AI-driven remote sensing enhances Mediterranean

seagrass monitoring and conservation to combat

climate change and anthropogenic impacts
Masuma Chowdhury (  masuma.chowdhury@quasarsr.com )
Quasar Science Resources, S.L.
Alejo Martínez-Sansigre
Quasar Science Resources, S.L.
Maruška Mole
Quasar Science Resources, S.L.
Eduardo Alonso Peleato
Quasar Science Resources, S.L.
Nadiia Basos
Quasar Science Resources, S.L.
Jose Manuel Blanco
Quasar Science Resources, S.L.
Maria Ramirez
Quasar Science Resources, S.L.
Isabel Caballero
Institute of Marine Sciences of Andalusia, Spanish National Research Council (ICMAN-CSIC)
Ignacio de la Calle
Quasar Science Resources, S.L.

Article

Keywords:

Posted Date: September 15th, 2023

DOI: https://doi.org/10.21203/rs.3.rs-3304270/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Page 1/19
Abstract
Seagrasses are undergoing widespread loss due to anthropogenic pressure and climate change. Since
1960, the Mediterranean seascape lost 13–50% of the areal extent of its dominant and endemic
seagrass- Posidonia oceanica, which regulates its ecosystem. Many conservation and restoration
projects failed due to poor site selection and lack of long-term monitoring. Here, we present a fast and
efficient operational approach based on a deep-learning artificial intelligence model using Sentinel-2 data
to map the spatial extent of the meadows, enabling short and long-term monitoring, and identifying the
impacts of natural and human-induced stressors and changes at different timescales. We apply ACOLITE
atmospheric correction to the satellite data and use the output to train the model along with the ancillary
data and map the extent of the meadows. We apply noise-removing filters to enhance the map quality.
We obtain 74–92% of overall accuracy, 72–91% of user’s accuracy, and 81–92% of producer’s accuracy,
where high accuracies are observed at 0-25m depth. Our model is easily adaptable to other regions and
can produce maps in in-situ data-scarce regions, providing a first-hand overview. Our approach can help
restoration practitioners, conservationists and ecosystem managers to make rational decisions to protect
this species and promote sustainability.

Introduction
Seagrasses, the submersed marine flowering plants, are undergoing substantial decline due to direct and
indirect anthropogenic activities and climate change. The recent publication1 estimated the global
seagrass area to date as 160,387 to 266,562 square km, which is significantly lower than the previously
reported global seagrass spatial distribution ranging from 177,000 to 600,000 square km2,3. Their loss
rates accelerate from 0.9% yr− 1 in the 1940s to 7% yr− 1 toward the end of the 20th century, making them
one of the most threatened ecosystems on Earth4,5. Since 1960, the Mediterranean seascape lost 13–
50% of the areal extent of its most common and endemic Posidonia oceanica seagrass meadows6. If the
recent heat waves in the Mediterranean Sea continues, this seagrass may disappear within 100 to 150
years7. Recently, efforts to protect P. oceanica have led to the creation of the Mediterranean Posidonia
Network (MPN, https://medposidonianetwork.com/) including more than 50 stakeholders from 10
countries, which seeks to effectively protect 100% of P. oceanica by 2030 by introducing innovative tools
and raising awareness at the local, national, and international levels about the importance of this
species8. However, the lack of a multi-temporal monitoring system over large areas, and identifying the
hotspot to mitigate the rapid degradation caused by anthropogenic pressures are becoming a major
concern for Mediterranean coastal managers, affecting the conservation program to protect this species.

P. oceanica covers 50,000 square km of coastal sandy and rocky areas consisting of 25% of the
Mediterranean Sea bottom at 0 to 40 m depth9,10. The meadows are home to about 20% of
Mediterranean sea life species and can accommodate up to 350 animal species per hectare11–13. They
contribute to different coastal processes like sediment deposition, and attenuate currents and wave
energy14. They are globally significant carbon sinks with greater organic carbon density compared to

Page 2/19
estuarine mangroves, peatlands and tropical forests15,16, thus play a key role in climate change
mitigation17. However, the meadows are susceptible to increasing heat waves and climate change18–20.
Besides, the meadows are under substantial threat by different anthropogenic activities, i.e., 22% of
construction of coastal infrastructure, 23% of water pollution, 14% of invasive species, 18% of Fishing, 5%
of shipping and 18% of modifications of marine currents and hydrography21 (Fig. 1). The European Union
(EU) Habitats Directive 92/43/EEC has listed P. oceanica meadows as a priority habitat. P. oceanica is
also a protected species in the Natura2000 networks within the EU because of its significant influences
on the surrounding biological, biogeochemical and physical processes in the littoral22.

