ArticlePDF Available

Abstract and Figures

The COVID-19 pandemic has led to reduced anthropogenic pressure on ecosystems in several world areas, but resulting ecosystem responses in these areas have not been investigated. This paper presents an approach to make quick assessments of potential habitat changes in 2020 of eight marine species of commercial importance in the Adriatic Sea. Measurements from floating probes are interpolated through an advection-equation based model. The resulting distributions are then combined with species observations through an ecological niche model to estimate habitat distributions in the past years (2015–2018) at 0.1° spatial resolution. Habitat patterns over 2019 and 2020 are then extracted and explained in terms of specific environmental parameter changes. These changes are finally assessed for their potential dependency on climate change patterns and anthropogenic pressure change due to the pandemic. Our results demonstrate that the combined effect of climate change and the pandemic could have heterogeneous effects on habitat distributions: three species (Squilla mantis, Engraulis encrasicolus, and Solea solea) did not show significant niche distribution change; habitat suitability positively changed for Sepia officinalis, but negatively for Parapenaeus longirostris, due to increased temperature and decreasing dissolved oxygen (in the Adriatic) generally correlated with climate change; the combination of these trends with an average decrease in chlorophyll, probably due to the pandemic, extended the habitat distributions of Merluccius merluccius and Mullus barbatus but reduced Sardina pilchardus distribution. Although our results are based on approximated data and reliable at a macroscopic level, we present a very early insight of modifications that will possibly be observed years after the end of the pandemic when complete data will be available. Our approach is entirely based on Findable, Accessible, Interoperable, and Reusable (FAIR) data and is general enough to be used for other species and areas.
Content may be subject to copyright.
Habitat Distribution Change of Commercial Species in the1
Adriatic Sea during the COVID-19 Pandemic2
Gianpaolo Coroa,1,2,
, Pasquale Bovea, Anton Ellenbroekb
aIstituto di Scienza e Tecnologie dell’Informazione “Alessandro Faedo” – CNR, Pisa, Italy4
bFood and Agriculture Organization of the United Nations, Viale delle Terme di Caracalla, 00153 Rome,5
The COVID-19 pandemic has led to reduced anthropogenic pressure on ecosystems in
several world areas, but resulting ecosystem responses in these areas have not been inves-
tigated. This paper presents an approach to make quick assessments of potential habitat
changes in 2020 of eight marine species of commercial importance in the Adriatic Sea.
Measurements from floating probes are interpolated through an advection-equation based
model. The resulting distributions are then combined with species observations through
an ecological niche model to estimate habitat distributions in the past years (2015-2018)
at 0.1° spatial resolution. Habitat patterns over 2019 and 2020 are then extracted and ex-
plained in terms of specific environmental parameter changes. These changes are finally
assessed for their potential dependency on climate change patterns and anthropogenic
pressure change due to the pandemic. Our results demonstrate that the combined effect
of climate change and the pandemic could have heterogeneous effects on habitat distri-
butions: three species (Squilla mantis,Engraulis encrasicolus, and Solea solea) did not
show significant niche distribution change; habitat suitability positively changed for Sepia
officinalis, but negatively for Parapenaeus longirostris, due to increased temperature and
Corresponding author
Email addresses: (Gianpaolo Coro),
(Pasquale Bove), (Anton Ellenbroek)
1Telephone Number: +39 050 315 8210
2Fax Number: +39 050 621 3464
Preprint submitted to Ecological Informatics May 10, 2022
decreasing dissolved oxygen (in the Adriatic) generally correlated with climate change;
the combination of these trends with an average decrease in chlorophyll, probably due
to the pandemic, extended the habitat distributions of Merluccius merluccius and Mullus
barbatus but reduced Sardina pilchardus distribution. Although our results are based on
approximated data and reliable at a macroscopic level, we present a very early insight of
modifications that will possibly be observed years after the end of the pandemic when
complete data will be available. Our approach is entirely based on Findable, Accessible,
Interoperable, and Reusable (FAIR) data and is general enough to be used for other species
and areas.
Keywords: Ecological Niche Modelling, Marine Science, COVID-19, Conservation8
1. Introduction10
The COVID-19 pandemic has directly affected human activities in many world areas11
(Coro and Bove, 2022), but its direct and indirect effects on the ecosystems of these areas12
are still under study. The reduced anthropogenic pressure on these ecosystems may have13
been beneficial for species habitats. However, the combined effects of the pandemic and14
climate change may have triggered complex reactions. Analysing natural pattern changes15
can reveal how ecosystems have responded to general climatic trends and inter-annual cli-16
matic variations within the context of human pressure reduction in 2020. In particular,17
marine ecosystems, especially in the Adriatic Sea, have benefited from the reduction of18
stress factors such as (i) fishing and vessel traffic (Depellegrin et al., 2020), (ii) distur-19
bance of species life (Kemp et al., 2020), (iii) nutrient load in coastal areas (Adwibowo,20
2020; Mishra et al., 2020; Shehhi and Samad, 2021), and (iv) water pollution (Yunus et al.,21
2020). Understanding these benefits is interesting to quantitatively assess the peculiar ma-22
rine ecosystem dynamics modifications that occurred at various levels (e.g., pollution,23
biodiversity, and ecosystems) and how these influenced human activities (e.g., fisheries,24
ecosystem services, social interaction and mobility, and illegal activities) (Snapshot-CNR,25
2020). Understanding these dynamics allows identifying correlations that would have26
been hidden without the lockdowns and help designing novel strategies for marine re-27
source sustainability. For example, the lockdowns have allowed scientists to better model28
the resilience of Adriatic fishing fleets to activity closure, i.e., the time to return to regime29
fishing activity and market saturation (Coro et al., 2022). Moreover, the 2020 lockdown30
restrictions to fishing activities in many areas (including the Adriatic) have limited scien-31
tific survey ranges and resulted in missing survey hauls with consequent information loss32
on stock biomass in 2020. This scenario calls for solutions to estimate biomass variation in33
2020 despite the data gaps, which in turn requires information about habitat modification34
as support to expert observations, biomass estimates, and fishing catch change understand-35
ing (Brown et al., 2010; Weatherdon et al., 2016; Trifonova et al., 2017; Coro et al., 2020,36
This paper analyses the potential habitat change, in 2020, of eight marine species38
of commercial importance in the Adriatic Sea: European hake (Merluccius merluccius),39
common sole (Solea solea), mantis shrimp (Squilla mantis), red mullet (Mullus barba-40
tus), common cuttlefish (Sepia officinalis), European anchovy (Engraulis encrasicolus),41
European pilchard (Sardina pilchardus), and deep-water rose shrimp (Parapenaeus lon-42
girostris). These species are target of beam (common sole, mantis shrimp, common cut-43
tlefish), bottom (red mullet, deep-water rose shrimp, European hake), and mid-water (Eu-44
ropean anchovy and pilchard) trawlers and purse seine vessels (European anchovy and45
pilchard). They currently account for about 70% of the total catch in the basin (FAO,46
2020). The related fishing grounds range from coastal and offshore waters to deeper wa-47
ters (e.g., the Pomo Pit) (Russo et al., 2020). The high fishing stress on these species48
and most Adriatic stocks (Froese et al., 2018) makes them relevant to understand how49
the combination of reduced anthropogenic stress during the COVID-19 pandemic and cli-50
matic changes influenced their distribution in the Adriatic. The study presented in this51
paper sheds light on the magnitude of change in one year of reduced anthropogenic pres-52
sure. Additionally, it indicates the sensitivity of the species’ habitats to environmental53
change and can be used to predict the economic and ecological impact of a return to the54
pre-pandemic human activity level.55
Habitat assessment often estimates the ecological niche of a species (Jones et al., 2012;56
Coro et al., 2016a; Weber et al., 2017; Deneu et al., 2021), i.e., the set of resources and en-57
vironmental conditions that foster its persistence and proliferation in an area. It indicates58
such conditions either in the species’ native habitat (native niche) or in other geographical59
areas (potential niche). Mathematically, a species’ ecological niche is the space within60
a hyper-volume, in a vector space made up of environmental parameters, associated to61
the species’ proliferation. Ecological niche models (ENMs) both estimate the parameters62
to use in the vector space and identify the hyper-volume boundaries. As a first step, an63
ENM uses statistical analysis or machine learning to estimate a predictive function be-64
tween species observation records and specific environmental parameters. As a second65
step (projection phase), it applies the predictive function to other environmental parameter66
values that refer to a new area or other environmental scenarios (Peterson et al., 2007).67
For example, a model trained on the environmental parameters of an area in 2015 can68
be projected onto the parameter values in 2020 (Coro et al., 2018c; Coro, 2020). In the69
experiment presented in this paper, individual ENMs for the eight selected species were70
estimated for average environmental parameter values of the 2015-2018 years. Then they71
were projected onto the environmental parameters of 2020 to see if the COVID-19 re-72
lated changes influenced habitat distribution change. Furthermore, the major parameters73
driving change were checked against other studies to assess if the observed variations po-74
tentially depended on climate change (rather than inter-annual climatic variations) or the75
pandemic. Our experiment was conducted in a context of minimal environmental and76
species-occurrence data available for the pandemic period. Information extraction tech-77
niques were therefore used to estimate enough information to feed the ENMs. Pattern78
recognition was finally used to infer habitat change information over the years.79
ENMs have been used to identify suitable areas for species (Peterson, 2003; Menchetti80
et al., 2019). The generality of the approach made them adopted in early predictions of the81
potential spread of COVID-19 due to environmental and meteorological conditions, e.g.,82
they foresaw the lower summer outbreak rate of 2020 (Araujo and Naimi, 2020; Coro,83
2020). These models have demonstrated a sufficient prediction effectiveness when work-84
ing with few data, for example to predict rare species distributions (de Siqueira et al., 2009;85
Coro et al., 2013a, 2015b; Chunco et al., 2013). The possibility to process environmental86
parameters over time also makes them effective to monitor long-term habitat change (Ben87
Rais Lasram et al., 2010; Friedlaender et al., 2011; Ashraf et al., 2017; Coro et al., 2016a,88
2018c; Chala et al., 2019). ENMs commonly require uniformly distributed environmental89
parameters estimated from real observations over the study area. These distributions can90
result from hydrodynamic models based on point observations coming from satellite (Du-91
rand et al., 2010; Werdell and Bailey, 2005; Alvera-Azcárate et al., 2005) or in situ probes92
(Peterson, 2001; Huang et al., 2008; Ravdas et al., 2018; Scarponi et al., 2018). Effective93
distributions are also obtainable through lower-complexity models, based on the advec-94
tion equation that simulates the dispersion of a quantity by currents (Lipizer et al., 2014;95
Troupin et al., 2012; Djakovac et al., 2015). Parameters estimated from these models com-96
monly find applications in ecological models (Toonen and Bush, 2020; Garcia et al., 2019;97
Blackford, 2002) and ecological niche models (Coll et al., 2007; Azzolin et al., 2020).98
Accurate parameter selection is also integral to ENMs, because these models are sensitive99
to mutually-dependent variables and achieve higher performance when using independent100
variables (Pearson, 2007). A correct variable selection is typically achieved through sta-101
tistical analysis (Sánchez-Tapia et al., 2017; Guo and Liu, 2010; Muscarella et al., 2014;102
Magliozzi et al., 2019; Schnase et al., 2021) or other ENMs (Warren and Seifert, 2011;103
Coro et al., 2013a, 2015b,a; Zeng et al., 2016; Bargain et al., 2017).104
This paper proposes a workflow based on the application of ENMs to in situ environ-105
mental parameter observations and expert-verified species observations to discover habitat106
change across 2015-2018, 2019, and 2020. The 2015-2018 period was used as an aggre-107
gated and meaningful reference for average environmental conditions and species presence108
in the near past, and 2019 data were used to assess if the variations observed in 2020 were109
due to the pandemic or climate change. First, punctual environmental observations were110
transformed into uniform parameter distributions through an advection equation-based111
model. Second, parameter selection per species was conducted to feed ENMs with the112
parameters mostly associable with the species habitat (e.g., its preferred depth range and113
environmental conditions). Third, the consistency of our ENMs was verified against other114
ENMs calculated independently. Fourth, habitat variation over the years, per species, was115
studied to identify habitat change trends. Finally, these trends were explained in terms of116
environmental parameter change potentially correlated with climate change and the pan-117
demic. Our study used only a few, but reliable, environmental and species data. This118
choice was made to investigate the viability of open data and thus to only use actual obser-119
vations whose modulations contained information on the reduced anthropogenic pressure120
in 2020 due to the pandemic.121
Our analysis identified robust patterns at the Adriatic scale but cannot be considered122
punctually reliable because it is based on few data (i.e., it is a data-poor approach). Nev-123
ertheless, it offers an unprecedented possibility to shed light on the modifications that the124
combined action of the COVID-19 pandemic and climate change brought to species’ dis-125
tribution in the Adriatic Sea, way ahead of the time when data will be collected, collated,126
and analysed after the end of the pandemic. The open data approach was possible thanks127
to the recent investments by international communities on Findable, Accessible, Interop-128
erable, and Reusable (FAIR) data, Open Science, and data collection networks addressing129
the realisation of digital twins of marine systems (EU Commission, 2020b).130
2. Methods131
2.1. Data132
Our experiment used the data of the international Argo float network (Argo, 2000).133
This network includes robotic probes that drift with ocean currents while moving and134
measuring biogeochemical parameters along the water column. These probes collect en-135
vironmental information with sampling frequencies ranging from 2s to several minutes,136
reaching down to 2000 m in 10-day data collection cycles. Data streams are transmitted137
