognize AMD with respect to FP items. In addi-

observations comparable to the ones performed by operators on the tion, it should be noted that, for Adangau island, the software perfor-
ground. In fact, the matching between the ground-assessment and the mance is quite stable, even when the number of real AMD-items on
visual screening of the images is higher than 87% (Table 1). In addi- the beach increases (Table 2). The pixel-wise classification heatmaps
tion, the results of our survey protocol shows that the majority of the for En'Boodhoo-island, shown in Fig. 6b, qualitatively confirm the
detected objects (Table 1) were plastics bottles and aluminum cans. good ability of deep learning algorithm, in recognizing the presence
Particularly abundant were also flip flops on En'Boodhoo and Adan- or absence of elements belonging to the “sea” class. Also, the soft-
R

gau. These three categories of debris were observed with different ware correctly reports on a low-probability with respect to the pres-
degradation level: from brand new, with labels and the colours still ence of vegetation (below 30%). On the contrary, the heatmaps for
intact, to partially disrupted. This can indicate the heterogeneity of the other classes draw attention to issues that become evident when
CO

the sources: some of them can be just washed up on the shore from looking at the zoomed bounding-boxes maps in Fig. 8, and to
UN

Fig. 8. Examples of shadows in the proximity of footprints (left image) and of items of AMD (right image), that are mistaken as AMD by the software, representing pitfalls for the
algorithm in its present version.
 Science of the Total Environment xxx (xxxx) xxx-xxx 9

the quantitative results shown in Table 2. In this case, the average soft- 4.2. Best-practices optimization and future improvements
ware-performance for En'Boodhoo drops of a factor of about 3 with
respect to the Adangau case. This limitation can be explained by con- In order to optimize and enhance best practices for AMD re-
sidering the different lighting conditions. In particular, Adangau-is- mote-sensing monitoring, further improvements should be applied to
land images were collected at 12 am of a sunny day, therefore with the adopted protocols and methods.
similar sunlight-conditions of the testing set of Jinnathuga island. On In terms of the UAV survey, we reckon that once all the flight pa-

F
the contrary, En'Boodhoo images were collected at 5 pm, and the shad- rameters have been set, the monitoring can be carried out from a small
ows in the proximity of footprints (Fig.8) or of real AMD, represent boat, in the proximity of the shores, without the need of reaching the

OO
pitfalls for the algorithm, as clarified by the high number of FP items beach, often inaccessible for the presence of coral reefs all around the
in Table 2. As a matter of facts, the software was not trained to recog- islands. Therefore, this methodology can be particularly useful in ge-
nize footprints or shadows, and therefore such a limitation restricts the ographical sites, such as the Maldives, where the presence of many
use of PlasticFinder, in its present version, to specific sunlight-condi- small remote uninhabited islands, and the need to optimize the AMD
tions. For these reasons, we suggest conducting the survey with the beach-monitoring, represents a pressing matter.
sun high on the horizon, in order to avoid excessive shadows on the In terms of the ground assessment protocol, such assessment is
surveyed areas. However, to date, and to the best of our knowledge, used to validate the reliability of the methodology, but, in implement-
there is only another algorithm presented to the scientific community, ing the protocol, it is not supposed to happen every time (otherwise

PR
that has been developed for the specific purpose of automatically de- the protocol would lose its time-efficiency). For this reason it has been
tect and quantify AMD along the shores by using a combination of not included in the Protocol recommendations (Table 4). In general,
UAV images and AI. In their pioneering work, Martin et al. (2018), the beach should be left untouched before the UAV survey to avoid
focused their efforts in the Saudi-Arabian shorelines. They faced the footprints or other environment manipulations that could affect the
highly-challenging task of both detecting and classifying the AMD ty- methodology performance, especially in remote places.
pology with a series of multi-class random-forest classifiers, based on Another important remark is that, for a relevant fraction of the
the extraction of HoG features. The authors validated the feasibility AMD, the deposition on the beach is only a transitory phase before be-