A successful conservation programme for P. oceanica involves knowledge about the spatial extent of the
meadows, understanding processes of degradation23; identifying causes of degradation24; designing and
implementing effective monitoring programmes to manage, restore, or create these seagrass
meadows25,26. This demands fast, cost-effective and validated methods to observe and monitor the
present state, spatial pattern, and dynamics at appropriate spatial and temporal scales27,28. Multiple
studies29–38 presented the potential of remote sensing data in mapping the extent of this seagrass
species, complementing the traditional in-situ monitoring system. However, they were mostly for proof of
concept to utilize satellite-derived data and hardly capable of supporting the continuous monitoring
system since they are computationally expensive, and require validated methods that can map across
large areas with adequate accuracy and precision.

With growing archives of freely accessible earth observation data of adequate spatial and temporal
resolution, along with fast-growing computational processing capabilities, we now have the potential
tools to support mapping and monitoring of P. oceanica over large regions. To our knowledge, there is no
use of remote sensing techniques coupled with deep-learning-based artificial intelligence, integrated
within a pipeline technique that facilitates reproducibility and scalability to monitor the multi-temporal
spatial extent of this seagrass species, thus supporting the conservation activity. In this study, we present
an operational approach aiming to fill this gap. Our approach corresponds to a Scientific Exploitation
Platform (SEP) dedicated to the extraction and processing of Earth observation (EO) products by adding
one more layer on top of the existing tools to take the exploitation of the raw satellite data one step
further. The proposed pipeline used Sentinel-2 data for the artificial intelligence deep-learning neural
network (DLNN) to produce multi-temporal maps of P. oceanica. Here, we present the results for four
major islands of the Balearic Islands, i.e. Menorca, Mallorca, Ibiza, and Formentera in Spain, and the
Maltese Islands (Fig. 2). For these regions, we have trained the DLNN model and evaluated its
performance under the scheme of long-term monitoring and evaluation to support conservation and
ecosystem management to minimize climate change and anthropogenic impacts. This study should be
of interest to scientists, conservationists and/or coastal managers, who are responsible for the
challenging short-term management and long-term policy decisions to protect this type of natural
resources.

Results and Discussion

Page 3/19
Maps of P. oceanica

We have detected 577.29 and 74.27 square km of P. oceanica respectively in the Balearic Islands and the
Maltese Islands during 2021 (Fig. 3). Our proposed method is based on pixel-by-pixel classification
schemes, and the final output map has 10m spatial resolution. We have stacked 3–9 atmospherically
corrected clear Sentinel-2 level-2 images per tile (Supplementary Table S1) and applied the DLNN model
with the corresponding in-situ and bathymetry data. The yearly stacked image reduced the potential
misclassification due to non-persistent and seasonal variations in algal growth that may have similar
signals. It was evident that depending on the region, our proposed approach was capable of mapping P.
oceanica with 74–92% of overall accuracy, 72–91% of user’s accuracy, and 81–92% producer’s accuracy
(Table 1). When a satellite pixel contained a region with sparsely distributed or unhealthy P. oceanica
(according to the in-situ data), it was classified as false negative (FN) due to weak spectral signal. In
contrast, when majority parts of the pixel contained healthy P. oceanica, it was classified as false
positives (FP) due to strong spectral signal. However, in the north and northeast parts of Ibiza, we have
observed swell noise in all the available Sentinel-2 images that caused residual noise in the yearly
stacked image and resulted in misclassification. In contrast, we have noticed a different scenario in
Malta. In the southeastern part of Malta, we observed the presence of spectral signals coming through
the water column, which were identified by the DLNN model as P. oceanica. In this context, there could be
two possible explanations, i.e., incomplete in-situ database, where the identified patches of P. oceanica
were not included; consequently, it was depicted as FP during accuracy assessment; or, the presence of
another seagrass species with similar signals. It should be noted that the in-situ data used in this study
came from several years of field campaigns and might not be up-to-date in every region. As a result, in
some places, the FP and FN identified in the final maps could be a true presentation of the current
scenario and stressed the need for further investigation.

Table 1
DLNN model performance in regional mapping of P. oceanica during 2021
Location Overall Accuracy (%) User’s accuracy (%) Producer’s accuracy (%)

Formentera 92 91 91

Ibiza 78 72 88

Mallorca 88 84 81

Menorca 86 83 86
Maltese Island 74 74 92

Our model obtained a higher user’s and producer’s accuracy at 0-25m depth (Table 2), which was the
most representative depth limit of the seabed detection using Sentinel-2 satellite in the Mediterranean
Sea36,39. As the water was transparent and Sentinel-2 could see accurately enough, the model showed
higher accuracy at this depth. With increasing depth and light attenuation, there could still be some P.

Page 4/19
oceanica living on the sea bottom up to 40m depth40,41 but the satellite could not see them accurately
enough. This explained the model’s performance with comparatively low accuracies at that depth.