via satellite to distributed information centres (Global Data Assembly Centers, GDACs).138
GDACs make the data freely available for download (Argo, 2000). Argo currently ex-139
poses over 20-years of data and manages 4000 operational floats. Floats are located140
worldwide except for ice zones, with a higher density in the equatorial belt. The collected141
environmental parameters include depth, pressure, dissolved oxygen, ocean-current speed142
components, practical salinity, temperature, wind-stress components, electrical conductiv-143
ity, chlorophyll-a, and fluorescence. Argo data can be included in the class of FAIR data as144
being free, timely, and unrestricted-access data (Tanhua et al., 2019). Data access has the145
only policy to acknowledge the Argo network in scientific publications. Ethical oversight146
is left to the individual scientists or organizations using the data.147
To use Argo data in our niche models, they were aggregated and processed to reduce148
noise and computational complexity. Three groups of data were selected and downloaded149
from the GDACs - in CSV format - for the Adriatic Sea (using a bounding box extension of150
[+8;+20] longitude and [+38;+46] latitude). The first dataset contained observations from151
2015 to 2018; the second included observations collected in 2019; the third contained152
observations collected in 2020. The 2015-2018 range represents an aggregated reference153
of environmental conditions in the near past. This aggregation was necessary to provide154
reference statistical averages for the environmental parameters and allowed collecting a155
meaningful set of species observations for training ENMs. The 2019 data were used as a156
reference to assess if the variations observed in 2020 were either due to the pandemic or157
continuing trends from the previous years (possibly related to climate change). The 2020158
data were assumed to contain observations with signals of the COVID-19 pandemic and159
climate change.160
Argo data were averaged at a 0.1° resolution to increase statistical viability (Coro et al.,161
2018b). The following parameters were extracted from the CSV data: temperature (°C),162
salinity (PSU), chlorophyll-a (mgm3), dissolved oxygen (DOX) (µmolkg). These are163
indeed the most abundant and reliable data downloadable from Argo. For each parameter,164
average values were calculated for surface range, seafloor (bottom), and the entire water165
column. Surface and bottom ranges were identified as the first and last ranges of a log-166
arithmic division, into five parts, of the maximum depth of each 0.1° cell in the Adriatic167
(Reyes, 2015; Coro et al., 2018b). Instead of using static ranges, this approach adapted168
the definition of surface and bottom ranges to the specific cell depth. It normally results169
in better niche modelling, especially for benthic and demersal species (Ready et al., 2010;170
Reyes, 2015). For each parameter, surface, bottom, and average (in the water column) val-171
ues were estimates at 0.1° resolution. Furthermore, locations outside of the Adriatic Sea172
were excluded by only using those within the geographical subareas 17 and 18 of the Gen-173
eral Fisheries Commission for the Mediterranean (GFCM, 2020). This process generated174
36 datasets overall, as the results of three aggregation types (surface, bottom, average), for175
each aggregation time (2015-2018, 2019, 2020), repeated for four parameters.176
As a final step, consistency between the observations from the different datasets was177
enhanced by constraining all datasets to cover the same areas. Different spatial coverage178
over the years can indeed be a source of bias. For example, if observations covered north179
Adriatic more extensively than south Adriatic in a particular year, sampling would be180
northward skewed with consequent over-representation of northern environmental values.181
If this is not the case for the other years, inconsistency between parameter sampling and182
representation will occur. To avoid this issue, only probes locations that were present in all183
reference years were retained. A 0.5° spatial tolerance was used in the selection of these184
The ENM used in the present experiment required environmental data uniformly dis-186
tributed over the Adriatic Sea. Consequently, all 0.1° cells required an environmental value187
assigned, either averaged from the Argo observations or estimated through a model. Given188
the low density and quantity of the available environmental observations (Section 3) and189
the importance of currents in the biogeochemical components’ drift and spread in the Adri-190
atic, parameter values were interpolated through a model based on the advection equation191
and depth information. In particular, the Data-Interpolating Variational Analysis (DIVA)192
was used (Barth et al., 2010). DIVA is commonly used to produce uniform distributions of193
environmental parameters (Coro et al., 2018a; Coro and Trumpy, 2020; Schaap and Lowry,194
2010) and solves the advection equation to simulate the transport of a substance or quantity195
by currents. DIVA also estimates the mutual spatial correlation between observations and196
requires minimal parametrisation to produce high-quality interpolation at a user-defined197
resolution (Troupin et al., 2010, 2012; Coro et al., 2016c). Internally, DIVA reconstructs198
a continuous field from discrete measurements through a numerical implementation of the199
Variational Inverse Model (Bennett, 1992). This algorithm fits a continuous field to the200
data through a minimization cost function (Watelet et al., 2016), using a finite-element201
statistical method that embeds topographic and dynamic constraints (based on bathymetry202
and oceanic-currents data). It can process irregularly-spaced observations to produce esti-203
mates on a regular grid. Based on this fit, DIVA estimates a triangular-element mesh over204
the interpolation area, where the characteristic length of each element is directly linked to205
the mutual spatial correlation between observations.206
For our experiment, DIVA was applied to all Argo-aggregated data described in Sec-207
tion 2.1. Data of ocean current components were taken as NetCDF files from the Global208
Ocean Physic Analysis dataset hosted by the Copernicus Marine Service (Von Schuck-209
mann et al., 2018). In addition, depth information was taken from the GEBCO-2020210
bathymetry dataset, a global terrain model for ocean and land with 0.0042° uniform spatial211
resolution (GEBCO, 2020). To execute DIVA, the D4Science e-Infrastructure computa-212
tional platform was used (Candela et al., 2016; Coro et al., 2015a, 2017; Assante et al.,213
2019). As a result, 36 uniform parameter distributions at 0.1° resolution for our environ-214
mental parameter aggregations were produced and represented with the ESRI-grid format215
2.2. Species observations217
In order to extract species observation data, we consulted the Ocean Biogeographic218
Information System (OBIS) (Grassle, 2000). OBIS contains taxonomic and occurrence219
information for 155,000 marine species and provides access to more than 163 million220
observation records, integrated from more than 4,000 sources. Its contributors include221
international research projects, national monitoring programs, museums, and individuals.222
OBIS is suitable for data mining and pattern recognition experiments, especially in data-223
poor scenarios where the quality of the data is fundamental to produce reliable analyses224
(Coro et al., 2013b, 2015c, 2016b, 2018c). The OBIS data quality checking is integral to225
ecological niche models that are particularly sensitive to data bias (Coro et al., 2015b).226
Furthermore, for each occurrence record, OBIS indicates if it underwent expert verifica-227
tion. This feature makes OBIS more suited for ecological niche modelling in data-poor228
scenarios than other data collections (Coro et al., 2015b,c). In our experiment, the OBIS229
observation records in the Adriatic Sea, between 2015 and 2018, that underwent expert230
verification were retrieved for the eight species under study. Their coordinates were stored231
as CSV files to feed ENMs later.232
2.3. Ecological Niche Modelling233
Maximum Entropy (MaxEnt) is a widely used ENM for marine species (Raybaud234
et al., 2015; Angeletti et al., 2020; Capezzuto et al., 2018). MaxEnt is a shallow ma-235
chine learning model that estimates a function π(¯x)defined over real-valued vectors ¯x236
of environmental parameters. This function is forced to reach maxima on the parameters237
associated with a species’ presence and minima on absence-related parameters. Follow-238
ing a common abuse of notation, π(¯x)can be considered a proxy of a probability density239
of a species’ presence given the ¯xenvironmental parameters (Phillips and Dudík, 2008;240
Elith et al., 2011; Merow et al., 2013). MaxEnt learns the relation between environmen-241
tal values in the species-observation locations and the general species’ presence (Pearson,242
2007; Coro et al., 2018c). One advantage of this model is that it can work with species-243
presence information only, but it is over-sensitive to biased data (Elith and Graham, 2009;244
Coro et al., 2015b). A MaxEnt model trained with parameters and species observations245
at 0.1° resolution will produce a probability distribution of species presence over the 0.1°246
cell subdivision of a study area. The π(¯x)function is thus the probability that a 0.1° cell247
is suitable species habitat. MaxEnt estimates π(¯x)after maximising the entropy func-248
tion H=π(¯x)ln(π(¯x)) on the training locations with respect to randomly-selected249
environmental parameter vectors in the study area (background points). In the present250
experiment, ¯xwas made up of 13 parameters associated with the 2015-2018 year range:251
temperature, salinity, chlorophyll-a, DOX (with related surface, bottom, water-column ag-252
gregations), and depth (from the GEBCO-2020 bathymetry data set). Although depth was253
constant through the years, it was included in our models because it is a fundamental pa-254
rameter to estimate the niches of the studied species correctly. Depth was used as a proxy255
to model species preference to different seabeds and water column heights. Thus, it en-256
hanced prediction reliability by adding complementary and valuable information about the257
species habitat. On the other hand, it was not functional to the subsequent pattern analysis.258
Training locations were those associated with the OBIS observations between 2015 and259
2018. The used MaxEnt implementation (Phillips et al., 2021) accepted environmental260
parameters in ASC-raster format and species observation data in CSV format.261
The training algorithm estimates the coefficients of a linear combination of the en-262
vironmental parameters. These coefficients represent the weight of each environmental263
parameter in the species’ environmental preferences (percent contribution). MaxEnt also264
estimates the permutation importance of each parameter in the ¯xvector. The training pro-265
cess is based on the following function definitions: f(¯x), the probability density over the266
background parameters; f1(¯x), the density on the training set; and pr, the prior distribu-267
tion (prevalence) of the species (equal to 0.5 when no prior assumption is available, as in268
our case). Based on these functions, π(¯x)is defined as269
In a maximum entropy condition, the optimal f1(¯x)is the closest function to f(¯x),270
because there would be no difference without species observations. Additionally, f1(¯x)271
should have maxima on the parameter means in the training set locations. With these272
constraints, the model minimises the Kullback-Leibler distance between f1(¯x)and f(¯x)273
d(f1(¯x), f (¯x))=
This minimisation is solved by Gibbs distribution functions in the form f1(¯x)=f(¯x)eη(¯x)
(Phillips et al., 2006a), with η(¯x)=α+β h(¯x);αbeing a normalization constant that275
makes f1(¯x)sum to 1; hbeing an optional transformation of ¯xthat simulates a com-276
plex relation between the environmental parameters; and βbeing the percent contribution277
coefficients. The minimisation of η(¯x)- which requires solving a log-linear equation -278
consequently minimises d(f1(¯x), f (¯x)). The used MaxEnt software automatically solves279
this minimisation problem. It also estimates percent parameter contribution through an280
iterative process that calculates and accumulates the percent performance gain provided281
by each parameter (Phillips et al., 2017).282
MaxEnt is generally preferred over linear and logistic regression for species habitat283
distribution modelling. It is equivalent to a Poisson regression (a generalized linear model)284
that is naturally suited for modelling the probability of a number of events in a fixed space285
(such as species occurrences) (Renner and Warton, 2013). Once the model parameters286
have been estimated, the π(¯x)function can be used to estimate probability distributions287
over new parameter values than those of the training set, e.g. the parameters of locations288
outside of the study area (to discover the potential species niche) or new environmental289
scenarios (to study niche change over time) (Elith and Graham, 2009; Phillips et al., 2017).290
MaxEnt is sensitive to sampling bias associated with species-observation locations and291
can over-fit small datasets (Merow et al., 2013; Wang et al., 2018). Our selected occur-292
rence datasets were indeed small, as only expert-verified records were selected. They293
also had potentially biased distributions, as they belonged to OBIS-included surveys with294
frequent and fixed paths (Coro et al., 2015c). One way to manage this issue is to select295
background points far away from the presence locations (Hengl et al., 2009). However,296
our analysed species are common and widely distributed in the Adriatic, with absence lo-297
cations potentially dense in the presence areas. Therefore, it was not possible to focus298
background point sampling on specific areas. Providing the model with precise absence299
and background locations would also have required more presence data and precise envi-300
ronmental parameter distributions. However, specific studies on MaxEnt parametrisation301
(Zaniewski et al., 2002; Dudík et al., 2005; Phillips and Dudík, 2008; Phillips et al., 2017)302
have indicated general strategies to reduce presence location sampling and over-fitting bi-303
ases, which include (i) selecting background points to reflect the same sampling bias as304
the presence locations, (ii) including presence points among background points, (iii) using305
hinge features to model complex species response to the environmental parameters and306
make model fitting more flexible. The MaxEnt software used for this experiment offers307
options to use hinge features and include presence locations among background points if308
these are associated with unique combinations of environmental parameters (Phillips et al.,309
2021). These options were used to attenuate over-fitting and sampling bias issues as far as310
In the present experiment, MaxEnt was trained with 2015-2018 Adriatic environmental312
data and species occurrence records to produce an ecological niche reference for the near313
past. Then it was projected onto the 2019 and 2020 environmental data to analyse prob-314
ability distribution change due to the different environmental parameters of these years.315
Since the βvector indicates the parameters that carry the highest quantity of informa-316
tion to understand species habitat preferences (Coro et al., 2018c; Coro, 2020), it can be317
used to remove poorly niche-correlated parameters from the ¯xvector. This operation opti-318
mally selects the variables associated with the species habitat (Section 3.1). For example,319