ED
of using AI for AMD detection, but pointed out that the use of deep ing taken up by the currents to resume the floating travel in the ocean.
learning would have been more beneficial with respect to their ap- Instead, other AMD can be trapped on the upper part of the beach,
proach that achieved a maximum sensitivity of only 44%. Therefore, where environmental factors and the erosive action of the sand can
our work represents the first implementation to automatically detect accelerate the plastic degradation processes. The microplastics parti-
and accurately quantify AMD, based on a deep-learning approach. Re- cles (<5 mm) produced by the degradation of the AMD trapped on
sults in Table 3 point out that PlasticFinder performances give better the upper part of the shore can enter in the sediments or can be re-
results, with respect to all the metrics, especially if used in the ap- leased as contaminants in the water of the lagoon (Saliu et al., 2018;
CT
propriate sunlight conditions. In particular, PlasticFinder PPV is much Saliu et al., 2019). Therefore, fast and efficient data collection and
higher than the one obtained by Martin et al. (2018), allowing for a image analysis of the distribution of AMD on the shore, as well as
more specific tool to alarm on and quantify the presence of AMD. In specific AI tools for its automatic and objective assessment are nec-
fact, it is important to highlight that, in order to monitor the presence essary, but not sufficient, since microplastic is lost from this detec-
of AMD and to know which are the areas that require an urgent in- tion and quantification. However, the distribution and quantitation of
tervention (i.e. those where AMD accumulate the most), it is essential AMD, as obtained from our protocol, could be used to understand
RE

to have a tool that is able to detect only AMD, without mistaken false
positives. To this extent, reaching a high PPV is more crucial than a
high Sensitivity. Also, the fact that the reached Sensitivity is constant, Table 4
despite the different loads of litter on the beach, is also a good result Optimal protocol and key recommendations for the optimization of the UAV survey,
because it shows that the more is the AMD, the higher is the litter de- the collection of gold standards and of UAV images suitable for the training and testing
of a deep-learning algorithm.
tected by PlasticFinder, i.e. this technology is able to detect accumu-
R

lation zones. UAV survey

These results reflect the major advantage of deep learning, with re- Flight altitude 10 m
spect to conventional ML methods, which is the fact that it is not nec- GSD 4.36 mm/pix
Camera Gimbal −90° (nadir orientation)
essary to pre-transform data (e.g. an image) into selected features to
CO

Orientation
feed models, but data can be input into neural-network models to let Images 80% frontal overlap
them automatically identify the best representations that allows tasks acquisition
such as detection or classification (LeCun et al., 2015). along fixed
paths
70% lateral overlap
Table 3 2 s of shooting interval
Comparison between results from Martin et al. (2018) (average on the overall re- 1,3 m/s constant velocity
Gold standard
UN

sults), PlasticFinder results for Adangau island and averaged results for Adangau and
En'Boodhoo island. Ground AMD inspection and subtype classification (to be limited to
assessment few representative areas for the validation of methodology)
Sens PPV F-score AMD size >5 cm
Algorithm TOT TP FN FP (%) (%) (%) Image screening AMD counting and subtype classification
Matching score >80%
Martin et al. 415 164 251 1941 0.40 0.08 0.13 Deep learning
PlasticFinder 61.3 40.3 21 3.3 0.67 0.94 0.78 Training # of images per class ~103
(Adangau) Validation accuracy for tailored CNN >95%
PlasticFinder 131.9 57.9 74 48.9 0.44 0.54 0.49 Testing # of images ~ # IOI
(AVG A/E) Use IS-GS to test performances with metrics
Use GA-GS to estimate error on the IS-quantified
performance
10 Science of the Total Environment xxx (xxxx) xxx-xxx

which are the most impacted areas, and the AMD depositional sea- Baron, J., Hill, D.J., Elmiligi, H., 2018. Combining image processing and machine
learning to identify invasive plants in high-resolution images. Int. J. Remote Sens.
sonal trends connected to the Indian Ocean currents patterns (Mheen
39, 5099–5118. https://doi.org/10.1080/01431161.2017.1420940.
et al., 2019). The creation of an integrated model could allow stake- Bonali, F.L., Tibaldi, A., Marchese, F., Fallati, L., Russo, E., Corselli, C., Savini, A.,
holders (e.g. governments, NGO) using this information in order to 2019. UAV-based surveying in volcano-tectonics: an example from the Iceland
promote mitigation actions, such as specific citizenship awareness ini- rift. J. Struct. Geol. https://doi.org/10.1016/j.jsg.2019.02.004.
tiatives, beach–clean up events, but also addressing – with a data-dri- Burns, J.H.R., Delparte, D., 2017. Comparison of commercial structure-from-motion

F
photogrammety software used for underwater three-dimensional modeling of coral
ven approach - the interception of the floating AMD, before reaching reef environments. In: International Archives of the Photogrammetry, Remote
the shorelines. Sensing and Spatial Information Sciences - ISPRS Archives. vol. 42, pp. 127–131.