Table 2
DLNN model performance by depth in regional mapping of P. oceanica during 2021
Menorca Mallorca Ibiza Formentera Maltese
Islands

Statistics 0- 25- 0- 25- 0- 25- 0- 25- 0- 25-

Depth 25m 40m 25m 40m 25m 40m 25m 40m 25m 40m

Overall Accuracy 86 86 86 90 85 70 93 90 77 71
(%)

User’s accuracy 86 79 85 81 81 56 93 88 77 71
(%)

Producer’s 90 82 86 69 95 76 96 83 92 92
accuracy (%)

One of the advantages of our proposed method is the classification scheme. We trained the DLNN model
to learn how the colour of the sea bottom with and without P. oceanica varied with depth. Besides, we
designed the DLNN model for pixel-by-pixel classification per tile. As a result, our model was easily
adaptable to other regions. We obtained accurate and reliable results while training the DLNN model with
in-situ data from the target region, however, without in-situ data, the model was still capable of producing
reliable maps which could provide a first-hand overview. To apprehend this, we reproduced the map of P.
oceanica in the Maltese Island using the training dataset from Formentera (Fig. 4). The final output
showed an overall accuracy of 64% with a user’s accuracy of 73% and a producer’s accuracy of 69%. As
expected, the model performance was lower than what we observed in Table 1, however, it was still
scientifically acceptable and could provide a general idea of the spatial extent of the meadows. It was
evident that in both cases (Figs. 3i-j and 4) the model identified P. oceanica in the southeastern part of
Malta, implying strong evidence of the presence of the seagrass meadows.

Yearly monitoring of P. oceanica and its change assessment

We applied our proposed approach to perform a time-series analysis (Fig. 5 and Supplementary Table S2)
in Formentera, as this region was highly renowned for conservation activities. While the seagrass patches
mostly looked stable during 2017–2021, we detected changes in the northern and northeastern parts of
Formentera. Especially, in the northern part of the island, we detected a loss of P. oceanica meadows until
2019 and then a subsequent gain after 2020. The change assessment between 2017 and 2021 showed
that the area of the meadow increased by about 8 square km and decreased by about 1 square km, with a
net gain of 7 square km, which might be a result of the current conservation activities. This showed the
promising capabilities of our proposed approach to support monitoring of the meadows in the short and
long term.

Page 5/19
Application in conservation and ecosystem management
In the recent past, the concept of natural recolonization (cutting and seedlings) has been introduced to
restore P. oceanica meadows in the Mediterranean basin. However, most of them failed due to either
choosing the wrong site where the species had never been before or a lack of long-term monitoring
system due to economic constraints42. In addition, conservation and ecosystem management demand
explicit and up-to-date knowledge to deal with real-time situations. Timely and reliable maps with a large
spatial and temporal coverage can provide substantial ground in capacity building and making decisions.
Unlike the available methodologies that were resource expensive and developed for a local scale, our
proposed method can generate maps of P. oceanica with adequate accuracy in a cost-effective manner.
This can add value to the current restoration programmes by identifying suitable sites and monitoring
them regularly for short-term goals and long-term evaluation. Besides, the information generated through
our approach can support conservationists and ecosystem managers in making rational decisions to
promote sustainability.

In this context, we present an example in Fig. 6. The figure demonstrates boat pressure over the seagrass
patches in Formentera and Ibiza during June-August 2021, a peak tourist season. Each hexagon mesh
presents the number of ships detected using Sentinel-1 satellite data. It was evident that the channel
between Formentera and Ibiza had the highest pressure within this period, thus showing the promising
capabilities to identify the hotspot of P. oceanica for conservation and sustainable coastal resource
management.

This study facilitates the continuous and long-term monitoring of P. oceanica in the Mediterranean Sea in
an efficient manner. Our approach can be used in data-scarce region, with environmental and other
constraints that make it impossible to collect in situ data, thus providing a first-hand overview. Besides,
our approach can be a stepping-stone for recording the long-term impact of climate change as well as
different environmental drivers and stressors on the seagrass meadows following a standardized and
uniform method, and ameliorate understanding of the connectivity between oceanography and seagrass
ecosystem, spatial and temporal management and protection, and finally mapping other benthic
ecosystems for conservation and management.

Methods
Scientific exploitation platform
To map P. oceanica in a fast, reliable and efficient way, we developed an operational pipeline
(Supplementary Figure S1) within a SEP. SEP integrated the hardware/software infrastructures that
supplied the computing and storage resources needed for the exploitation and provision of the tools to
manage EO datasets in a distributed environment. The platform comprised a Kubernetes cluster with
many nodes hosted on-premises. The pipeline consisted of several scientific modules, such as data
acquisition, data pre-processing, followed by artificial intelligence-based DLNN, and post-processing.
Page 6/19
Within the SEP, all these modules could be executed isolated or sequentially within Docker containers,
encapsulating the entire environment required for the scientific workflow. This ensured scalability and the
pipeline’s results could easily be reproduced with consistent outcomes. The pipeline also integrated a
quality control system where a Network-file-system (NFS) volume held the raw data and the processed
products, which could later be downloaded and analysed within the SEP web interface.