deep-water and benthic species will likely be modelled with bottom-averaged parameters,320
whereas pelagic species habitat will likely be modelled with water-column or surface re-321
lated parameters. Furthermore, reducing the number of input environmental parameters322
decreases the inter-dependence between the variables and improves the model accuracy323
(Coro et al., 2015b). In the present experiment, the MaxEnt models of the studied species324
were first trained with all parameters and then re-trained using only those parameters hav-325
ing a percent contribution within 95% from the maximum contribution.326
In summary, MaxEnt ENMs were produced for the 8 Adriatic species through the fol-327
lowing steps: (i) MaxEnt models were trained with 2015-2018 OBIS observations and328
interpolated environmental data; (ii) after a first training phase, the parameters with the329
95% highest percent contributions were retained (thus, different parameter sets were as-330
sociated to the different species); (iii) the models were re-trained only with the retained331
parameters; (iv) the models were projected onto the 2019 and 2020 environmental param-332
eters. The produced models will be referred to as floating sensor (FS) based models -333
i.e., FS 2015-2018, FS 2019, and FS 2020 - to distinguish them from the baseline models334
used for evaluation. A total of 24 models was thus produced, i.e., three models for each335
analysed species.336
2.4. Evaluation and pattern recognition337
The ENM distributions were used to discover driving factors of species habitat change338
over the years. The first goal of our quality evaluation was to assess the consistency of the339
produced maps. As our second goal, the principal environmental drivers of habitat suit-340
ability change were checked against evidence from general climate change and COVID-19341
pandemic related trends. The entire evaluation process was managed through four evalua-342
tion questions:343
Question 1:Are the produced distributions consistent?344
This question was answered by verifying the similarity between our models and other345
ENMs. This operation confirmed that our models consistently captured the species’ envi-346
ronmental preferences, although they were trained on scarce and scattered data and tested347
on the same training set (Section 3.1). Indeed, the partial reliability of our MaxEnt model348
was assessed using the training data, but this was insufficient to state they were consistent,349
due to the few data at hand. Thus, we set two consistency boundaries for our model: one350
similarity and one dissimilarity reference. We used the similarity reference to confirm that351
the produced distributions agreed with an independent habitat distribution. Instead, we352
used the dissimilarity reference to check for significant difference with respect to a known353
improbable scenario based on unlikely environmental parameter distributions.354
The AquaMaps distributions were used for these tasks (Kaschner et al., 2006). They355
were downloaded (not re-calculated) from the AquaMaps website (AquaMaps, 2020).356
AquaMaps is a presence-only ENM that incorporates scientific expert knowledge into357
species habitat modelling to account for known limitations of species occurrence records358
(Corsi et al., 2000; Ready et al., 2010). We used AquaMaps as a mechanistic model359
to estimate species distributions independently of the data available in our experiment.360
Moreover, AquaMaps uses a complementary approach with respect to machine-learning-361
based approaches because it explicitly models the causality between species presence and362
environmental parameters (Pearson, 2007; Baker et al., 2018). AquaMaps has comparable363
accuracy to GAM- and GLM-based ecological niche models (Ready et al., 2010). It is364
particularly effective for large areas (e.g., the size of the Adriatic Sea) and when expert365
knowledge about the species is available at the global scale. Moreover, it is reliable for366
extracting macro-patterns of climate change influence on species distributions (Coro et al.,367
The AquaMaps native algorithm estimates the species niche distribution in its known369
habitat. It uses a multiplication of environmental parameter envelopes whose ranges are370
either statistically estimated or defined by an expert. The environmental parameters inte-371
grated with the model are 0.5° resolution distributions of depth, salinity, temperature, pri-372
mary production, distance from land, and sea ice concentration. In the present experiment,373
the AquaMaps native model based on 2019 annual environmental parameters (hereafter374
referred as AquaMaps 2019) was used as a similarity reference for our models.375
As a dissimilar reference model, the AquaMaps native-2050 model was used (hereafter376
referred as AquaMaps 2050). This model integrates environmental parameters estimated377
under the Special Report on Emissions Scenario (SRES) A2 of the Intergovernmental Panel378
on Climate Change (IPCC). This scenario describes a future world with independent, self-379
reliant nations with a continuously increasing population. Economic and technological380
development are assumed to increase non uniformly across the world countries. Of key381
importance are average surface temperature and salinity that have increasing trends (with382
localised decreases for salinity), whereas ice concentration decreases globally and wa-383
ter level increases. Our models were checked to be significantly distant from AquaMaps384
2050 because this model represents an unlikely scenario for all selected species today.385
Using the AquaMaps 2050 distributions as unlikely scenarios was particularly consistent386
for our studied species because their 2050 distributions were significantly different from387
the AquaMaps native distributions (Section 3). The AquaMaps native models were down-388
loaded from the AquaMaps website (AquaMaps, 2020; Scarponi et al., 2018), whereas a389
NetCDF FAIR version of the AquaMaps 2050 model was used, whose consistency and va-390
lidity was confirmed by other experiments (Coro et al., 2018a). GDAL and CDO software391
(OSGeo, 2019) was used to downsample the models to 0.1° resolution, through first-order392
conservative remapping (Schulzweida, 2020), in order to be able to compare them with393
our models.394
Question 2:Can habitat patterns be identified in 2020 with respect to the previous395
A map comparison procedure was used to answer this question (described in Coro397
et al. (2014)). This process calculates discrepancy and agreement between two maps. It398
allows setting a threshold over each probability distribution to conduct presence/absence399
comparison. Absences are values under the threshold and presences are values over the400
threshold. The process then uses this classification to calculate discrepancy as the percent-401
age cells where the two distributions disagree. It also calculates Cohen’s kappa (Cohen402
et al., 1960) to estimate agreement with respect to chance. Kappa is classified as poor,403
slight, fair, moderate, substantial, or excellent according to the Landis and Koch range404
classifications (Landis and Koch, 1977).405
The three FS distributions of each species had different probability ranges. This issue406
made it difficult to find a common threshold to compare low and high probability cells,407
which is a common problem when comparing different distributions (Coro et al., 2014;408
Phillips et al., 2006b). MaxEnt suggested potential habitat suitability thresholds out of a409
training session over the 2015-2018 data, using a sensitivity-specificity analysis that con-410
sidered only the observations and environmental data in 2015-2018. However, after this411
training session, the MaxEnt model was projected onto the 2019 and 2020 data without412
re-training, and this operation normally produces distributions with new probability ranges413
(Phillips et al., 2006b; Coro and Bove, 2022). One approach to accommodate for this is-414
sue is to allow MaxEnt to extend estimates beyond the parameter ranges observed on the415
training set (i.e., to disable the model’s clamping option). However, this technique should416
be used with caution because it could generate inconsistent results or unnatural projec-417
tions (Elith et al., 2011). Moreover, the approach assumes that the projection conditions418
represent a completely different environmental scenario (e.g., in the far past or future). In419
contrast, our projection scenarios fell within the clamped ranges for most variables (Sec-420
tion 3.3). We also experimentally verified that clamping was not useful in overcoming this421
issue with the data at hand.422
Thus, the thresholds suggested by the sensitivity-specificity analysis over the 2015-423
2018 data could not be used for the 2019 and 2020 distributions. Therefore, conducting a424
fair comparison between the MaxEnt distributions required setting appropriate thresholds425
for habitat suitability/unsuitability on each distribution separately; to transform a numer-426
ical comparison into a consistent classification comparison. In this case, one possible427
threshold to use is the first-quartile probability value, as also suggested by O’Brien (1980)428
and Theil (1982). This property comes out of the observation that although the distribution429
ranges and shapes can differ between the models, one comparable measure of MaxEnt430
probability abundance (and thus of habitat suitability extent) is the number of elements431
with MaxEnt output value over the first quartile. Therefore, we used the first-quartile432
probability value of each FS distribution to identify areas of low and high suitability. Our433
results demonstrate that this approach generated comparable FS distributions (Section 3).434
As for AquaMaps, the log-linear nature of this model allows setting a 0.2 probability value435
as the threshold (Coro et al., 2013a, 2016a).436
Since discrepancy and agreement calculation does not indicate if one distribution cor-437
responds to more suitable habitat than the other, a new metric was introduced for this438
scope. In particular, a suitability score was defined on the discrepancy cells:439
where irefers to cells on which the two dichotomic Pand P′′ distributions differ;440
Nis the total number of cells involved in the comparison; and P
H(i)and P′′
H(i)are the441
compared habitat distributions using new thresholds that identify very high probability442
zones. These thresholds were set to the 3rd quartiles of the FS distributions and to 0.8 for443
AquaMaps. The rationale behind the suitability score calculation is that if one distribution444
indicates very high suitability in the discrepancy areas more often than the other, that445
distribution is overall more favourable. Thus, S>0indicates that the first distribution is446
more suitable than the second (habitat gain) - and vice-versa when S<0(habitat loss)447
- whereas S=0indicates overall equal suitability between the two distributions (stable448
Discrepancy, agreement, and suitability scores over the years can identify habitat change.450
Increasing habitat suitability from 2015-2018 to 2019 and 2020 may indicate overall habi-451
tat expansion (gain) in 2020, stable suitability may indicate unchanged habitat, and incon-452
stant habitat gain and loss over the years can be associated with potential habitat change.453
Question 3:Which parameters drove habitat change in 2020?454
MaxEnt also produces single-parameter distributions by training the model with one455
parameter at-a-time. These parameter distributions allow inferring the parameter ranges456
that correspond to higher suitability. The inference is straightforward when the involved457
parameters are independent or bring a high contribution (Coro et al., 2013a, 2015b, 2018c).458
Our approach enhances parameter independence by re-training MaxEnt after removing459
low-contributing parameters. Intersecting environmental parameter trends with MaxEnt460
single-parameter distributions identifies the key responsible parameters for habitat change.461
Question 4:Do environmental parameter changes in 2020 depend on the COVID-19462
pandemic or also on climate change?463
The change in key parameters for our selected species’ habitat change could be due464
to statistical inter-annual fluctuations, or to general global-scale changes such as climate465
change or the reduction of anthropogenic pressure due to the COVID-19 pandemic. The466
key factors were investigated by searching for other studies that specifically analysed these467
parameters in other locations and correlated their trends to climate change or the pandemic.468
This analysis, combined with the results from the previous evaluation phases, clarified the469
correlation between anthropogenic pressure on ecosystems due to the COVID-19 pan-470
demic, the coupling with climate change, and potential species habitat change.471
2.5. Complete workflow472
The complete workflow can be summarised as the production and comparison of Max-473
Ent distributions of eight selected Adriatic Sea species out of OBIS species observations474
and Argo data. Each step of the workflow has code and data associated in the open-source475
repository linked to this paper (see Supplementary Material). The steps can be summarised476
through the following phases:477
Phase 1: Retrieve Argo data for the Adriatic and aggregate them at 0.1° spatial res-478
olution (from Select probes across479
years that have a mutual distance under 0.5° . Produce surface, bottom, and water-column480
average values for each environmental parameter in every reference time frame, i.e., 2015-481
2018, 2019, and 2020. This phase generated 9 datasets (3 aggregations by 3 years) for482
Argo parameters (4 in total), i.e., 36 datasets overall. All processing R code and results of483
this phase are available in the repository linked in the Supplementary Material, within the484
"Phase 1 - Argo Data Preparation" folder.485
Phase 2: Interpolate the 36 environmental parameter datasets through DIVA, using486
data on ocean current speed components and depth, to obtain uniform 0.1° distributions for487
the entire Adriatic. Prepare the data as ASC files for MaxEnt. The used DIVA notebook488
and the results of this phase are available in the repository linked in the Supplementary489
Material, within the "Phase 2 - Environmental Parameter Distributions" folder.490
Phase 3: Retrieve species occurrence records from OBIS (
manual/access/) and prepare them for MaxEnt. For each species, use 2015-2018492
OBIS species occurrence records and environmental datasets (plus depth from GEBCO)493
within a MaxEnt model to produce 8 floating-sensor-based full-variable models for 2015-494
2018 at 0.1° resolution. The retrieved and pre-processed OBIS occurrences, the data prepa-495
ration scripts, the link to the MaxEnt software, and the MaxEnt results are available in the496
repository linked in the Supplementary Material, within the "Phase 3 - Occurrence Records497
and First MaxEnt Run" folder.498
Phase 4: Execute MaxEnt again, for each species, using only the parameters that499
had the highest percent contribution, i.e., those within 95% relative difference from the500
maximum. This phase produced 8 final FS 2015-2018 models, one for each species. It501
also modelled each species with an optimal selection of parameters associated with their502
preferred depth ranges. For example, it selected depth and bottom-level parameters for503
deep-water and benthic species (Section 3.3). As a further step, project the MaxEnt models504
over the 2019 and 2020 parameter data to obtain FS 2019 and FS 2020 models for the 8505
species. The MaxEnt re-execution results are available in the repository linked in the506
Supplementary Material, within the "Phase 4 - MaxEnt Re-application" folder.507
Phase 5: Retrieve AquaMaps 2019 and 2050 distributions and downsample them to508