OO
Finally, we would suggest some improvement for PlasticFinder, https://doi.org/10.5194/isprs-archives-XLII-2-W3-127-2017.
for example, by a more in-depth training, with the aim of avoiding Casella, E., Collin, A., Harris, D., Ferse, S., Bejarano, S., Parravicini, V., … Rovere,
A., 2016. Mapping coral reefs using consumer-grade drones and structure from
sunlight-conditions dependence for its use, which represents, to date, motion photogrammetry techniques. Coral Reefs https://doi.org/10.1007/s00338-
one of its major limitations. Also, optimization should be implemented 016-1522-0.
to scale up the algorithm speed and ability to process full orthomosaic Chollet, F., 2017. Deep Learning With Python, 1st edition Manning Publications Co.,
images, overcoming time-scale limitations due to processing of a large Greenwich, CT, USA.
Cinner, J.E., Maire, E., Huchery, C., MacNeil, M.A., Graham, N.A.J., Mora, C., …
amount of data. In Table 4, we summarize and suggest an optimal pro- Mouillot, D., 2018. Gravity of human impacts mediates coral reef conservation
tocol, with key recommendations.

PR
gains. Proc. Natl. Acad. Sci.Retrieved from http://www.pnas.org/content/early/
2018/06/12/1708001115.
5. Conclusion Colefax, A.P., Butcher, P.A., Kelaher, B.P., 2018. The Potential for Unmanned Aerial
Vehicles (UAVs) to Conduct Marine Fauna Surveys in Place of Manned Aircraft.
ICES Journal of Marine Science. Oxford University Presshttps://doi.org/10.1093/
Our work was aimed at proposing an efficient and reliable moni- icesjms/fsx100.
toring protocol, to address a pressing worldwide environmental issue Cook, K.L., 2017. An evaluation of the effectiveness of low-cost UAVs and structure
such as AMD deposition along the shores. Low altitude remote-sens- from motion for geomorphic change detection. Geomorphology 278, 195–208.
https://doi.org/10.1016/j.geomorph.2016.11.009.
ing data are essential for obtaining a synoptic overview of extended
Corcoran, P.L., Biesinger, M.C., Grifi, M., 2009. Plastics and beaches: a degrading re-

ED
areas, and UAVs are powerful tools to acquire them. Our study, con- lationship. Mar. Pollut. Bull. https://doi.org/10.1016/j.marpolbul.2008.08.022.
firmed the use of a commercial drone for AMD monitoring as a fast Deidun, A., Gauci, A., Lagorio, S., Galgani, F., 2018. Optimising beached litter moni-
and reliable surveys methodology. The use of UAV is instrumen- toring protocols through aerial imagery. Mar. Pollut. Bull. 131, 212–217. https://
tal to survey remote areas and the spatial resolution achieved in the doi.org/10.1016/J.MARPOLBUL.2018.04.033.
Department of Economic and Social Affairs Division for Sustainable Development,
collected images allowed detecting a percentage of the objects on 2011. Trends in Sustainable Development Small Island Developing States
the shores higher than 87.8%. A deep-learning based software, Plas- (SIDS). Retrieved from https://sustainabledevelopment.un.org/content/documents/
ticFinder, has been used for the automatic detection and quantifica- 313Trends_in_Sustainable_Development_SIDS.pdf.
CT
tion of AMD, providing analysis of the UAV collected images. In the Ebbesmeyer, C.C., Ingraham, W.J., Jones, J.A., Donohue, M.J., 2012. Marine debris
from the oregon dungeness crab fishery recovered in the Northwestern Hawaiian
Maldivian case study, the overall performances were good, reaching Islands: identification and oceanic drift paths. Mar. Pollut. Bull. https://doi.org/10.
a PPV of 94% with the better sunlight conditions, much greater than 1016/j.marpolbul.2011.09.037.
the only state-of-the-art AI algorithm so far published in literature. Eriksen, M., Lebreton, L.C.M., Carson, H.S., Thiel, M., Moore, C.J., Borerro, J.C., …
The only critical limitations, observed in our study, are determined Reisser, J., 2014. Plastic pollution in the world's oceans: more than 5 trillion plas-
tic pieces weighing over 250,000 tons afloat at sea. PLoS ONE https://doi.org/10.
by environmental circumstances encountered during the survey, and 1371/journal.pone.0111913.
RE