Data acquisition
Sentinel-2 Level-1C data were downloaded from Google Cloud Storage using their Application
Programming Interface (API). This data corresponded to the top of the atmospheric reflectance (TOA) for
7 Sentinel-2 tiles covering the Balearic Islands, and 1 tile covering the Maltese Islands (Supplementary
Table S1). After visual inspection, images with high cloud coverage or sun glint effects were discarded
from further processing. While mapping P. oceanica, bathymetry data plays a significant role, as this
species is highly depth-dependent. In this study, the freely available bathymetry data at 100m spatial
resolution were obtained from the European Marine Observation and Data Network (EMODnet,
https://www.emodnet-bathymetry.eu/). Besides, the in-situ data were collected from the local government
of the Balearic Islands and the Marine Database of Environment and Resources Authority (ERA) in Malta
(https://meps.eraportal.org.mt/) to develop and train the DLNN model. These data corresponded to single
cartography in the GIS layer combining several years of field campaigns during 2008–2019 for the
Balearic Islands and 2017–2019 for the Maltese Islands. The data from the Balearic Islands
(https://atlasposidonia.com/en/home/) had information for different marine habitats following the
nomenclature and coding of the Standard List of Marine Habitats of Spain (LPHME), whereas the data
from the Maltese Islands only had information regarding the habitat of P. oceanica.

Data pre-processing
Pre-processing is a crucial step while producing benthic maps using satellite data for artificial
intelligence. It is even more important while using Sentinel-2 data for aquatic applications as it has low-
signal-to-noise-ratio that requires a good atmospheric correction (AC) to separate the top-of-atmospheric
(TOA) reflectance observed by the satellite into the signal from the atmosphere and the signal from the
surface (bottom-of-atmosphere, BOA)43. In this study, we applied the ACOLITE (v2020) AC processor
developed by the Royal Belgian Institute of Natural Sciences (RBINS). Based on the literature and the
analysis results, we chose ACOLITE because it better reproduced the shape of the reflectance spectra
than the other widely used AC processors, i.e., C2RCC (a SNAP plugin). As ACOLITE processor is image-
based, it does not require any in-situ measurements. By default, ACOLITE performs AC using the ‘dark
spectrum fitting’ (DSF) algorithm for aerosol corrections. However, this algorithm avoids severe glints for
which the glint effects can still be found in the DSF-derived surface reflectance44. As sun glint was a
prominent optical phenomenon in the study region, especially during spring and summertime, the
optional glint correction within ACOLITE was also applied. The land and clouds were masked during the
AC processing to retain the water-leaving reflectance. Therefore, the data were inspected visually. In some
cases where high clouds (i.e. cirrus, cirrostratus, and cirrocumulus) and haze were left unmasked, and
when these were found to degrade the region of interest near the coast, the images were discarded.
Page 7/19
Occasionally, images with clear meteorological conditions were also discarded after visual inspection due
to significant turbidity being visible in the shallow waters. This was done as it significantly affected the
colour of the water and made it impossible to observe through the water column. These strict criteria only
admitted images that were of very good quality. To remove boats, very large waves, and other anomalies,
as well as to reduce noisy patterns and artifacts to improve the overall quality of the image, a pixel-by-
pixel median stacking was performed for the blue (B), green (G), and red (R) bands of all the good quality
atmospherically processed data (3–9 images per year).

The EMODnet bathymetry data was reprojected and resampled to 10m (using the bilinear interpolation
technique) to match the projection system and the spatial resolution of the Sentinel-2 data. The data was
cropped after 40m depth to remove open and deep ocean pixels to avoid potential FP of P. oceanica, as
both classes had similar spectral values36.

The marine habitats of the in-situ data were classified into two classes, i.e. "PO" (indicating the habitat of
P. oceanica) and "Non-PO" (indicating other than the habitat of P. oceanica). In-situ data corresponding to
land, offshore, and shallow water with a sandy/rocky habitat or other seagrass habitats were included in
the Non-PO class. The PO and Non-PO classes were labelled as 1 and 0, respectively, and rasterized using
the same projection system of Sentinel-2 scenes.

Artificial intelligence-based DLNN approach

The identification of P. oceanica is a binary classification problem where pixels correspond to either PO or
Non-PO. Therefore, the aim of using a DLNN model was to train a function based on the given satellite
and bathymetry dataset (X) as well as target in-situ data (Y) to output a predicted class (Ŷ) of PO and
Non-PO, which should match as closely as possible with the in-situ data (Y).