0.1° for consistent comparison with the MaxEnt distributions. The retrieved AquaMaps509
distributions are available as ESRI-grid files in the repository linked in the Supplementary510
Material, within the "Phase 5 - AquaMaps Distributions" folder.511
Phase 6: Extract parameter quantiles to study trends over the years. Compare Max-512
Ent distributions to quantify discrepancy and estimate habitat change (though suitability513
score). The results and the used scripts are available in the repository linked in the Sup-514
plementary Material, within the "Phase 6 - Estimate Quantiles" folder.515
Phase 7: Identify patterns of habitat change (gain, loss, stability). The extracted516
patterns are available in the repository linked in the Supplementary Material, within the517
"Phase 7 - Patterns" folder.518
Phase 8: Study the main parameter trends to identify those that influenced habitat519
change. Understand the relation between these trends and climate change and COVID-19520
pandemic (Sections 3.3-3.4).521
3. Results522
Our method produced distribution maps for 2015-2018, 2019, and 2020 for each of523
the eight analysed species (Figure 2). Referring to our evaluation questions (Section 2.4),524
Section 3.1 addresses question 1; Section 3.2 addresses question 2; Section 3.3 addresses525
question 3; and Section 3.4 addresses question 4.526
For the present experiment, our workflow processed overall 2,166,025 in situ observa-527
tions for 2015-2018, 364,219 observations for 2019, and 463,352 observations for 2020.528
These observations covered from 600 (for chlorophyll-a and DOX) to 2100 (for tem-529
perature and salinity) 0.1° cells in the Adriatic Sea. OBIS occurrence records that had530
undergone expert review were extracted for these cells to increase observation reliability531
(at the expense of their quantity). The extracted records between 2015 and 2018 were532
47 for Sepia officinalis, 189 for Merluccius merluccius, 166 for Mullus barbatus, 39 for533
Sardina pilchardus, 30 for Parapenaeus longirostris, 28 for Solea solea, 40 for Squilla534
mantis, and 27 for Engraulis encrasicolus. These observations were distributed across535
the species’ Adriatic habitats (Figure 1). Although they were theoretically unsuitable for536
building a detailed model, they were useful for a macroscopic pattern-change analysis of537
species distributions, in agreement with other ENM approaches that use even a lower num-538
ber of observations to trace viable environmental envelopes for pattern analyses (Kaschner539
et al., 2006; Rees, 2008; Ready et al., 2010; Kaschner et al., 2011; Coro et al., 2016a).540
3.1. Model consistency541
3.1.1. Variable selection and model optimisation542
Our feature selection criterion was evaluated using the Kuenm R package (Cobos et al.,543
2019), which also allowed us to fine-tune the models. This software exhaustively tests544
the performance of MaxEnt with multiple sets of environmental parameters and finds the545
optimal configuration of (i) the analytical form of h- among linear, quadratic, product,546
threshold, hinge, and their combinations (feature classes) - and (ii) a penalty factor on the547
βvector (regularisation multiplier) (Merow et al., 2013; Morales et al., 2017). Kuenm548
allows selecting the optimal model based on the highest Akaike Information Criterion549
value (AIC) calculated on a test set. To select the optimal parametrisations of our 2015-550
2018 models, several sets of environmental variables were prepared and evaluated in two551
ways: (i) on the entire training set (self-performance) and (ii) based on the average AIC552
over ten randomly extracted observation sets, with an 80-20% training-test set ratio for553
each extraction and considering only models with omission rate below 5%. The prepared554
sets of environmental variables included the entire set, the 95% percent contribution-based555
set (Section 2.3), and ten randomly chosen subsets.556
The Kuenm evaluation estimated that the optimal regularisation multipliers for all anal-557
ysed species ranged around 1. Thus, this parameter was fixed to 1 for all models for558
simplicity, i.e. no penalty was set on β. Moreover, both self-performance and 80-20%559
validation indicated that the optimal set of environmental variables was the one obtained560
using a 95% threshold percent contribution from the maximum contribution. Finally, us-561
ing a complex hfunction that combined all feature classes was optimal for 80-20% valida-562
tion and also gained high self-accuracy performance. The average AIC over all tests was563
990, whereas the average optimal models’ AIC was 860. These results likely derive564
from the fact that our selection criterion discards the predictor variables that bring poor565
and potentially confounding information to the model. Moreover, using complex feature566
classes reduced the over-fitting bias (Section 2.3) and thus likely increased validation per-567
As a further evaluation step, the Receiver Operating Characteristic (ROC) curve was569
traced for each optimal model to conduct a sensitivity analysis. This analysis calculated570
the true-positive rate and the false-positive rate using various decision-thresholds on the571
model output. Consequently, all optimal models were verified to achieve an Area Under572
the Curve (AUC) (i.e., the integral of the ROC curve) over 0.95. Specifically, AUC was573
averagely 0.96 [0.954;0.97] for the optimal models, and 0.83 [0.78;0.95] for sub-optimal574
models. This property guaranteed that the probability distributions simulated by each575
model were significantly higher on species-presence locations than on random locations.576
All these quality checks aimed to optimise model robustness in a context of scattered577
environmental data and few observation data.578
It is worth noting that using AIC as a selection criterion can be prone to criticisms,579
especially because AIC tends to select models with a higher number of parameters among580
equal-likelihood models (Guthery et al., 2005; Arnold, 2010). However, issues especially581
arise if AIC were used (i) as the only selection criterion, (ii) without adding prior infor-582
mation to guide selection, and (iii) to build models that pretend to assess ecological reality583
(Zhang et al., 2018; Reside et al., 2019; Roy-Dufresne et al., 2019). Therefore, our use of584
AIS, through Kuenm, can be tolerated because we (i) did not assume the optimal models585
to be punctually reliable, but generally reliable to assess macroscopic changes when com-586
pared to each other, (ii) used a prior condition to evaluate only the models with omission587
rates below 5%, (iii) forcibly introduced a further parametrisation that involved the 95%588
percent contribution-based set; (iv) added sensitivity analysis to assess model validity fur-589
ther; (v) checked model consistency through comparison with AquaMaps; (vi) introduced590
constraints to avoid over-fitting. Indeed, the optimal models did not use the highest number591
of environmental parameters and complex regularisation and penalty conditions.592
The optimal parametrisations estimated for the FS 2015-2018 models were also used593
for the FS 2019 and FS 2020 projections. The resulting optimal distributions are reported594
in Figure 2.595
3.1.2. Comparison with AquaMaps596
The dissimilarity between our maps and AquaMaps 2019 was reasonably low, i.e., av-597
eraging below 20% (19.14%, Table 1). Furthermore, a fair kappa agreement (according598
to Landis and Koch classification, Landis and Koch (1977)) occurred for 81.3% of the599
comparisons. The greatest discrepancy, corresponding to slight agreement, was found for600
Engraulis encrasicolus and Merluccius merluccius. For these species (Figures 2-h and601
-b), AquaMap 2019 extended more into south Adriatic. As for AquaMaps 2050, the IPCC602
SRES A2 scenario was found to be significantly distant from our distributions, with a603
30% average discrepancy and poor/marginal agreement with 87.5% of the distributions.604
The highest similarity - with moderate kappa agreement - occurred for Squilla mantis605
(19.2% discrepancy vs FS 2015-2018, 17.57% vs FS 2019, and 19.07% vs FS 2020). The606
FS models indicated that this species had a stable habitat concentrated in northern Adriatic,607
whereas AquaMaps 2019 estimated a possible presence in south Adriatic. Notably, OBIS608
does not report expert-verified occurrences of Squilla mantis in south Adriatic, which en-609
forces the consistency of our model.610
Overall, this assessment indicates that our distributions generally agreed with an in-611
dependent reference model (AquaMaps 2019) and were far from an unlikely scenario612
(AquaMaps 2050). Thus, despite the poor data, the predictions of our models were not613
poor, which permitted us to conduct further analyses and extract general patterns over the614
3.2. Habitat change classification616
Based on the discrepancy (Table 1) and the suitability score (Table 2) calculations,617
detailed habitat gain and loss trends were traced per species. In particular, Sepia of-618
ficinalis habitat expanded in 2020 with respect to both 2015-2018 (+3.95%) and 2019619
(+0.14%) with significant discrepancy (12.36% vs. 2015-2018 and 7.18% vs. 2019) (Fig-620
ure 2-a). Distributional differences were found off the Apulian coasts and in the south621
Balkans. The FS 2020 distribution was also similar to AquaMaps 2019, with substantial622
kappa agreement, because both the distributions indicated extension towards south-east623
and south-west. In northern Adriatic, the FS 2020 map presented a similar distribution624
to the other FS maps, with substantial kappa agreement. This distribution was differ-625
ent from AquaMaps 2050 (24.72% discrepancy), which predicted habitat loss throughout626
south Adriatic. Overall, this analysis indicates habitat gain for this species in 2020.627
Merluccius merluccius habitat expanded in 2020 with respect to 2015-2018 (+5.68%)628
but minimally lost habitat with respect to 2019 (-0.36%) (Figure 2-b). The discrepancy vs629
2019 (5.89%) was lower than vs 2015-2018 (17.82%). The similarity between FS 2020630
and FS 2019 was due to minimal differences in the south-eastern Adriatic. Furthermore,631
FS 2019 reported habitat gain (+7.04%) against FS 2015-2018, which indicated an increas-632
ing habitat extension trend over the years. The greatest discrepancy between FS 2020 and633
AquaMaps 2019 was in the south Adriatic, where AquaMaps reported high suitability.634
The FS 2020 distribution was also different from AquaMaps 2050 (41.03% discrepancy)635
due to the AquaMaps-predicted habitat loss throughout south Adriatic in 2050. Overall,636
this analysis suggests habitat gain for this species in 2020 because its habitat substantially637
expanded with respect to 2015-2018 and was similar to a habitat-favourable 2019.638
Similarly, Mullus barbatus habitat expanded in 2020 with respect to 2015-2018 (+3.38%)639
and slightly lost habitat with respect to 2019 (-1.94%) (Figure 2-c). The discrepancy vs640
2019 (9.20%) was lower than vs 2015-2018 (16.24%). The similarity between FS 2020641
and FS 2019 was due to minimal differences in middle Adriatic. Furthermore, FS 2019642
resulted in habitat gain (+7.61%) against FS 2015-2018, which indicated an increasing643
habitat extension trend over the years. The FS 2020 was also similar to AquaMaps 2019644
(19.6% discrepancy and moderate agreement) because both models reported high suit-645
ability for south Adriatic. For this reason, FS 2020 was different from AquaMaps 2050646
(27.42% discrepancy and poor agreement), which foresaw habitat loss in south Adriatic.647
Overall, this analysis indicates habitat gain for Mullus barbatus in 2020 because its habi-648
tat substantially expanded with respect to 2015-2018 and was similar to an advantageous649
Sardina pilchardus habitat expanded with respect to 2015-2018 (+4.6%) but substan-651
tially lost habitat with respect to 2019 (-5.46%) (Figure 2-d). The discrepancy between FS652
2020 and FS 2019 (29.6%) was concentrated off Apulian coasts (with gain in 2020) and in653
the Balkans (with gain in 2019). Furthermore, FS 2019 reported habitat gain (+4.31%) vs654
2015-2018 especially in south-western Adriatic and off central Italian coasts. Thus, habi-655
tat trend was not stable, and the FS 2020 habitat suitability patterns changed with respect656
to FS 2015-2018 and FS 2019. Due to the high suitability reported in south Adriatic, all657
FS distributions had moderate agreement with AquaMaps 2019. The discrepancy between658
FS 2020 and AquaMaps 2050 (20.89%) was lower than the one of the previous species659
because also AquaMaps 2050 foresaw suitable habitat in 2050 in south Adriatic. Overall,660
this analysis indicates habitat change for Sardina pilchardus in 2020 because no definite661
trend and pattern was present across the models.662
Similarly, Parapenaeus longirostris habitat expanded with respect to 2015-2018 (+8.33%)663
but substantially lost habitat with respect to 2019 (-7.04%) (Figure 2-e). The discrepancy664
between FS 2020 and FS 2019 (20.83%) was concentrated in the south and middle Adri-665
atic (with gain in 2019). In the same areas, FS 2019 reported substantial habitat gain666
(+14.87%) vs 2015-2018. Thus, habitat trend was unstable since the FS 2020 habitat suit-667
ability patterns were substantially different with respect to FS 2015-2018 and FS 2019.668
All FS distributions had moderate kappa agreement with AquaMaps 2019 due to the high669
habitat suitability AquaMaps indicated in south Adriatic. In contrast, since AquaMaps670
2050 indicated great habitat loss in south Adriatic, the discrepancy with FS distributions671
was large (42.37% average). Overall, this analysis indicates habitat change for Parape-672
naeus longirostris in 2020 because no definite trend and pattern was present across the673
Solea solea slightly gained habitat with respect to 2015-2018 (+0.5%) and presented675
stable habitat suitability with respect to 2019 (Figure 2-f). The discrepancy between FS676
2020 and FS 2015-2018 (6.75%) was due to a slightly higher suitability area off Apulian677
coasts by FS 2020. The habitat change trend was thus stable, and the similarity and the678
kappa agreement between the FS 2020 and the other distribution was substantial. The679
FS distributions also had substantial kappa agreement with AquaMaps 2019, with very680
similar patterns throughout the Adriatic. Since AquaMaps 2050 foresaw great habitat loss681
in south Adriatic (except for a small area in southern Balkans), its discrepancy with respect682
to the FS distributions was high (34.63%). Overall, this analysis indicates stable habitat683
for Solea solea from 2015-2018 to 2020.684
Squilla mantis slightly gained habitat with respect to 2015-2018 (+0.36%) and slightly685
lost habitat with respect to 2019 (-0.72%) (Figure 2-g). The discrepancy between FS686
2020 and the other FS distributions was concentrated off the Apulian coasts. The habitat687
change trend was overall stable, and kappa agreement between the FS 2020 and the other688
distribution was substantial. The FS distributions also had moderate kappa agreement689
with AquaMaps 2019, which reported habitat suitability for most of the Adriatic. Since690
AquaMaps 2050 reported high probability areas in northern and middle Adriatic and off691
northern Albanian coasts, kappa agreement with the FS maps was moderate. Overall,692
Solea solea presented an overall stable habitat from 2015-2018 to 2020.693
Engraulis encrasicolus presented stable habitat distribution with respect to 2015-2018694
and a slight suitability loss with respect to 2019 (-1.15%) (Figure 2-h). The discrep-695
ancy between FS 2020 and FS 2019 was due to a higher probability area off Albanian696
coasts. The habitat change trend was overall stable, and the mutual similarity had sub-697
stantial kappa agreement. The FS distributions also had moderate kappa agreement with698