especially sunlight conditions and the associated terrain shading ef- Fallati, L., Savini, A., Sterlacchini, S., Galli, P., 2017. Land use and land cover
fects: restrictions are given for the images that can be analysed with (LULC) of the republic of the Maldives: first national map and LULC change
the deep-learning algorithm in its present version, where the PPV is analysis using remote-sensing data. Environ. Monit. Assess. 189 (8), 417. https://
doi.org/10.1007/s10661-017-6120-2.
reduced to 54%.
Flynn, K.F., Chapra, S.C., 2014. Remote sensing of submerged aquatic vegetation in a
shallow non-turbid river using an unmanned aerial vehicle. Remote Sens. 6 (12),
Uncited reference 12815–12836. https://doi.org/10.3390/rs61212815.
R

Galgani, Francois, Ryan, P.G., Moore, C.J., Eriksen, M., Borerro, J.C., Carson, H.S.,
Department of Economic and Social Affairs Division for … Thiel, M., 2014. Plastic pollution in the world's oceans: more than 5 trillion
plastic pieces weighing over 250,000 tons afloat at sea. PLoS ONE 9 (12),
Sustainable Development, 2011 e111913. https://doi.org/10.1371/journal.pone.0111913.
CO

Galgani, François, Hanke, G., Maes, T., 2015. Global distribution, composition and
Acknowledgments abundance of marine litter. In: Marine Anthropogenic Litter. Springer International
Publishing, Cham, pp. 29–56. https://doi.org/10.1007/978-3-319-16510-3_2.
Galgani, F., Leaute, J.P., Moguedet, P., Souplet, A., Verin, Y., Carpentier, A., … Ner-
The authors are grateful to MaRHE Center and Planhotel Hospital- isson, P., 2000. Litter on the sea floor along European coasts. Mar. Pollut. Bull.
ity Group, in particular, Diamonds Thudufushi Resort, for the logistic https://doi.org/10.1016/S0025-326X(99)00234-9.
support provided during the field work. Geyer, R., Jambeck, J.R., Law, K.L., 2017. Production, use, and fate of all plastics ever
made. Sci. Adv. https://doi.org/10.1126/sciadv.1700782.
Guest, D., Cranmer, K., Whiteson, D., 2018. Deep learning and its application to LHC
References
UN