The training was done through two processes, namely, forward propagation and backpropagation to
recognize patterns based on the given dataset. In forward propagation, the process moved in the forward
direction through the neural network (NN) to produce a final value at the output layer, whereas at the
same time, in backpropagation, the process moved from the output layer to the input layer to improve the
weight (w) value generated in the forward process. Forward propagation was a value-processing step
from the input layer that involved calculating a linear combination per neuron using adjustable linear
coefficients and then calculating an activation function by performing a non-linear transformation for the
corresponding neuron within the layers. The activation functions of all the layers except the last one were
simply rectified linear unit activations, meaning that the output was set to zero if it was negative. Hence,
the output coefficients from the neurons never had negative values. In the DLNN model, the last layer was
a binary classifier, which means that it classified pixels as belonging to either class 1 or class 0. In this
case, the activation function was a sigmoid function, which predicted the probability (Ŷ) of a pixel being
class 1 or 0 based on a threshold value of 0.5.

To optimize the training process, we performed mini-batch processing, where we split the training set into
smaller sets and implemented gradient descent on each batch chronologically to make the algorithm

Page 8/19
work faster. Therefore, based on the input data, each 10m pixel of the Sentinel-2 image was classified as
containing PO (class 1) or Non-PO (class 0). The result, Ŷ, was compared to Y, and the difference was
quantified in a cost function. Minimizing the cost function equates to making Ŷ as similar as possible to
Y. The differences between Ŷ and Y per pixel allowed the calculation of the gradient that quantified how
much each coefficient in each neuron must be altered to match Ŷ to Y. This gradient was then used to
update the coefficients in the neurons of all the layers through backpropagation. This was an iterative
process, and therefore, at each iteration, a new prediction was made, compared to the target Y, a gradient
was calculated, and the coefficients were updated. This process was carried out continuously until it
produced the smallest loss value, and the output was saved in the form of a file containing the weight (w)
and bias (b) values. Afterward, the derived set of coefficients from the DLNN model was used to map P.
oceanica.
Post-processing and accuracy assessment
In the post-processing step, we evaluated several filters, i.e. sieve (replacing small isolated groups of
pixels with the pixel value of the largest neighbour), median (ranking values of the pixels in the moving
window with a specified radius and taking the middle-ranking value), bilateral (replacing the value of each
pixel with a weighted average of the pixel values in the moving window), and principal component
analysis (removing noise by discarding the last principal components that explain the least of variance).
Among them, we chose the median filter with 3x3 pixels moving window as it significantly reduced the
noise while preserving the shape of the seagrass patches. We applied the filter to the binary maps and
merged the resultant maps per region to produce the regional map. Then we calculated the overall,
producer’s, and user’s accuracies of the final map using the corresponding geo-referenced in-situ data.
The overall accuracy expressed the ratio of the number of correctly classified pixels to the total number of
given pixels regardless of the class. Producer’s accuracy expressed how often P. oceanica could be
identified correctly on the classified map; whereas, the user’s accuracy expressed how often the P.
oceanica class on the map would be present on the ground. Producer’s accuracy is the map accuracy
from the mapmaker’s viewpoint and so it is a great statistical metric for the remote sensing scientist
creating the habitat map. In contrast, the user’s accuracy is the accuracy from the map user’s viewpoint.
Hence, it is more significant in a management context of a given region as it reports a quantitative
probability for the actual presence of the studied habitat (in this case, P. oceanica) in the given region37.
The equations for all the aforementioned classification metrics are given below:

T P +T N
Overall accuracy = n
(Eq. 1)

Producer’s accuracy = (Eq. 2)

TP

T P +F N

User’s accuracy = (Eq. 3)

TP

T P +F P

Here, TP is the number of true positives, i.e., pixels correctly identified as PO; TN is the number of true
negatives, i.e., pixels correctly identified as Non-PO; FP is the number of false positives, i.e., the number of

Page 9/19
actual Non-PO pixels classified as PO; and FN is the number of false negatives, i.e., the number of actual
PO pixels classified as Non-PO.

Declarations
Acknowledgements
This study is a part of SIMBAD project (QSR-ESABIC-2018-001) incubated by the European Space Agency
Business Incubation Centre (ESA-BIC) Madrid Region, which facilitates incentives for business projects
and start-ups to use space technologies or to develop applications based on these technologies with the
aim of developing new products and services unrelated to space. The authors are thankful to the support
from the Parque Científico de Madrid where Quasar Science Resources, S.L. is incubated. The first author
is supported by the Industrial Doctorate Program of the Spanish Ministerio de Ciencia e Innovación (ref.
DIN2020-010979/AEI/10.13039/501100011033). This work is part of the first author’s PhD within the
SIMBAD project. The authors would like to thank the rest of the SIMBAD team. The authors are thankful
to the European Space Agency, the European Commission, and the Copernicus programme for
distributing Sentinel-2 imagery. This research would not be possible without the in-situ data kindly
provided by ERA and Govierno Balear. The in-situ data for the Maltese Islands was collected within the
scope of the Marine Environmental Monitoring: Towards Effective Management of Malta’s Marine Waters
project (ERA 2017). Details on the survey performed can be found in Dewey and MacMillan (2021). We
would like to thank Marcial Bardolet Richter (Técnico de IBANAT. Consellería de Medi Ambient i Territori),
from D.G. Biodiversidad y Espacios Naturales, Consellería de Medi Ambient i Territori de les Illas Baleares,
Govern Balear, for providing the in-situ data for the Balearic Islands region.