AquaMaps 2019, which presented a decreasing gradient from north to south. Since AquaMaps699
2050 reported habitat loss for middle and south Adriatic, kappa agreement with the FS700
maps was poor. Overall, Engraulis encrasicolus presented an approximately stable habi-701
tat from 2015-2018 to 2020.702
3.3. Habitat change due to environmental parameter change703
The key driving parameters for habitat change in 2020 were identified through the704
analysis of their percent contributions (Table 4). Notably, the MaxEnt parameter selection705
corresponded to known environmental preferences of the studies species. For example,706
Mullus barbatus lives in sandy, muddy bottoms near river mouths (Esposito et al., 2014),707
and indeed its key parameters were bottom temperature and depth, but also chlorophyll-a708
and DOX averages in the upper water column. Sardina pilchardus habitat-depth ranges709
between 10 and 100 m (Santos et al., 2006), and indeed it was associated with bottom and710
water-column averaged parameters. Parapenaeus longirostris is a deep-water species, and711
its habitat was indeed highly dependent on depth. However, its distribution also depends712
on temperature and DOX in the water column (Ardizzone et al., 1990) as confirmed by our713
MaxEnt model.714
The single-parameter charts of FS 2015-2018 - produced by MaxEnt after training -715
were used to identify the most significant driving factors of the change (Figure 3). In addi-716
tion, parameter quartiles were extracted to understand if variation trends could be identified717
among the driving factors (Table 3). To enhance readability, only the parameter distribu-718
tions that were sensitive to parameter change over the years, i.e., with probability density719
variation over 0.05 - were reported in Figure 3. Other probability distributions indicated720
non-significant variation in correspondence of the median parameter change over the years721
(e.g., they reported a plateau over the variation range), and were omitted. Since this anal-722
ysis was conducted on the optimal models, only the parameters that showed significant723
percent contribution were analysed for each species’ distribution.724
As regards the species that expanded habitat, Sepia officinalis was mainly supported by725
a general decreasing trend, from 2015 to 2020, of average DOX (with median going from726
234.1 to 213.7 µmolkg, Table 3) and an increasing trend of bottom temperature over the727
years (with median rising from 14.15 to 14.32 °C, Table 3). These two parameters sig-728
nificantly contributed to the MaxEnt model, and their trends went towards maxima of the729
single-parameter densities (Figure 3-a). Change in the other parameters did not influence730
habitat gain and thus was not discussed. Merluccius merluccius and Mullus barbatus ex-731
panded habitat especially because of increasing bottom temperature trend and decreasing732
average chlorophyll-a over time (from 0.039 to 0.034 mgm3, Table 3). These changes733
moved the habitat to higher MaxEnt probability values and consequently increased habitat734
gain (Figures 3-b and -c).735
As regards the species that changed habitat, the inconstant trend of Sardina pilchardus736
was due to average DOX and average chlorophyll-a decrease (Table 3). This decrease737
changed habitat suitability in 2020 with respect to 2015-2018 (Figure 3-d), and also gen-738
erated different patterns between the FS 2019 and 2020 distributions. Habitat change for739
Parapenaeus longirostris was mainly driven by surface temperature modulations (from740
16.6 °C in 2015-2018 to 19.7 °C in 2019 and 18.4 °C in 2020, Table 3) and surface DOX741
modulations (from 228.36 µmolkg in 2015-2018 to 227.8 µmolkg in 2019 and 214.7742
µmolkg in 2020, Table 3). For this species, this parameter combination resulted in a less743
favourable habitat in 2020 than the previous years (Figure 3-e).744
The species with stable habitat distributions presented a robust response to environ-745
mental change, and no parameter could be highlighted over the others.746
3.4. Environmental parameter relation with climate change and COVID-19 pandemic747
The parameters that principally drove distribution changes - i.e., temperature, chlorophyll-748
a, and DOX - were analysed to understand if their change depended on inter-annual cli-749
matic variations, general climate change trends or the COVID-19 pandemic (Table 5).750
The general change of temperature positively affected the distributions of Sepia offic-751
inalis,Merluccius merluccius,Mullus barbatus, but negatively the one of Parapenaeus752
longirostris. Despite the cooling effect of La Niña since August 2020 - which mainly af-753
fected surface temperature - global temperature increased up to 1.2 ° C above pre-industrial754
value (DownToEarth; United Nations, 2021a; World Meterological Organization, 2021).755
Similarly, the general decrease of DOX positively affected the habitat of Sepia of-756
ficinalis, but negatively the habitats of Sardina pilchardus and Parapenaeus longirostris.757
Although in 2020 DOX increased in several world areas, as the consequence of the qual-758
ity improvement of coastal environments during the pandemic (Arif et al., 2020), in the759
Adriatic Sea the trend has been strongly decreasing in the last two decades (Kralj et al.,760
2019b). The Adriatic has a generally increasing DOX gradient from north to south conse-761
quent to its water circulation, a decreasing nutrient concentration provided by rivers, and762
a higher phytoplankton development in northern regions (especially in autumn and win-763
ter) (Zavatarelli et al., 1998). The overall average DOX decrease trend is probably due to a764
general DOX depletion at the Adriatic Sea floor. DOX level correlates with plankton respi-765
ration and benthic oxygen consumption, which has been exceeding the oxygen produced766
by microalgae and the one coming from oxygenated water (Kralj et al., 2019b; Lipizer767
et al., 2014). This condition has been assessed as being a probable consequence of bottom768
temperature and salinity increase due to climate change (Marasovi´
c et al., 2005; Lipizer769
et al., 2014; Kralj et al., 2019a), and indeed was never observed before 1984 (Justi´
c et al.,770
Conversely, the strong chlorophyll-a decrease in 2020 - i.e., -6% in the water column,772
-50% at the sea bottom, and -14% at the surface than 2019, based on the Argo data (Table773
3) - could be correlated with the COVID-19 pandemic. Although this correlation cannot774
be demonstrated with our data, some supporting conjectures can be reported from other775
studies. Chlorophyll-a is indeed one of the main indicators of ocean productivity and is an776
integral part of the carbon cycle and oxygen production. The carbon cycle indeed depends777
on carbon dioxide consumption during photosynthetic primary production and inorganic778
carbon production during biomineralisation. The global balance of the natural carbon779
cycle implies that a large decrease of carbon dioxide (CO2) in the atmosphere likely corre-780
sponds to a lower chlorophyll-a level because of the lower demand for CO2uptake (Shehhi781
and Samad, 2021). In 2020, a 7% reduction in the global carbon dioxide emissions was782
measured from satellite and in situ estimates due to big industry closure in several world783
countries with high industrial activity and large population (Le Quéré et al., 2020). As784
a probable consequence (Adwibowo, 2020; Mishra et al., 2020), a consistent decrease of785
chlorophyll-a was observed in many areas throughout 2020. For example, a 123 tonne786
reduction of CO2emission in south China corresponded to a measured 5% reduction of787
chlorophyll-a during the pandemic (Shehhi and Samad, 2021). This phenomenon was also788
observed in north Europe, South Korea, south-east United States, the Pacific Ocean, Mid-789
dle East, western Africa, and south-east Australia. Thus, the chlorophyll-a decrease was790
probably a global phenomenon correlated with anthropogenic activity reduction (Shehhi791
and Samad, 2021).792
Thus, our analysis indicates that the COVID-19 pandemic likely resulted in modify-793
ing three species habitats among those studied: it positively affected the distributions of794
Merluccius merluccius and Mullus barbatus, but negatively the one of Sardina pilchardus.795
4. Discussion and Conclusions796
This paper has presented an analysis of habitat change in 2020 with respect to the pre-797
vious years (2015-2018 aggregated and 2019), based on floating sensor information and798
species occurrence records from the OBIS data collection. Our experiment estimated the799
habitat of 8 commercial species of the Adriatic Sea over this period. The produced eco-800
logical niche distributions were sufficiently reliable when compared to those produced by801
an independent model. They were similar to a model based on 2019 environmental con-802
ditions (AquaMaps 2019) and very distant from a model based on a currently improbable803
environmental scenario (AquaMaps 2050).804
Our distributions were suitable for a pattern analysis to investigate if habitat change805
depended on climate change or the COVID-19 pandemic. The main parameters that influ-806
enced habitat change were the general increase of temperature and the overall decrease of807
dissolved oxygen and chlorophyll-a. Although the observed temperature and DOX trends808
depend on climate change, the chlorophyll-a decrease in 2020 was likely a consequence809
of the COVID-19 pandemic.810
Although some species - Solea solea,Squilla mantis, and Engraulis encrasicolus -811
were not significantly affected by these changes, heterogeneous effects on the other species812
habitat were observed. The increasing temperature and decreasing DOX trends - i.e., the813
potential effects of climate change - negatively affected the distribution of Parapenaeus814
longirostris by making its habitat overall unstable and less suitable in 2020 than in 2019.815
This potential negative dependency on climate change finds confirmation by several stud-816
ies on this species (Ungaro and Gramolini, 2006; Colloca et al., 2014; Sbrana et al., 2019;817
Quattrocchi et al., 2020). Conversely, these trends favoured Sepia officinalis and extended818
its potential habitat, in agreement with other studies that analysed its response to the single819
parameter changes (Palmegiano and d’Apote, 1983; Capaz et al., 2017).820
The potential coupling between climate change and COVID-19 - manifested as a821
simultaneous decreasing trend of DOX and chlorophyll-a - negatively affected the dis-822
tribution of Sardina pilchardus. Other studies have also reported habitat instability of823
this species’ habitat as the consequence of the variation of these parameters (Sinovˇ
2001; Ganias, 2009). However, the combination of rising temperature and decreasing825
chlorophyll-a positively affected the habitats of Merluccius merluccius and Mullus barba-826
tus. This observation agrees with parameter-specific indications by other studies (Gucu827
and Bingel, 2011; García-Rodríguez et al., 2011; Sabates et al., 2015; Sion et al., 2019).828
These two species were the major beneficiary of the two parameter trend combination.829
Thus, reduced anthropogenic stress on ecosystems in 2020 was beneficial for some species’830
4.1. Reusability and limitations of the approach832
Our approach predicted potential general consequences of climate change on species833
habitat and its coupling with the COVID-19 pandemic. In this view, it can be useful for834
integrated environmental assessments (Antunes and Santos, 1999; Kristensen, 2004). For835
example, it can be combined with human activity analysis and when estimating available836
biomass, and can be used in models that predict risk of regime shift caused by habitat loss837
(deyoung et al., 2008; Graham et al., 2015; Wernberg et al., 2016). Notably, the potential838
effects of reduced fishing activity - due to sanitary restrictions and market closure - on839
habitat distributions are yet unclear. Only a 10% reduction of fishing hours with respect to840
the 2019 level has been estimated globally (for large and small scale fisheries) (Clavelle,841
2020; WWF, 2020). Furthermore, the overall fishing activity reduction was just 4% in the842
Italian seas (Clavelle, 2020). Such a low reduction possibly had minor effects on the habi-843
tat distributions of our analysed species and will be the subject of our future investigations.844
Our approach is also general enough to be applied to other species and areas. To this aim,845
our workflow uses FAIR data that have a global-scale coverage. Furthermore, our software846
is open source, and all data are reported under the ESRI-grid format (see Supplementary847
Material). Specifically, the optimal MaxEnt models and the data are all available as raster848
ESRI-grid files in the repository linked in the Supplementary Material, within the "Phase849
4 - MaxEnt Re-application/MaxEnt Distributions and Statistics" folder, for re-use in GIS850
software and other experiments.851
The main limitation of our experiment is the low amount of data used, due to current852
data availability, which was partially compensated by accurate data selection and model853
optimisation. Although the proposed Adriatic-scale pattern analysis is reliable enough to854
extract habitat change trends, the produced maps cannot be considered punctually reliable855
(Queiroz et al., 2021). Conducting a precise analysis will require collecting, collating,856
and analysing a massive amount of data that will be available only years after the end of857
the pandemic. Nevertheless, data-poor approaches like ours can predict realistic macro-858
scopic patterns and indicate priority directions for investigating species modifications in859
the search for confirmation or confutation of the reported results (Coro et al., 2015b,860
2016a). In this view, our model allows looking ahead to the possible significant modi-861
fications that will possibly be observed in the Adriatic in the following years due to the862
impact of the combined action of the COVID-19 pandemic and climate change on species863
distributions. Small-scale reliability can also be enhanced in our model when marine en-864
vironmental data and species records will be more dense and uniform in the study area.865
Several initiatives are promoting the collection of these data (EU Commission, 2020a;866
Snapshot-CNR, 2020; EU Commission, 2020b), but they are ongoing and main address867
regional scales. These data will be a fundamental source of information to repeat our anal-868
ysis and validate its predictions. We believe that these activities are justified to understand869
the effects of natural and man-made pressure on marine ecosystems in current and future870
scenarios. Our study also confirmed that in order to realise the UN Decade on Ecosystem871
Restoration motto "the science we need for the ocean we want" (United Nations, 2021b)872
an Open Science approach can be successful.873
Supplementary Material874
Experimental data and source code are publicly available on the D4Science e-Infrastructure875
The authors acknowledge Enrico Nicola Armelloni and Giuseppe Scarcella for indi-878
cations about Adriatic fisheries. This research was partially funded by the SNAPSHOT879
project of the National Research Council of Italy (CNR) and by the Blue Cloud EU project880
(Grant Agreement No.862409).881
Adwibowo, A., 2020. Does social distancing have an effect on water quality?: An evidence883
from chlorophyll-a level in the water of populated southeast asian coasts. https:884
Alvera-Azcárate, A., Barth, A., Rixen, M., Beckers, J.M., 2005. Reconstruction of incom-886
plete oceanographic data sets using empirical orthogonal functions: application to the887
adriatic sea surface temperature. Ocean Modelling 9, 325–346.888