physics. Annu. Rev. Nucl. Part. Sci. 68 (1), 161–181. https://doi.org/10.1146/

annurev-nucl-101917-021019.
Andrades, R., Martins, A.S., Fardim, L.M., Ferreira, J.S., Santos, R.G., 2016. Origin of Hengstmann, E., Gräwe, D., Tamminga, M., Fischer, E.K., 2017. Marine litter abun-
marine debris is related to disposable packs of ultra-processed food. Mar. Pollut. dance and distribution on beaches on the Isle of Rügen considering the influence
Bull. 109 (1), 192–195. https://doi.org/10.1016/J.MARPOLBUL.2016.05.083. of exposition, morphology and recreational activities. Mar. Pollut. Bull. 115 (1–2),
Andrady, A.L., 2011. Microplastics in the marine environment. Mar. Pollut. Bull. 62 297–306. https://doi.org/10.1016/J.MARPOLBUL.2016.12.026.
(8), 1596–1605. https://doi.org/10.1016/J.MARPOLBUL.2011.05.030. Imhof, H.K., Sigl, R., Brauer, E., Feyl, S., Giesemann, P., Klink, S., … Laforsch, C.,
Barnes, D.K.A., 2004. Surveys from Negombo in West Sri Lanka, Ari Atoll in the 2017. Spatial and temporal variation of macro-, meso- and microplastic abundance
Maldives. Retrieved from http://www.cordell.org/HD/HD_documents/HE_ on a remote coral island of the Maldives, Indian Ocean. Mar. Pollut. Bull. 116
Library/Debris/RIA.pdf. (1–2), 340–347. https://doi.org/10.1016/j.marpolbul.2017.01.010.
Barnes, D.K.A., Galgani, F., Thompson, R.C., Barlaz, M., 2009. Accumulation and Jambeck, J.R., Geyer, R., Wilcox, C., Siegler, T.R., Perryman, M., Andrady, A., …
fragmentation of plastic debris in global environments. Philos. Trans. R. Soc. B Law, K.L., 2015. Marine pollution. Plastic waste inputs from land into the ocean.
364 (1526), 1985–1998. https://doi.org/10.1098/rstb.2008.0205. Science (New York, N.Y.) 347 (6223), 768–771. https://doi.org/10.1126/science.
1260352.
 Science of the Total Environment xxx (xxxx) xxx-xxx 11

Kako, S., Isobe, A., Seino, S., Kojima, A., 2010. Inverse estimation of drifting-object Pierdomenico, M., Casalbore, D., Chiocci, F.L., 2019. Massive benthic litter funnelled
outflows using actual observation data. J. Oceanogr. https://doi.org/10.1007/ to deep sea by flash-flood generated hyperpycnal flows. Sci. Rep. 9 (1), 5330.
s10872-010-0025-9. https://doi.org/10.1038/s41598-019-41816-8.
Kako, S., Isobe, A., Magome, S., 2012. Low altitude remote-sensing method to moni- Programme, U. N. E, 2005. Marine Litter an Analytical Overview www.unep.org. Re-
tor marine and beach litter of various colors using a balloon equipped with a digital trieved from http://wedocs.unep.org/handle/20.500.11822/8348.
camera. Mar. Pollut. Bull. 64 (6), 1156–1162. https://doi.org/10.1016/j.marpolbul. Ramirez-Llodra, E., De Mol, B., Company, J.B., Coll, M., Sardà, F., 2013. Effects of
2012.03.024. natural and anthropogenic processes in the distribution of marine litter in the deep

F
Kako, S., Isobe, A., Kataoka, T., Hinata, H., 2014. A decadal prediction of the quantity Mediterranean Sea. Prog. Oceanogr. https://doi.org/10.1016/j.pocean.2013.07.027.
of plastic marine debris littered on beaches of the East Asian marginal seas. Mar. Rochman, C.M., Browne, M.A., Halpern, B.S., Hentschel, B.T., Hoh, E., Karapanagi-
Pollut. Bull. https://doi.org/10.1016/j.marpolbul.2014.01.057. oti, H.K., … Thompson, R.C., 2013. Policy: classify plastic waste as hazardous.

OO
Kataoka, T., Murray, C.C., Isobe, A., 2018. Quantification of marine macro-debris Nature https://doi.org/10.1038/494169a.
abundance around Vancouver Island, Canada, based on archived aerial pho- Ryan, P.G., 2015. A brief history of marine litter research. In: Marine Anthropogenic
tographs processed by projective transformation. Mar. Pollut. Bull. 132 (Septem- Litter. https://doi.org/10.1007/978-3-319-16510-3_1.
ber 2017), 44–51. https://doi.org/10.1016/j.marpolbul.2017.08.060. Ryan, P.G., Moore, C.J., Van Franeker, J.A., Moloney, C.L., 2009. Monitoring the
Kiszka, J.J., Mourier, J., Gastrich, K., Heithaus, M.R., 2016. Using unmanned aerial abundance of plastic debris in the marine environment. Philos. Trans. R. Soc. B
vehicles (UAVs) to investigate shark and ray densities in a shallow coral lagoon. 364 (1526), 1999–2012. https://doi.org/10.1098/rstb.2008.0207.
Mar. Ecol. Prog. Ser. 560, 237–242. https://doi.org/10.3354/meps11945. Saliu, F., Montano, S., Garavaglia, M.G., Lasagni, M., Seveso, D., Galli, P., 2018. Mi-
Laist, D.W., 1987. Overview of the biological effects of lost and discarded plastic de- croplastic and charred microplastic in the Faafu Atoll, Maldives. Mar. Pollut. Bull.
bris in the marine environment. Mar. Pollut. Bull. https://doi.org/10.1016/S0025- 136, 464–471. https://doi.org/10.1016/j.marpolbul.2018.09.023.