Authors’ contributions
MC, AM, MM, NB, IC, MR and IdC contributed with ideas to design the research objectives. AM led the
design of the architecture of the scientific modules. MC and AM developed the scientific modules for data
pre-processing, the DLNN model and its accuracy assessment, whereas NB developed the filters and time-
series analysis, with support from IdC. EP and JB contributed to the SEP’s development. MM conducted
all the model runs within the Docker environment for the operational approach, with support from EP. MC
wrote the original manuscript with input from MM. All authors contributed to the interpretation and
discussion of results and the final manuscript. IdC and JB provided part of the financial support leading
to this publication.

Data Availability
The necessary procedures to generate the data and reproduce the methodology have been outlined in the
manuscript. The data that support the findings of this study are available from the authors on reasonable
request.

Page 10/19
Competing Interest
The author(s) declare no competing interests.

References
1. McKenzie, L. J. et al. The global distribution of seagrass meadows. Environ. Res. Lett. 15, 074041
(2020).
2. Duarte, C. M. et al. Seagrass community metabolism: Assessing the carbon sink capacity of
seagrass meadows. Glob. Biogeochem. Cycles 24, (2010).
3. Short, F. T. World Atlas of Seagrasses. (University of California Press, 2003).
4. de los Santos, C. B. et al. Recent trend reversal for declining European seagrass meadows. Nat.
Commun. 10, 3356 (2019).
5. Waycott, M. et al. Accelerating loss of seagrasses across the globe threatens coastal ecosystems.
Proc. Natl. Acad. Sci. 106, 12377–12381 (2009).
6. Marbà, N., Díaz-Almela, E. & Duarte, C. M. Mediterranean seagrass (Posidonia oceanica) loss
between 1842 and 2009. Biol. Conserv. 176, 183–190 (2014).
7. Ibiza News. Posidonia Will Disappear From Ibiza And Formentera If Sea Temperatures Continue To
Rise. Ibiza News https://www.diariodeibiza.com/ibiza/posidonia-sea-temperatures-rise/ (2023).
8. MPN. Mediterranean Posidonia Network. Mediterranean Posidonia Network
https://medposidonianetwork.com/ (2023).
9. Bethoux, J. & Copin Montegut, G. BIOLOGICAL FIXATION OF ATMOSPHERIC NITROGEN IN THE
MEDITERRANEAN-SEA. Limnol. Oceanogr. 31, 1353–1358 (1986).
10. MedWet. Posidonia, the lung of the Mediterranean | MedWet.
https://medwet.org/2017/10/mediterranean-posidonia/ (2017).
11. Calvo, S., Ciraolo, G. & Loggia, G. L. Monitoring Posidonia oceanica meadows in a Mediterranean
coastal lagoon (Stagnone, Italy) by means of neural network and ISODATA classification methods.
Int. J. Remote Sens. 24, 2703–2716 (2003).
12. Francour, P. Fish Assemblages of Posidonia oceanica Beds at Port-Cros (France, NW Mediterranean):
Assessment of Composition and Long‐Term Fluctuations by Visual Census. Mar. Ecol. 18, 157–173
(1997).
13. Hemminga, M. A. & Duarte, C. M. Seagrass Ecology. (Cambridge University Press, 2000).
14. Gacia, E., Granata, T. C. & Duarte, C. M. An approach to measurement of particle flux and sediment
retention within seagrass (Posidonia oceanica) meadows. Aquat. Bot. 65, 255 (1999).
15. Fourqurean, J. W. et al. Seagrass ecosystems as a globally significant carbon stock. Nat. Geosci. 5,
505–509 (2012).
16. IUCN. Manual for the Creation of Blue Carbon Projects in Europe and the Mediterranean. at
https://iucn.org/resources/file/manual-creation-blue-carbon-projects-europe-and-mediterranean
Page 11/19
 (2021).
17. Duarte, C. M., Losada, I. J., Hendriks, I. E., Mazarrasa, I. & Marbà, N. The role of coastal plant
communities for climate change mitigation and adaptation. Nat. Clim. Change 3, 961–968 (2013).
18. Arias-Ortiz, A. et al. A marine heatwave drives massive losses from the world’s largest seagrass
carbon stocks. Nat. Clim. Change 8, 338–344 (2018).
19. Jordà, G., Marbà, N. & Duarte, C. M. Mediterranean seagrass vulnerable to regional climate warming.
Nat. Clim. Change 2, 821–824 (2012).
20. Wernberg, T. et al. An extreme climatic event alters marine ecosystem structure in a global
biodiversity hotspot. Nat. Clim. Change 3, 78–82 (2013).
21. EEA. Grouped threats to Posidonia oceanica beds as reported by Mediterranean EU Member States
under the EU Habitats Directive — European Environment Agency. https://www.eea.europa.eu/data-
and-maps/figures/grouped-threats-to-posidonia-oceanica (2015).
22. Mateo, M. A., Romeo, J., Perez, M., Littler, M. M. & Littler, D. S. Dynamics of millenary organic deposits
resulting from the growth of the Mediterranean seagrass < I > Posidonia oceanica.</I>. (1997).
23. Lee, K.-S., Park, S. & Kim, Y. Effects of irradiance, temperature, and nutrients on growth dynamics of
seagrasses: A review. J. Exp. Mar. Biol. Ecol. - J EXP MAR BIOL ECOL 350, 144–175 (2007).
24. Collado-Vides, L., Caccia, V. G., Boyer, J. N. & Fourqurean, J. W. Tropical seagrass-associated
macroalgae distributions and trends relative to water quality. Estuar. Coast. Shelf Sci. 73, 680–694
(2007).
25. Duarte, C. M. The future of seagrass meadows. Environ. Conserv. 29, 192–206 (2002).
26. Short, F., Carruthers, T., Dennison, W. & Waycott, M. Global seagrass distribution and diversity: A
bioregional model. J. Exp. Mar. Biol. Ecol. 350, 3–20 (2007).
27. Kenworthy, W., Wyllie-Echeverria, S., Coles, R., Pergent, G. & Pergent-Martini, C. Seagrass conservation
biology: an interdisciplinary science for protection of the seagrass biome. Seagrasses Biol. Ecol.
Conserv. (2006).
28. Sagawa, T. et al. Using bottom surface reflectance to map coastal marine areas: a new application
method for Lyzenga’s model. Int. J. Remote Sens. 31, 3051–3064 (2010).
29. Borfecchia, F. et al. Landsat 8 OLI satellite data for mapping of the Posidonia oceanica and benthic
habitats of coastal ecosystems. Int. J. Remote Sens. 40, 1548–1575 (2019).
30. Cozza, R. et al. Biomonitoring of Posidonia oceanica beds by a multiscale approach. Aquat. Bot. 156,
14–24 (2019).
31. Dattola, L. et al. Comparison of Sentinel-2 and Landsat-8 OLI satellite images vs. high spatial
resolution images (MIVIS and WorldView-2) for mapping Posidonia oceanica meadows. in Remote
Sensing of the Ocean, Sea Ice, Coastal Waters, and Large Water Regions 2018 vol. 10784 252–262
(SPIE, 2018).
32. Fornes, A. et al. Mapping Posidonia oceanica from IKONOS. ISPRS J. Photogramm. Remote Sens.
60, 315–322 (2006).