Angeletti, L., Prampolini, M., Foglini, F., Grande, V., Taviani, M., 2020. Cold-water coral889
habitat in the bari canyon system, southern adriatic sea (mediterranean sea), in: Seafloor890
Geomorphology as Benthic Habitat. Elsevier, pp. 811–824.891
Antunes, P., Santos, R., 1999. Integrated environmental management of the oceans. Eco-892
logical Economics 31, 215–226.893
AquaMaps, 2020. AquaMaps Web site.
Araujo, M.B., Naimi, B., 2020. Spread of sars-cov-2 coronavirus likely to be constrained895
by climate. doi:10.1101/2020.03.12.20034728. preprint.896
Ardizzone, G., Gravina, M., Belluscio, A., Schintu, P., 1990. Depth-size distribution897
pattern of parapenaeus longirostris (lucas, 1846)(decapoda) in the central mediterranean898
sea. Journal of Crustacean Biology 10, 139–147.899
Argo, 2000. Argo float data and metadata from global data assembly centre (argo gdac).900
Arif, M., Kumar, R., Parveen, S., Verma, N., 2020. Reduction in water pollution in yamuna902
river due to lockdown under covid-19 pandemic. The Pharma Innovation Journal 9, 84–903
Arnold, T.W., 2010. Uninformative parameters and model selection using akaike’s infor-905
mation criterion. The Journal of Wildlife Management 74, 1175–1178.906
Ashraf, U., Peterson, A.T., Chaudhry, M.N., Ashraf, I., Saqib, Z., Rashid Ahmad, S., Ali,907
H., 2017. Ecological niche model comparison under different climate scenarios: a case908
study of olea spp. in asia. Ecosphere 8, e01825.909
Assante, M., Candela, L., Castelli, D., Cirillo, R., Coro, G., Frosini, L., Lelii, L., Man-910
giacrapa, F., Pagano, P., Panichi, G., et al., 2019. Enacting open science by d4science.911
Future Generation Computer Systems 101, 555–563.912
Azzolin, M., Arcangeli, A., Cipriano, G., Crosti, R., Maglietta, R., Pietroluongo, G.,913
Saintingan, S., Zampollo, A., Fanizza, C., Carlucci, R., 2020. Spatial distribution mod-914
elling of striped dolphin (stenella coeruleoalba) at different geographical scales within915
the eu adriatic and ionian sea region, central-eastern mediterranean sea. Aquatic Con-916
servation: Marine and Freshwater Ecosystems 30, 1194–1207.917
Baker, R.E., Peña, J.M., Jayamohan, J., Jérusalem, A., 2018. Mechanistic models versus918
machine learning, a fight worth fighting for the biological community? Biology letters919
14, 20170660.920
Bargain, A., Marchese, F., Savini, A., Taviani, M., Fabri, M.C., 2017. Santa maria di leuca921
province (mediterranean sea): Identification of suitable mounds for cold-water coral922
settlement using geomorphometric proxies and maxent methods. Frontiers in Marine923
Science 4, 338.924
Barth, A., Alvera-Azcárate, A., Troupin, C., Ouberdous, M., Beckers, J.M., 2010. A925
web interface for griding arbitrarily distributed in situ data based on data-interpolating926
variational analysis (diva). Advances in Geosciences 28, 29–37.927
Ben Rais Lasram, F., Guilhaumon, F., Albouy, C., Somot, S., Thuiller, W., Mouillot, D.,928
2010. The mediterranean sea as a ‘cul-de-sac’for endemic fishes facing climate change.929
Global Change Biology 16, 3233–3245.930
Bennett, A.F., 1992. Inverse methods in physical oceanography. Cambridge university931
Blackford, J., 2002. The influence of microphytobenthos on the northern adriatic ecosys-933
tem: a modelling study. Estuarine, Coastal and Shelf Science 55, 109–123.934
Brown, C., Fulton, E., Hobday, A., Matear, R., Possingham, H., Bulman, C., Christensen,935
V., Forrest, R., Gehrke, P., Gribble, N., et al., 2010. Effects of climate-driven primary936
production change on marine food webs: implications for fisheries and conservation.937
Global Change Biology 16, 1194–1212.938
Candela, L., Castelli, D., Coro, G., Pagano, P., Sinibaldi, F., 2016. Species distribution939
modeling in the cloud. Concurrency and Computation: Practice and Experience 28,940
Capaz, J.C., Tunnah, L., MacCormack, T.J., Lamarre, S.G., Sykes, A.V., Driedzic, W.R.,942
2017. Hypoxic induced decrease in oxygen consumption in cuttlefish (sepia officinalis)943
is associated with minor increases in mantle octopine but no changes in markers of944
protein turnover. Frontiers in physiology 8, 344.945
Capezzuto, F., Sion, L., Ancona, F., Carlucci, R., Carluccio, A., Cornacchia, L., Maiorano,946
P., Ricci, P., Tursi, A., D’Onghia, G., 2018. Cold-water coral habitats and canyons as947
essential fish habitats in the southern adriatic and northern ionian sea (central mediter-948
ranean). Ecological Questions 29, 9–23.949
Chala, D., Roos, C., Svenning, J.C., Zinner, D., 2019. Species-specific effects of climate950
change on the distribution of suitable baboon habitats–ecological niche modeling of951
current and last glacial maximum conditions. Journal of human evolution 132, 215–952
Chunco, A.J., Phimmachak, S., Sivongxay, N., Stuart, B.L., 2013. Predicting environ-954
mental suitability for a rare and threatened species (lao newt, laotriton laoensis) using955
validated species distribution models. PLoS One 8, e59853.956
Clavelle, T., 2020. Global fisheries during COVID-957
Cobos, M.E., Peterson, A.T., Barve, N., Osorio-Olvera, L., 2019. kuenm: an r package for960
detailed development of ecological niche models using maxent. PeerJ 7, e6281.961
Cohen, J., et al., 1960. A coefficient of agreement for nominal scales. Educational and962
psychological measurement 20, 37–46.963
Coll, M., Santojanni, A., Palomera, I., Tudela, S., Arneri, E., 2007. An ecological model964
of the northern and central adriatic sea: analysis of ecosystem structure and fishing965
impacts. Journal of Marine Systems 67, 119–154.966
Colloca, F., Mastrantonio, G., Lasinio, G.J., Ligas, A., Sartor, P., 2014. Parapenaeus lon-967
girostris (lucas, 1846) an early warning indicator species of global warming in the cen-968
tral mediterranean sea. Journal of Marine Systems 138, 29–39. URL: https://www.969,970
doi: the wrapping971
up of the IDEADOS project: International Workshop on Environment, Ecosystems and972
Demersal Resources, and Fisheries.973
Coro, G., 2020. A global-scale ecological niche model to predict sars-cov-2 coronavirus974
infection rate. Ecological Modelling 431, 109187.975
Coro, G., Bove, P., 2022. A high-resolution global-scale model for covid-19 infection rate.976
ACM Transactions on Spatial Algorithms and Systems (TSAS) 8, 1–24.977
Coro, G., Candela, L., Pagano, P., Italiano, A., Liccardo, L., 2015a. Parallelizing the978
execution of native data mining algorithms for computational biology. Concurrency979
and Computation: Practice and Experience 27, 4630–4644.980
Coro, G., Ellenbroek, A., Pagano, P., 2021. An open science approach to infer fishing981
activity pressure on stocks and biodiversity from vessel tracking data. Ecological Infor-982
matics 64, 101384.983
Coro, G., Magliozzi, C., Ellenbroek, A., Kaschner, K., Pagano, P., 2016a. Automatic984
classification of climate change effects on marine species distributions in 2050 using985
the aquamaps model. Environmental and ecological statistics 23, 155–180.986
Coro, G., Magliozzi, C., Ellenbroek, A., Pagano, P., 2015b. Improving data quality to build987
a robust distribution model for architeuthis dux. Ecological modelling 305, 29–39.988
Coro, G., Magliozzi, C., Vanden Berghe, E., Bailly, N., Ellenbroek, A., Pagano,989
P., 2016b. Estimating absence locations of marine species from data of scien-990
tific surveys in obis. Ecological Modelling 323, 61–76. URL: https://www.991,992
Coro, G., Pagano, P., Ellenbroek, A., 2013a. Automatic procedures to assist in manual994
review of marine species distribution maps, in: International Conference on Adaptive995
and Natural Computing Algorithms, Springer. pp. 346–355.996
Coro, G., Pagano, P., Ellenbroek, A., 2013b. Combining simulated expert knowledge with997
neural networks to produce ecological niche models for latimeria chalumnae. Ecological998
modelling 268, 55–63.999
Coro, G., Pagano, P., Ellenbroek, A., 2014. Comparing heterogeneous distribution maps1000
for marine species. GIScience & Remote Sensing 51, 593–611.1001
Coro, G., Pagano, P., Ellenbroek, A., 2018a. Detecting patterns of climate change in long-1002
term forecasts of marine environmental parameters. International Journal of Digital1003
Earth , 1–19.1004
Coro, G., Pagano, P., Ellenbroek, A., 2020. Detecting patterns of climate change in long-1005
term forecasts of marine environmental parameters. International Journal of Digital1006
Earth 13, 567–585.1007
Coro, G., Pagano, P., Napolitano, U., 2016c. Bridging environmental data providers and1008
seadatanet diva service within a collaborative and distributed e-infrastructure. Bollettino1009
di Geofisica , 23–25.1010
Coro, G., Panichi, G., Scarponi, P., Pagano, P., 2017. Cloud computing in a distributed1011
e-infrastructure using the web processing service standard. Concurrency and Computa-1012
tion: Practice and Experience 29, e4219.1013
Coro, G., Scarponi, P., Pagano, P., 2018b. Enhancing argo floats data re-usability. Bollet-1014
tino di Geofisica 53.1015
Coro, G., Tassetti, A.N., Armelloni, E.N., Pulcinella, J., Ferrà, C., Sprovieri, M., Trincardi,1016
F., Scarcella, G., 2022. Covid-19 lockdowns reveal the resilience of adriatic sea fisheries1017
to forced fishing effort reduction. Scientific Reports 12, 1–14.1018
Coro, G., Trumpy, E., 2020. Predicting geographical suitability of geothermal power1019
plants. Journal of Cleaner Production 267, 121874.1020
Coro, G., Vilas, L.G., Magliozzi, C., Ellenbroek, A., Scarponi, P., Pagano, P., 2018c.1021
Forecasting the ongoing invasion of lagocephalus sceleratus in the mediterranean sea.1022
Ecological Modelling 371, 37–49.1023
Coro, G., Webb, T.J., Appeltans, W., Bailly, N., Cattrijsse, A., Pagano, P.,1024
2015c. Classifying degrees of species commonness: North sea fish as a1025
case study. Ecological Modelling 312, 272–280. URL: https://www.1026,1027
Corsi, F., de Leeuw, J., Skidmore, A., 2000. Modeling species distribution with GIS.1029
Research Techniques in Animal Ecology. Columbia University Press, New York n/a,1030
Deneu, B., Servajean, M., Bonnet, P., Botella, C., Munoz, F., Joly, A., 2021. Convolu-1032
tional neural networks improve species distribution modelling by capturing the spatial1033
structure of the environment. PLOS Computational Biology 17, e1008856.1034
Depellegrin, D., Bastianini, M., Fadini, A., Menegon, S., 2020. The effects of covid-191035
induced lockdown measures on maritime settings of a coastal region. Science of the1036
Total Environment 740, 140123.1037
deyoung, B., Barange, M., Beaugrand, G., Harris, R., Perry, R.I., Scheffer, M., Werner,1038
F., 2008. Regime shifts in marine ecosystems: detection, prediction and management.1039
Trends in Ecology & Evolution 23, 402–409.1040
Djakovac, T., Supi´
c, N., Aubry, F.B., Degobbis, D., Giani, M., 2015. Mechanisms of1041
hypoxia frequency changes in the northern adriatic sea during the period 1972–2012.1042
Journal of Marine Systems 141, 179–189.1043
DownToEarth, . Upper oceans hottest in 2020 despite lower emissions due to COVID-1044
19 lockdowns.
Dudík, M., Phillips, S., Schapire, R.E., 2005. Correcting sample selection bias in maxi-1047
mum entropy density estimation. Advances in neural information processing systems1048
18, 323–330.1049
Durand, M., Fu, L.L., Lettenmaier, D.P., Alsdorf, D.E., Rodriguez, E., Esteban-Fernandez,1050
D., 2010. The surface water and ocean topography mission: Observing terrestrial sur-1051
face water and oceanic submesoscale eddies. Proceedings of the IEEE 98, 766–779.1052
Elith, J., Graham, C.H., 2009. Do they? how do they? why do they differ? on finding1053
reasons for differing performances of species distribution models. Ecography 32, 66–1054
Elith, J., Phillips, S.J., Hastie, T., Dudík, M., Chee, Y.E., Yates, C.J., 2011. A statistical1056
explanation of maxent for ecologists. Diversity and distributions 17, 43–57.1057
Esposito, V., Andaloro, F., Bianca, D., Natalotto, A., Romeo, T., Scotti, G., Castriota, L.,1058
2014. Diet and prey selectivity of the red mullet, mullus barbatus (pisces: Mullidae),1059
from the southern tyrrhenian sea: the role of the surf zone as a feeding ground. Marine1060
Biology Research 10, 167–178.1061
EU Commission, 2020a. A European Green Deal.
EU Commission, 2020b. Mission Starfish 2030: Restore our Ocean1064
and Waters.
FAO, 2020. The State of Mediterranean and Black Sea Fisheries 2020. https://doi.1067
Friedlaender, A.S., Johnston, D.W., Fraser, W.R., Burns, J., Costa, D.P., et al., 2011.1069
Ecological niche modeling of sympatric krill predators around marguerite bay, west-1070
ern antarctic peninsula. Deep Sea Research Part II: Topical Studies in Oceanography1071
58, 1729–1740.1072
Froese, R., Winker, H., Coro, G., Demirel, N., Tsikliras, A.C., Dimarchopoulou, D., Scar-1073
cella, G., Quaas, M., Matz-Lück, N., 2018. Status and rebuilding of european fisheries.1074
Marine Policy 93, 159–170.1075
Ganias, K., 2009. Linking sardine spawning dynamics to environmental variabil-1076
ity. Estuarine, Coastal and Shelf Science 84, 402–408. URL: https://www.1077,1078
Garcia, D.A., Amori, M., Giovanardi, F., Piras, G., Groppi, D., Cumo, F., De Santoli, L.,1080
2019. An identification and a prioritisation of geographic and temporal data gaps of1081
mediterranean marine databases. Science of the total environment 668, 531–546.1082
García-Rodríguez, M., Fernández, A., Esteban, A., 2011. Biomass response to environ-1083
mental factors in two congeneric species of mullus, m. barbatus and m. surmuletus,1084
off catalano–levantine mediterranean coast of spain: a preliminary approach. Animal1085
Biodiversity and Conservation 34, 113–122.1086
GEBCO, 2020. Gridded bathymetry data.
GFCM, 2020. Geographical subareas of the General Fisheries Commission for the1089
Graham, N.A., Jennings, S., MacNeil, M.A., Mouillot, D., Wilson, S.K., 2015. Predicting1091
climate-driven regime shifts versus rebound potential in coral reefs. Nature 518, 94–97.1092
Grassle, J., 2000. The ocean biogeographic information system (obis): an on-line,1093
worldwide atlas for accessing, modeling and mapping marine biological data in1094
a multidimensional geographic context. OCEANOGRAPHY-WASHINGTON DC-1095
Gucu, A., Bingel, F., 2011. Hake, merluccius merluccius l., in the northeastern mediter-1097
ranean sea: a case of disappearance. Journal of Applied Ichthyology 27, 1001–1012.1098
Guo, Q., Liu, Y., 2010. Modeco: an integrated software package for ecological niche1099
modeling. Ecography 33, 637–642.1100
Guthery, F.S., Brennan, L.A., Peterson, M.J., Lusk, J.J., 2005. Information theory in1101
wildlife science: critique and viewpoint. The Journal of Wildlife Management 69, 457–1102
Hengl, T., Sierdsema, H., Radovi´
c, A., Dilo, A., 2009. Spatial prediction of species’1104
distributions from occurrence-only records: combining point pattern analysis, enfa and1105
regression-kriging. Ecological modelling 220, 3499–3511.1106
Huang, Y.P., Kao, L.J., Sandnes, F.E., 2008. Efficient mining of salinity and temperature1107
association rules from argo data. Expert Systems with Applications 35, 59–68.1108
Jones, M.C., Dye, S.R., Pinnegar, J.K., Warren, R., Cheung, W.W., 2012. Modelling1109
commercial fish distributions: Prediction and assessment using different approaches.1110
Ecological Modelling 225, 133–145.1111
c, D., Legovi´
c, T., Rottini-Sandrini, L., 1987. Trends in oxygen content1112
1911–1984 and occurrence of benthic mortality in the northern adriatic sea.1113
Estuarine, Coastal and Shelf Science 25, 435–445. URL: https://www.1114,1115
Kaschner, K., Ready, J., Agbayani, E., Kesner-Reyes, K., Rius-Barile, J., Eastwood, P.,1117
South, A., Kullander, S., Rees, T., Watson, R., et al., 2011. Using ’aquamaps’ for1118
representing species distribution in regional seas, in: Fisheries Centre Research Reports.1119
Fisheries Centre, University of British Columbia, pp. 17–21.1120
Kaschner, K., Watson, R., Trites, A., Pauly, D., 2006. Mapping world-wide distributions of1121
marine mammal species using a relative environmental suitability (res) model. Marine1122