PR
326X(87)80019-X. Saliu, F., Montano, S., Leoni, B., Lasagni, M., Galli, P., 2019. Microplastics as a threat
Laist, D.W., 2011. Impacts of Marine Debris: Entanglement of Marine Life in Marine to coral reef environments: detection of phthalate esters in neuston and scleractin-
Debris Including a Comprehensive List of Species With Entanglement and Inges- ian corals from the Faafu Atoll, Maldives. Mar. Pollut. Bull. 142, 234–241. https://
tion Records. https://doi.org/10.1007/978-1-4613-8486-1_10. doi.org/10.1016/j.marpolbul.2019.03.043.
Lavers, J.L., Bond, A.L., 2017. Exceptional and rapid accumulation of anthropogenic Savini, A., Vertino, A., Marchese, F., Beuck, L., Freiwald, A., 2014. Mapping
debris on one of the world's most remote and pristine islands. Proc. Natl. Acad. cold-water coral habitats at different scales within the Northern Ionian Sea (Cen-
Sci. 114 (23), 6052–6055. https://doi.org/10.1073/pnas.1619818114. tral Mediterranean): an assessment of coral coverage and associated vulnerability.
Lavers, J.L., Oppel, S., Bond, A.L., 2016. Factors influencing the detection of beach PLoS One 9 (1), e87108https://doi.org/10.1371/journal.pone.0087108.
plastic debris. Mar. Environ. Res. https://doi.org/10.1016/j.marenvres.2016.06.009. Sha, J., Shifaw, E., Bao, Z., Li, X., Hanchiso, T., 2018. Monitoring of beach litter by
Law, K.L., Morét-Ferguson, S., Maximenko, N.A., Proskurowski, G., Peacock, E.E., automatic interpretation of unmanned aerial vehicle images using the segmentation

ED
Hafner, J., Reddy, C.M., 2010. Plastic accumulation in the North Atlantic subtropi- threshold method. Mar. Pollut. Bull. 137, 388–398. https://doi.org/10.1016/j.
cal gyre. Science https://doi.org/10.1126/science.1192321. marpolbul.2018.08.009.
Law, K.L., Morét-Ferguson, S.E., Goodwin, D.S., Zettler, E.R., Deforce, E., Kukulka, Shimizu, T., Nakai, J., Nakajima, K., Kozai, N., Takahashi, G., Matsumoto, M., Kikui,
T., Proskurowski, G., 2014. Distribution of surface plastic debris in the eastern pa- J., 2008. Seasonal variations in coastal debris on Awaji Island, Japan. Mar. Pollut.
cific ocean from an 11-year data set. Environ. Sci. Technol. https://doi.org/10. Bull. https://doi.org/10.1016/j.marpolbul.2007.10.005.
1021/es4053076. Thiel, M., Hinojosa, I.A., Miranda, L., Pantoja, J.F., Rivadeneira, M.M., Vásquez, N.,
Lebreton, L.C.M., Van Der Zwet, J., Damsteeg, J.W., Slat, B., Andrady, A., Reisser, J., 2013. Anthropogenic marine debris in the coastal environment: a multi-year com-
2017. River plastic emissions to the world's oceans. Nat. Commun. 8, 1–10. https: parison between coastal waters and local shores. Mar. Pollut. Bull. 71 (1–2),
CT
//doi.org/10.1038/ncomms15611. 307–316. https://doi.org/10.1016/J.MARPOLBUL.2013.01.005.
Lebreton, L., Slat, B., Ferrari, F., Sainte-Rose, B., Aitken, J., Marthouse, R., … Thompson, R.C., Moore, C.J., Saal, F.S.V., Swan, S.H., 2009. Plastics, the environ-
Reisser, J., 2018. Evidence that the Great Pacific Garbage Patch is rapidly accumu- ment and human health: current consensus and future trends. Philos. Trans. R. Soc.
lating plastic. Sci. Rep. 8 (1), 1–15. https://doi.org/10.1038/s41598-018-22939-w. B https://doi.org/10.1098/rstb.2009.0053.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436–444. UNEP, 1999. Western Indian Ocean Environment Outlook-1999. UNEP, Nairobi.
https://doi.org/10.1038/nature14539. Van Cauwenberghe, L., Vanreusel, A., Mees, J., Janssen, C.R., 2013. Microplastic pol-
Levy, J., Hunter, C., Lukacazyk, T., Franklin, E.C., 2018. Assessing the spatial distrib- lution in deep-sea sediments. Environ. Pollut. https://doi.org/10.1016/j.envpol.
ution of coral bleaching using small unmanned aerial systems. Coral Reefs 1–15. 2013.08.013.
RE