Page 12/19
33. Matarrese, R., Acquaro, M., Morea, A., Tijani, K. & Chiaradia, M. T. Applications of Remote Sensing
Techniques for Mapping Posidonia Oceanica Meadows. in IGARSS 2008–2008 IEEE International
Geoscience and Remote Sensing Symposium vol. 4 IV-906-IV–909 (2008).
34. Pasqualini, V., Pergent-Martini, C., Clabaut, P. & Pergent, G. Mapping ofPosidonia oceanicausing
Aerial Photographs and Side Scan Sonar: Application off the Island of Corsica (France). Estuar.
Coast. Shelf Sci. 47, 359–367 (1998).
35. Topouzelis, K., Makri, D., Stoupas, N., Papakonstantinou, A. & Katsanevakis, S. Seagrass mapping in
Greek territorial waters using Landsat-8 satellite images. Int. J. Appl. Earth Obs. Geoinformation 67,
98–113 (2018).
36. Traganos, D. et al. Spatially Explicit Seagrass Extent Mapping Across the Entire Mediterranean. Front.
Mar. Sci. 9, (2022).
37. Traganos, D. & Reinartz, P. Mapping Mediterranean seagrasses with Sentinel-2 imagery. Mar. Pollut.
Bull. 134, 197–209 (2018).
38. Traganos, D. & Reinartz, P. Machine learning-based retrieval of benthic reflectance and Posidonia
oceanica seagrass extent using a semi-analytical inversion of Sentinel-2 satellite data. Int. J. Remote
Sens. 39, 9428–9452 (2018).
39. Poursanidis, D., Traganos, D., Reinartz, P. & Chrysoulakis, N. On the use of Sentinel-2 for coastal
habitat mapping and satellite-derived bathymetry estimation using downscaled coastal aerosol
band. Int. J. Appl. Earth Obs. Geoinformation 80, 58–70 (2019).
40. Colantoni, P., Gallignani, P., Fresi, E. & Cinelli, F. Patterns of Posidonia oceanica (L.) Delile Beds
around the Island of Ischia (Gulf of Naples) and in Adjacent Waters. Mar. Ecol. 3, 53–74 (1982).
41. Meinesz, A. & Laurent, R. Cartographie et état de la limite inferieure de l’herbier de Posidonia
oceanica dans les Alpes-maritimes (France) – Campagne Poseïdon 1976 —. Bot. Mar. 21, 513–526
(1978).
42. Boudouresque, C.-F., Blanfuné, A., Pergent, G. & Thibaut, T. Restoration of Seagrass Meadows in the
Mediterranean Sea: A Critical Review of Effectiveness and Ethical Issues. Water 13, 1034 (2021).
43. Vanhellmont, Q. & Ruddick, K. Acolite for Sentinel-2: Aquatic Applications of MSI Imagery. (2016).
44. Vanhellemont, Q. & Ruddick, K. Atmospheric correction of metre-scale optical satellite data for inland
and coastal water applications. Remote Sens. Environ. 216, 586–597 (2018).