Ecology Progress Series 316, 285–310.1123
Kemp, P.S., Froese, R., Pauly, D., 2020. Covid-19 provides an opportunity to advance1124
a sustainable uk fisheries policy in a post-brexit brave new world. Marine Policy 120,1125
Kralj, M., Lipizer, M., ˇ
Cermelj, B., Celio, M., Fabbro, C., Brunetti, F., Francé, J., Mozetiˇ
P., Giani, M., 2019a. Hypoxia and dissolved oxygen trends in the northeastern adriatic1128
sea (gulf of trieste). Deep Sea Research Part II: Topical Studies in Oceanography 164,1129
Kralj, M., Lipizer, M., ˇ
Cermelj, B., Celio, M., Fabbro, C., Brunetti, F., Francé,1131
J., Mozetiˇ
c, P., Giani, M., 2019b. Hypoxia and dissolved oxygen trends1132
in the northeastern adriatic sea (gulf of trieste). Deep Sea Research Part1133
II: Topical Studies in Oceanography 164, 74–88. URL: https://www.1134,1135
doi: revisiting the1136
Eastern Mediterranean: Recent knowledge on the physical, biogeochemical and1137
ecosystemic states and trends (Volume I).1138
Kristensen, P., 2004. The dpsir framework, european topic centre on water. European1139
Environment Agency , 1–10.1140
Landis, J.R., Koch, G.G., 1977. The measurement of observer agreement for categorical1141
data. biometrics , 159–174.1142
Le Quéré, C., Jackson, R.B., Jones, M.W., Smith, A.J., Abernethy, S., Andrew, R.M., De-1143
Gol, A.J., Willis, D.R., Shan, Y., Canadell, J.G., et al., 2020. Temporary reduction in1144
daily global co 2 emissions during the covid-19 forced confinement. Nature Climate1145
Change 10, 647–653.1146
Lipizer, M., Partescano, E., Rabitti, A., Giorgetti, A., Crise, A., 2014. Qualified temper-1147
ature, salinity and dissolved oxygen climatologies in a changing adriatic sea. Ocean1148
Science 10, 771–797.1149
Magliozzi, C., Coro, G., Grabowski, R.C., Packman, A.I., Krause, S., 2019. A multiscale1150
statistical method to identify potential areas of hyporheic exchange for river restoration1151
planning. Environmental Modelling & Software 111, 311–323.1152
c, I., Ninˇ
c, Ž., Kušpili´
c, G., Marinovi´
c, S., Marinov, S., 2005. Long-term1153
changes of basic biological and chemical parameters at two stations in the middle adri-1154
atic. Journal of Sea Research 54, 3–14.1155
Menchetti, M., Guéguen, M., Talavera, G., 2019. Spatio-temporal ecological niche mod-1156
elling of multigenerational insect migrations. Proceedings of the Royal Society B 286,1157
Merow, C., Smith, M.J., Silander Jr, J.A., 2013. A practical guide to maxent for modeling1159
species’ distributions: what it does, and why inputs and settings matter. Ecography 36,1160
Mishra, D.R., Kumar, A., Muduli, P.R., Equeenuddin, S.M., Rastogi, G., Acharyya, T.,1162
Swain, D., 2020. Decline in phytoplankton biomass along indian coastal waters due1163
to covid-19 lockdown. Remote Sensing 12. URL:
2072-4292/12/16/2584, doi:10.3390/rs12162584.1165
Morales, N.S., Fernández, I.C., Baca-González, V., 2017. Maxent’s parameter configu-1166
ration and small samples: are we paying attention to recommendations? a systematic1167
review. PeerJ 5, e3093.1168
Muscarella, R., Galante, P.J., Soley-Guardia, M., Boria, R.A., Kass, J.M., Uriarte, M.,1169
Anderson, R.P., 2014. Enm eval: An r package for conducting spatially independent1170
evaluations and estimating optimal model complexity for maxent ecological niche mod-1171
els. Methods in ecology and evolution 5, 1198–1205.1172
O’Brien, P.C., 1980. The quartiles of the maximum entropy distribution. Economics1173
Letters 6, 49–52.1174
OSGeo, 2019. GDAL - Geospatial Data Abstraction Library. https://www.gdal.1175
Palmegiano, G., d’Apote, M., 1983. Combined effects of temperature and salinity on1177
cuttlefish (sepia officinalis l.) hatching. Aquaculture 35, 259–264.1178
Pearson, R.G., 2007. Species’ distribution modeling for conservation educators and prac-1179
titioners. Synthesis. American Museum of Natural History 50, 54–89.1180
Peterson, A.T., 2001. Predicting Species’ Geographic Distributions Based on Ecological1181
Niche Modeling. The Condor 103, 599–605. URL:
condor/103.3.599, doi:10.1093/condor/103.3.599.1183
Peterson, A.T., 2003. Predicting the geography of species’ invasions via ecological niche1184
modeling. The quarterly review of biology 78, 419–433.1185
Peterson, T., Pape¸s, M., Eaton, M., 2007. Transferability and model evaluation in ecolog-1186
ical niche modeling: a comparison of garp and maxent. Ecography 30, 550–560.1187
Phillips, S., Dudik, M., Schapire, R., 2021. Maxent software for modeling species niches1188
and distributions (Version 3.4.1). http://biodiversityinformatics.amnh.1189
Phillips, S.J., Anderson, R.P., Dudík, M., Schapire, R.E., Blair, M.E., 2017. Opening the1191
black box: An open-source release of maxent. Ecography 40, 887–893.1192
Phillips, S.J., Anderson, R.P., Schapire, R.E., 2006a. Maximum entropy modeling of1193
species geographic distributions. Ecological Modelling 190, 231–259.1194
Phillips, S.J., Anderson, R.P., Schapire, R.E., 2006b. Maximum entropy mod-1195
eling of species geographic distributions. Ecological Modelling 190, 231 –1196
259. URL:
Phillips, S.J., Dudík, M., 2008. Modeling of species distributions with maxent: new ex-1199
tensions and a comprehensive evaluation. Ecography 31, 161–175.1200
Quattrocchi, F., Fiorentino, F., Lauria, V., Garofalo, G., 2020. The increasing tempera-1201
ture as driving force for spatial distribution patterns of parapenaeus longirostris (lucas1202
1846) in the strait of sicily (central mediterranean sea). Journal of Sea Research 158,1203
101871. URL:
pii/S1385110119302485, doi:
Queiroz, N., Humphries, N.E., Couto, A., Vedor, M., Da Costa, I., Sequeira, A.M., Mu-1207
cientes, G., Santos, A.M., Abascal, F.J., Abercrombie, D.L., et al., 2021. Reply to:1208
Caution over the use of ecological big data for conservation. Nature 595, E20–E28.1209
Ravdas, M., Zacharioudaki, A., Korres, G., 2018. Implementation and validation of a1210
new operational wave forecasting system of the mediterranean monitoring and forecast-1211
ing centre in the framework of the copernicus marine environment monitoring service.1212
Natural Hazards and Earth System Sciences 18, 2675–2695.1213
Raybaud, V., Beaugrand, G., Dewarumez, J.M., Luczak, C., 2015. Climate-induced range1214
shifts of the american jackknife clam ensis directus in europe. Biological Invasions 17,1215
Ready, J., Kaschner, K., South, A.B., Eastwood, P.D., Rees, T., Rius, J., Agbayani, E.,1217
Kullander, S., Froese, R., 2010. Predicting the distributions of marine organisms at the1218
global scale. Ecological Modelling 221, 467–478.1219
Rees, T., 2008. 18.9. using aquamaps for biodiversity assessment, including a prototype1220
mpa (marine protected area) network design tool. The Proceedings of TDWG , 73.1221
Renner, I.W., Warton, D.I., 2013. Equivalence of maxent and poisson point process models1222
for species distribution modeling in ecology. Biometrics 69, 274–281.1223
Reside, A.E., Critchell, K., Crayn, D.M., Goosem, M., Goosem, S., Hoskin, C.J., Sydes,1224
T., Vanderduys, E.P., Pressey, R.L., 2019. Beyond the model: expert knowledge im-1225
proves predictions of species’ fates under climate change. Ecological Applications 29,1226
Reyes, K., 2015. AquaMaps: Algorithm and Data Sources for Aquatic Organ-1228
Roy-Dufresne, E., Saltré, F., Cooke, B.D., Mellin, C., Mutze, G., Cox, T., Fordham, D.A.,1231
2019. Modeling the distribution of a wide-ranging invasive species using the sampling1232
efforts of expert and citizen scientists. Ecology and Evolution 9, 11053–11063.1233
Russo, E., Monti, M.A., Mangano, M.C., Raffaeta, A., Sara, G., Silvestri, C., Pranovi, F.,1234
2020. Temporal and spatial patterns of trawl fishing activities in the adriatic sea (central1235
mediterranean sea, gsa17). Ocean & Coastal Management 192, 105231.1236
Sabates, A., Zaragoza, N., Raya, V., 2015. Distribution and feeding dynamics of larval1237
red mullet (Mullus barbatus) in the NW Mediterranean: the important role of cladocera.1238
Journal of Plankton Research 37, 820–833. URL:
plankt/fbv040, doi:10.1093/plankt/fbv040.1240
Sánchez-Tapia, A., de Siqueira, M.F., Lima, R.O., Barros, F.S.M., Gall, G.M., Gadelha,1241
L.M., da Silva, L.A.E., Osthoff, C., 2017. Model-r: a framework for scalable and repro-1242
ducible ecological niche modeling, in: Latin American High Performance Computing1243
Conference, Springer. pp. 218–232.1244
Santos, A.M.P., Ré, P., Dos Santos, A., Peliz, Á., 2006. Vertical distribution of the euro-1245
pean sardine (sardina pilchardus) larvae and its implications for their survival. Journal1246
of Plankton Research 28, 523–532.1247
Sbrana, M., Zupa, W., Ligas, A., Capezzuto, F., Chatzispyrou, A., Follesa, M.C., Ganci-1248
tano, V., Guijarro, B., Isajlovic, I., Jadaud, A., et al., 2019. Spatiotemporal abun-1249
dance pattern of deep-water rose shrimp, parapenaeus longirostris, and norway lobster,1250
nephrops norvegicus, in european mediterranean waters. Scientia Marina 83, 71–80.1251
Scarponi, P., Coro, G., Pagano, P., 2018. A collection of aquamaps native layers in netcdf1252
format. Data in brief 17, 292–296.1253
Schaap, D.M., Lowry, R.K., 2010. Seadatanet–pan-european infrastructure for marine and1254
ocean data management: unified access to distributed data sets. International Journal of1255
Digital Earth 3, 50–69.1256
Schnase, J.L., Carroll, M.L., Gill, R.L., Tamkin, G.S., Li, J., Strong, S.L., Maxwell, T.P.,1257
Aronne, M.E., Spradlin, C.S., 2021. Toward a monte carlo approach to selecting climate1258
variables in maxent. PloS one 16, e0237208.1259
Schulzweida, U., 2020. CDO User Guide.
Shehhi, M.R.A., Samad, Y.A., 2021. Effects of the covid-19 pandemic on the oceans.1262
Remote Sensing Letters 12, 325–334. doi:10.1080/2150704X.2021.1880658.1263
c, G., 2001. Biotic and abiotic factors influencing sardine, sardina pilchardus1264
(walb.) abundance in the croatian part of the eastern adriatic. FAO Adriamed paper1265
Sion, L., Zupa, W., Calculli, C., Garofalo, G., Hidalgo, M., Jadaud, A., Lefkaditou, E.,1267
Ligas, A., Peristeraki, P., Bitetto, I., et al., 2019. Spatial distribution pattern of european1268
hake, merluccius merluccius (pisces: Merlucciidae), in the mediterranean sea. Scientia1269
Marina 83, 21–32.1270
de Siqueira, M.F., Durigan, G., de Marco Júnior, P., Peterson, A.T., 2009. Something from1271
nothing: using landscape similarity and ecological niche modeling to find rare plant1272
species. Journal for Nature Conservation 17, 25–32.1273
Snapshot-CNR, 2020. Synoptic Assessment of Human Pressures on Key Mediterranean1274
Hot Spots: The Snapshot-CNR project.
Tanhua, T., Pouliquen, S., Hausman, J., O’brien, K., Bricher, P., De Bruin, T., Buck, J.J.,1276
Burger, E.F., Carval, T., Casey, K.S., et al., 2019. Ocean fair data services. Frontiers in1277
Marine Science 6, 440.1278
Theil, H., 1982. Some recent and new results on the maximum entropy distribution. Statis-1279
tics & Probability Letters 1, 17–22.1280
Toonen, H.M., Bush, S.R., 2020. The digital frontiers of fisheries governance: Fish at-1281
traction devices, drones and satellites. Journal of environmental policy & planning 22,1282
Trifonova, N., Maxwell, D., Pinnegar, J., Kenny, A., Tucker, A., 2017. Predicting ecosys-1284
tem responses to changes in fisheries catch, temperature, and primary productivity with1285
a dynamic bayesian network model. ICES Journal of Marine Science 74, 1334–1343.1286
Troupin, C., Barth, A., Sirjacobs, D., Ouberdous, M., Brankart, J.M., Brasseur, P., Rixen,1287
M., Alvera-Azcárate, A., Belounis, M., Capet, A., et al., 2012. Generation of analysis1288
and consistent error fields using the data interpolating variational analysis (diva). Ocean1289
Modelling 52, 90–101.1290
Troupin, C., Machin, F., Ouberdous, M., Sirjacobs, D., Barth, A., Beckers, J.M., 2010.1291
High-resolution climatology of the northeast atlantic using data-interpolating variational1292
analysis (diva). Journal of Geophysical Research: Oceans 115.1293
Ungaro, N., Gramolini, R., 2006. Possible effect of bottom temperature on distribution1294
of parapenaeus longirostris (lucas, 1846) in the southern adriatic (mediterranean sea).1295
Turkish Journal of Fisheries and Aquatic Sciences 6, –.1296
United Nations, 2021a. Cooling La Niña is on the wane, but temperatures set to rise: UN1297
weather agency.
United Nations, 2021b. UN Decade on Ecosystem Restoration. https://www.1299
Von Schuckmann, K., Le Traon, P.Y., Smith, N., Pascual, A., Brasseur, P., Fennel, K.,1301
Djavidnia, S., Aaboe, S., Fanjul, E.A., Autret, E., et al., 2018. Copernicus marine1302
service ocean state report. Journal of Operational Oceanography 11, S1–S142.1303
Wang, L., Kerr, L.A., Record, N.R., Bridger, E., Tupper, B., Mills, K.E., Armstrong, E.M.,1304
Pershing, A.J., 2018. Modeling marine pelagic fish species spatiotemporal distributions1305
utilizing a maximum entropy approach. Fisheries Oceanography 27, 571–586.1306
Warren, D.L., Seifert, S.N., 2011. Ecological niche modeling in maxent: the importance1307
of model complexity and the performance of model selection criteria. Ecological appli-1308
cations 21, 335–342.1309
Watelet, S., Back, Ö., Barth, A., Beckers, J.M., 2016. Data-interpolating variational anal-1310
ysis (diva) software: recent development and application, in: Proceedings of the EGU1311
General Assembly 2016, European Geosciences Union General Assembly. p. 1.1312
Weatherdon, L.V., Magnan, A.K., Rogers, A.D., Sumaila, U.R., Cheung, W.W., 2016.1313
Observed and projected impacts of climate change on marine fisheries, aquaculture,1314
coastal tourism, and human health: an update. Frontiers in Marine Science , 48.1315
Weber, M.M., Stevens, R.D., Diniz-Filho, J.A.F., Grelle, C.E.V., 2017. Is there a corre-1316
lation between abundance and environmental suitability derived from ecological niche1317
modelling? a meta-analysis. Ecography 40, 817–828.1318
Werdell, P.J., Bailey, S.W., 2005. An improved in-situ bio-optical data set for ocean color1319
algorithm development and satellite data product validation. Remote sensing of envi-1320
ronment 98, 122–140.1321
Wernberg, T., Bennett, S., Babcock, R.C., De Bettignies, T., Cure, K., Depczynski, M.,1322
Dufois, F., Fromont, J., Fulton, C.J., Hovey, R.K., et al., 2016. Climate-driven regime1323
shift of a temperate marine ecosystem. Science 353, 169–172.1324
World Meterological Organization, 2021. Climate change indicators and impacts wors-1325
ened in 2020.
WWF, 2020. World Wide Fund for Nature - Impact of COVID-19 on Mediter-1328
ranean Fisheries.
Yunus, A.P., Masago, Y., Hijioka, Y., 2020. Covid-19 and surface water quality: Improved1332
lake water quality during the lockdown. Science of the Total Environment 731, 139012.1333
Zaniewski, A.E., Lehmann, A., Overton, J.M., 2002. Predicting species spatial distribu-1334
tions using presence-only data: a case study of native new zealand ferns. Ecological1335
modelling 157, 261–280.1336
Zavatarelli, M., Raicich, F., Bregant, D., Russo, A., Artegiani, A., 1998. Cli-1337
matological biogeochemical characteristics of the adriatic sea. Journal of1338
Marine Systems 18, 227–263. URL: https://www.sciencedirect.1339
com/science/article/pii/S0924796398000141, doi:https:1340
Zeng, Y., Low, B.W., Yeo, D.C., 2016. Novel methods to select environmental variables1342
in maxent: A case study using invasive crayfish. Ecological Modelling 341, 5–13.1343
Zhang, G., Zhu, A.X., Windels, S.K., Qin, C.Z., 2018. Modelling species habitat suitabil-1344
ity from presence-only data using kernel density estimation. Ecological Indicators 93,1345
Table 1: Discrepancy between the ecological niche models of the eight species involved in our experiment. Model names refer to floating sensor
models for 2015-2018 (FS 2015-2018), 2019 (FS 2019), 2020 (FS 2020), and AquaMaps 2019 and 2050. Coloured numbers refer to Cohen’s
kappa values corresponding to at-least-moderate (green), slight (orange), or poor agreement (red) according to Landis & Koch interpretation.