https://doi.org/10.1007/s00338-018-1662-5. Ventura, D., Bruno, M., Jona Lasinio, G., Belluscio, A., Ardizzone, G., 2016. A
Martin, C., Parkes, S., Zhang, Q., Zhang, X., McCabe, M.F., Duarte, C.M., 2018. Use low-cost drone based application for identifying and mapping of coastal fish nurs-
of unmanned aerial vehicles for efficient beach litter monitoring. Mar. Pollut. Bull. ery grounds. Estuar. Coast. Shelf Sci. 171, 85–98. https://doi.org/10.1016/j.ecss.
131, 662–673. https://doi.org/10.1016/j.marpolbul.2018.04.045. 2016.01.030.
Mheen, M., Pattiaratchi, C., Sebille, E., 2019. Role of Indian Ocean dynamics on accu- Ventura, D., Bonifazi, A., Gravina, M.F., Belluscio, A., Ardizzone, G., 2018. Mapping
mulation of buoyant debris. J. Geophys. Res. Oceans2018JC014806https://doi.org/ and classification of ecologically sensitive marine habitats using unmanned aerial
10.1029/2018JC014806. vehicle (UAV) imagery and Object-Based Image Analysis (OBIA). Remote Sens.
R

Mlambo, R., Woodhouse, I.H., Gerard, F., Anderson, K., 2017. Structure from motion 10 (9), 1331. https://doi.org/10.3390/rs10091331.
(SfM) photogrammetry with drone data: a low cost method for monitoring green- Verhoeven, G., 2011. Taking computer vision aloft - archaeological three-dimensional
house gas emissions from forests in developing countries. Forests 8 (3), 68. https:// reconstructions from aerial photographs with photoscan. Archaeol. Prospect. 18
doi.org/10.3390/f8030068. (1), 67–73. https://doi.org/10.1002/arp.399.
CO

Thompson, Richard C., Olsen, Ylva, Mitchell, Richard P., Davis, Anthony, Rowland, Vlachogianni, T., Fortibuoni, T., Ronchi, F., Zeri, C., Mazziotti, C., Tutman, P., …
Steven J., John, Anthony W.G., Mcgonigle, Daniel, Russell, A.E., Thompson, Scoullos, M., 2018. Marine litter on the beaches of the Adriatic and Ionian Seas:
R.C., Olsen, Y., Mitchell, R.P., Davis, A., Rowland, S.J., … Russell, A.E., 2004. an assessment of their abundance, composition and sources. Mar. Pollut. Bull. 131,
Lost at sea: where is all the plastic?. Science https://doi.org/10.1126/science. 745–756. https://doi.org/10.1016/J.MARPOLBUL.2018.05.006.
1094559. Watts, A.J.R., Porter, A., Hembrow, N., Sharpe, J., Galloway, T.S., Lewis, C., 2017.
Penca, J., 2018. European Plastics Strategy: what promise for global marine litter?. Through the sands of time: beach litter trends from nine cleaned north cornish
Mar. Policy 97, 197–201. https://doi.org/10.1016/J.MARPOL.2018.06.004. beaches. Environ. Pollut. https://doi.org/10.1016/j.envpol.2017.05.016.
UN