Figures

Page 13/19
Figure 1

Why conservation of P. oceanica is important?

Conceptual diagram depicting the ecosystem services of P. oceanica (outlined by blue lines) and the
anthropogenic pressures on the meadows (outlined by red lines). The meadows are endemic and
dominant seagrass species in the Mediterranean Sea that controls its coastal and aquatic ecosystem.
They provide shelter as well as act as a nursery for thousands of marine animals and plants, several of
which are endangered species such as sea turtles and dugongs. The plant itself oxygenates the water
and stabilizes the sandy shores and sea beds. They provide food directly through grazing or indirectly
through detritus cycle9. They create ecological corridors between different habitats and are also globally
significant carbon sinks13. The meadows are threatened by different anthropogenic activities, i.e. coastal
infrastructure development, water pollution, fishing, shipping and anchoring. Over the last 50 years, the
Mediterranean basin lost 13-50% of the areal extent of this meadow4. The recent warming of the
Mediterranean Sea exaggerates the situation even more, which calls for continuous monitoring of the

Page 14/19
meadows, identifying potential areas to mitigate the human impacts as well as selection of sites for
conservation and restoration.

Figure 2

Study regions

The study regions include four Islands, namely Menorca, Mallorca, Ibiza and Formentera from the
Balearic Islands in Spain, and the Maltese Islands located in the western and the central Mediterranean
Sea, respectively. The continental shelf of the Balearic Islands is nearly horizontal and situated at 93m
depth on an average. The coastal waters are highly transparent due to the absence of rivers and the
infiltration of rainfall in the karstic soils. Currently, 73% of the Balearic coastline is under protection
through the Natural Areas of Special Interest, National Natural Park, Natura2000 sites, and the UNESCO
Biosphere Reserve. In contrast, the Maltese Islands consist of shallow water, and the surrounding seafloor
comprises an irregularly shaped continental shelf. Sandy bottoms with P. oceanica and C. nodosa

Page 15/19
characterize the dominant benthic habitats in shallow waters, mainly the offshore east coast of the
Maltese archipelago.

Figure 3

Regional maps of P. oceanica meadows

Regional maps of P. oceanica obtained through training the DLNN model with the corresponding in-situ
data for the Balearic and Maltese Islands. The maps have 10m spatial resolution. Each pixel was
identified as either PO or Non PO by the DLNN model. Subplots a, c, e, g, and i demonstrate the spatial
extent of the meadows identified by the DLNN model in Menorca, Mallorca, Ibiza, Formentera, and the
Maltese Islands, respectively. Subplots b, d, f, h, and j demonstrate model’s performance per pixel for the
corresponding regions. TN is True Negatives (pixels correctly identified as Non-PO), TP is True Positives
(pixels correctly identified as PO), FP is False Positives (Non-PO pixels classified as PO), and FN is False
Negatives (PO pixels classified as Non-PO).

Page 16/19
Figure 4

DLNN model trained with Formentera in-situ data

Map of P. oceanica in the Maltese Islands using the training dataset from Formentera. The maps have
10m spatial resolution. Each pixel was identified as either PO or Non PO by the DLNN model. Subplot a
demonstrates the spatial extent of the meadows identified by the DLNN model, whereas subplot b
demonstrates the model’s performance per pixel for the Islands. TN is True Negatives (pixels correctly
identified as Non-PO), TP is True Positives (pixels correctly identified as PO), FP is False Positives (Non-
PO pixels classified as PO), and FN is False Negatives (PO pixels classified as Non-PO).

Page 17/19
Figure 5

Yearly monitoring of P. oceanica and its change assessment

Yearly maps of P. oceanica in Formentera during 2017-2021. The seagrass patches mostly look stable
during this period, except the northern part of the Island, where we detected a loss of the meadows until
2019 and then a subsequent gain since 2020.

Page 18/19
Figure 6

Boat intensity over the P. oceanicameadows

Boat pressure over the meadows in Formentera and Ibiza during June-August 2021, a peak tourist
season. Each hexagon mesh presents the number of ships detected through Sentinel-1 satellite.

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download.

Supplementarymaterials.pdf
Page 19/19