Bold-highlighted text indicates the most similar distribution for each model. Coloured species names indicate habitat gain (green), change (red),
or stability (blue) in 2020 with respect to 2015-2018.
Sepia officinalis
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 10.06% 12.36% 15.24% 21.71%
FS 2019 10.06% -7.18% 15.50% 22.71%
FS 2020 12.36% 7.18% -14.87% 24.72%
AquaMaps 2019 15.24% 15.50% 14.87% -43.14%
AquaMaps 2050 21.71% 22.71% 24.72% 43.14% -
Merluccius merluccius
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 18.53% 17.82% 25.92% 34.25%
FS 2019 18.53% -5.89% 22.61% 40.90%
FS 2020 17.82% 5.89% -22.47% 41.03%
AquaMaps 2019 25.92% 22.61% 22.47% - 52.26%
AquaMaps 2050 34.25% 40.90% 41.03% 52.26% -
Mullus barbatus
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 18.53% 16.24% 22.61% 24.28%
FS 2019 18.53% -9.20% 18.88% 30.43%
FS 2020 16.24% 9.20% -19.60% 27.42%
AquaMaps 2019 22.61% 18.88% 19.60% -38.25%
AquaMaps 2050 24.28% 30.43% 27.42% 38.25% -
Sardina pilchardus
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 14.51% 22.84% 16.69% 21.64%
FS 2019 14.51% -29.60% 24.22% 29.17%
FS 2020 22.84% 29.60% -18.32% 20.89%
AquaMaps 2019 16.69% 24.22% 18.32% -32.69%
AquaMaps 2050 21.64% 29.17% 20.89% 32.69% -
Parapenaeus longirostris
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 47.13% 34.34% 21.71% 26.35%
FS 2019 47.13% -20.83% 21.83% 58.85%
FS 2020 34.34% 20.83% -20.70% 41.91%
AquaMaps 2019 21.71% 21.83% 20.70% -47.88%
AquaMaps 2050 26.35% 58.85% 41.91% 47.88% -
Solea solea
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 6.18% 6.75% 11.23% 34.63%
FS 2019 6.18% -2.73% 11.23% 34.63%
FS 2020 6.75% 2.73% -10.85% 34.63%
AquaMaps 2019 11.23% 11.23% 10.85% -52.46%
AquaMaps 2050 34.63% 34.63% 34.63% 52.46% -
Squilla mantis
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 12.07% 4.74% 19.70% 19.20%
FS 2019 12.07% -10.63% 19.70% 17.57%
FS 2020 4.74% 10.63% -19.70% 19.07%
AquaMaps 2019 19.70% 19.70% 19.70% -41.87%
AquaMaps 2050 19.20% 17.57% 19.07% 41.87% -
Engraulis encrasicolus
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - 12.79% 12.07% 21.90% 30.80%
FS 2019 12.79% -4.45% 21.90% 30.80%
FS 2020 12.07% 4.45% -21.90% 30.65%
AquaMaps 2019 21.90% 21.90% 21.90% - 53.00%
AquaMaps 2050 30.80% 30.80% 30.65% 53.00% -
Table 2: Suitability score comparison between the ecological niche models of the eight species involved in our experiment. Model names indicate
floating sensor models for 2015-2018 (FS 2015-2018), 2019 (FS 2019), 2020 (FS 2020), and AquaMaps 2019 and 2050. Scores are reported only
for the FS models to ease the reading. Coloured numbers highlight habitat gain (green), loss (red), or stability (blue) in 2020. Coloured species
names indicate habitat gain (green), change (red), or stability (blue) in 2020 with respect to 2015-2018.
Sepia officinalis
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-0.29%) Loss (-3.95%) Loss Gain
FS 2019 Gain (+0.29%) - Loss(-0.14%) Loss Gain
FS 2020 Gain (+3.95%) Gain (+0.14%) -Loss Gain
AquaMaps 2019 Gain Gain Gain - Gain
AquaMaps 2050 Loss Loss Loss Loss -
Merluccius merluccius
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-7.04%) Loss (-5.68%) Loss Gain
FS 2019 Gain (+7.04%) - Gain(+0.36%) Loss Gain
FS 2020 Gain (+5.68%) Loss (-0.36%) -Loss Gain
AquaMaps 2019 Gain Gain Gain - Stable
AquaMaps 2050 Loss Loss Loss Stable -
Mullus barbatus
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-7.61%) Loss (-3.38%) Loss Gain
FS 2019 Gain (+7.61%) - Gain(+1.94%) Loss Gain
FS 2020 Gain (+3.38%) Loss (-1.94%) -Loss Gain
AquaMaps 2019 Gain Gain Gain - Gain
AquaMaps 2050 Loss Loss Loss Loss -
Sardina pilchardus
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-4.31%) Loss (-4.6%) Loss Gain
FS 2019 Gain (+4.31%) - Gain(+5.46%) Gain Gain
FS 2020 Gain (+4.6%) Loss(-5.46%) -Gain Gain
AquaMaps 2019 Gain Loss Loss - Gain
AquaMaps 2050 Loss Loss Loss Loss -
Parapenaeus longirostris
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-14.87%) Loss (-8.33%) Loss Gain
FS 2019 Gain (+14.87%) - Gain (+7.04%) Loss Gain
FS 2020 Gain (+8.33%) Loss (-7.04%) - Loss Gain
AquaMaps 2019 Gain Gain Gain - Gain
AquaMaps 2050 Loss Loss Loss Loss -
Solea solea
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Stable Loss (-0.5%) Gain Gain
FS 2019 Stable - Stable Gain Gain
FS 2020 Gain (+0.5%) Stable -Gain Gain
AquaMaps 2019 Loss Loss Loss - Gain
AquaMaps 2050 Loss Loss Loss Loss -
Squilla mantis
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-1.22%) Loss (-0.36%) Loss Gain
FS 2019 Gain (+1.22%) - Gain (+0.72%) Loss Gain
FS 2020 Gain (+0.36%) Loss (-0.72%) -Loss Gain
AquaMaps 2019 Gain Gain Gain - Gain
AquaMaps 2050 Loss Loss Loss Loss -
Engraulis encrasicolus
FS 2015-2018 FS 2019 FS 2020 AquaMaps
FS 2015-2018 - Loss (-5.6%) Stable Gain Gain
FS 2019 Gain (+5.6%) - Gain (+1.15%) Stable Gain
FS 2020 Stable Loss (-1.15%) -Gain Gain
AquaMaps 2019 Loss Stable Loss - Stable
AquaMaps 2050 Loss Loss Loss Stable -
Table 3: Median, 1st and 3rd quartiles of the environmental parameter distributions used in our experiment over the Adriatic Sea, estimated from
Argo data. Average aggregation type indicates parameter average over the entire water column.
Parameter name Aggregation
Years Median 1st Quartile 3rd Quartile
Temperature (° C)
2015-2018 14.95 14.94 15.02
2019 14.74 14.74 14.75
2020 15.26 15.25 15.26
2015-2018 14.15 14.14 14.16
2019 14.10 14.09 14.10
2020 14.32 14.31 14.32
2015-2018 16.58 16.50 18.48
2019 19.67 18.51 19.71
2020 18.40 18.35 18.55
Salinity (PSU)
2015-2018 38.83 38.83 38.83
2019 38.90 38.90 38.90
2020 38.97 38.97 38.97
2015-2018 38.82 38.82 38.82
2019 38.86 38.85 38.86
2020 38.90 38.89 38.90
2015-2018 38.78 38.77 38.78
2019 38.80 38.80 38.82
2020 39.01 39.00 39.01
Chlorophyll-a (mgm3)
2015-2018 0.0391 0.0389 0.0392
2019 0.0366 0.0365 0.0377
2020 0.0343 0.0331 0.0344
2015-2018 0.0051 0.0027 0.0052
2019 0.0056 0.0056 0.0057
2020 0.0028 0.0027 0.0029
2015-2018 0.0436 0.0432 0.0438
2019 0.2213 0.2202 0.2222
2020 0.1896 0.1888 0.1907
Dissolved oxygen (µmolkg)
2015-2018 234.12 228.72 234.26
2019 220.50 219.88 220.53
2020 213.70 213.67 213.72
2015-2018 214.32 212.41 214.39
2019 216.81 216.40 216.84
2020 210.33 210.16 210.35
2015-2018 228.36 228.25 228.64
2019 227.80 227.66 227.92
2020 214.73 214.47 214.84
Table 4: Percent contribution and permutation importance of the most habitat-predictive parameters for the 8 analysed species. Bold-highlighted
text indicates, for each species, the major drivers of habitat change from 2015-2018 to 2020. Coloured species names indicate habitat gain
(green), change (red), or stability (blue) in 2020 with respect to 2015-2018.
Species name Parameter Percent con-
tribution (%)
Sepia officinalis
depth 77.6 59.3
average dissolved oxygen 5.4 8.4
average salinity 5.3 21.2
bottom dissolved oxygen 4.9 0
bottom temperature 4.6 0.1
bottom salinity 1.4 5.6
surface chlorophyll-a 0.8 5.5
Merluccius merluccius
bottom temperature 48.2 27.6
average chlorophyll-a 24.6 14.2
depth 7.8 26.4
surface chlorophyll-a 6.4 16.5
average salinity 4.7 5.4
surface dissolved oxygen 3.9 3
surface salinity 2.4 1.5
average temperature 1.9 5.4
Mullus barbatus
bottom temperature 49.5 24.7
average chlorophyll-a 24.7 13.6
surface dissolved oxygen 6.7 6.8
depth 6.5 20.1
bottom chlorophyll-a 5.5 17.4
surface chlorophyll-a 3.1 10.4
bottom dissolved oxygen 2.3 3.8
surface salinity 1.7 3.2
Sardina pilchardus
bottom chlorophyll-a 66.6 54.4
average dissolved oxygen 16.7 0.5
average chlorophyll-a 11.9 20
bottom dissolved oxygen 4.2 0
depth 0.6 25.1
Parapenaeus longirostris
depth 66.2 45
surface temperature 12.9 40.4
average temperature 9.6 14.5
average dissolved oxygen 8.1 0
surface dissolved oxygen 3.2 0.1
Solea solea
depth 80.6 84.9
average temperature 9.7 0
average dissolved oxygen 5.1 0
bottom chlorophyll-a 2.8 9.7
average salinity 1.8 5.4
Squilla mantis
depth 66 77.3
bottom chlorophyll-a 14.4 6.3
average temperature 14.1 16.1
surface temperature 4.1 0.3
bottom salinity 1.5 0
Engraulis encrasicolus
depth 63 31
surface dissolved oxygen 20.1 43.6
bottom chlorophyll-a 5.6 25.4
bottom dissolved oxygen 5.5 0
average chlorophyll-a 3.5 0
average dissolved oxygen 2.4 0
Table 5: Summary of the principal environmental parameters that drove species distribution change in 2020. For each parameter, the table
reports (i) the general (increasing/decreasing) trend with respect to the past years, (ii) the main reasons of the change, (iii-iv) the species whose
distributions were positively affected (i.e. they increased in 2020) or negatively affected by that parameter change.
Principal parameters that
drove selected-species distri-
bution change in 2020
General trend in
2020 wrt past years
Possible reason of the
Species with positively af-
fected distribution by the
Species with negatively af-
fected distribution by the
Temperature Increasing Climate change Sepia officinalis, Merluccius
merluccius, Mullus barbatus
Parapenaeus longirostris
Dissolved Oxygen Decreasing Climate change and pollution Sepia officinalis Sardina pilchardus, Parape-
naeus longirostris
Chlorophyll-a Decreasing COVID-19 pandemic Merluccius merluccius, Mullus
Sardina pilchardus
Figure 1: Distribution of the analysed species’ occurrence records, used for our floating sensor based ecological niche models.
Figure 2: Ecological niches estimated by our floating sensor based (FS) models for 2015-2018, 2019, and 2020, and AquaMaps 2019 and 2050
over the eight analysed species. Coloured species names indicate habitat gain (green), change (red), or stability (blue) in 2020 with respect to
Figure 3: Single-parameter MaxEnt probability densities across the studied species. Only the charts of the key parameters driving habitat gain
and change are reported. Coloured species names in the chart titles indicate those that gained (green) or changed (red) habitat in 2020 with respect
to 2015-2018. Vertical bars highlight the values in 2015-2018 and 2020 at the intersection with medians as dashed lines and quartiles 1 and 3 as
dotted lines. A green horizontal arrow, from a red to a green vertical line, indicates a general habitat suitability increase from 2015-2018 to 2020.
Conversely, a yellow horizontal arrow, from a green to a red vertical line, indicates habitat suitability decrease from 2015-2018 to 2020.
... However, these approaches require huge amounts of high-quality data to produce meaningful knowledge 11,12 . In particular, environmental, geophysical, world-population, and marine-region data are crucial to model species habitats 13,14 , understand the response and resilience of marine areas to climate change [15][16][17] , assess stock status and fisheries pressure on stocks [18][19][20] , and build ecosystem models [21][22][23][24] . ...
... Moreover, sea temperature had a general increasing trend at the global scale (more than linearly for sea-bottom temperature), which several other studies have confirmed in the last decades 56,57 . Net primary production presented a globally increasing trend, in agreement with other studies 58,59 , and an overall decreasing trend in the Adriatic, Aegean, and the Black Sea also highlighted by other studies 14,60,61 . Sea-bottom dissolved oxygen presented a non-linear global-scale decrease in 2020 in all regions, also confirmed by other studies [62][63][64] . ...
... • The global scale had similar trends to those of the Baltic Sea (because of a similar ice concentration trend), Bay of Biscay, and Western Mediterranean because they shared averagely increasing net primary production and temperature trends [56][57][58][59] . Conversely, the global scale had different trends with respect to those of the Adriatic due to different parameter signal-phases; 14 ...
Full-text available
Measurement(s) habitat similarity • Time Series Analysis • Principal Component Technology Type(s) Habitat Representativeness Score • Cross-Correlation • Principal Component Analysis Factor Type(s) Sea-bottom and sea-surface dissolved oxygen, salinity, and temperature • Sea net primary production • Sea ice concentration • Average, minimum, maximum sea depth • Average, minimum, maximum elevation • Distance of a square marine area from land and its fraction covered by water • Indication of the belonging Large Marine Ecosystem (LME), Exclusive Economic Zone (EEZ), Marine Ecoregions of the World (MEOW), Major Ocean Basins, and Marine Protected Area (MPA) • Number of islands • Water area that lies within the shelf, slope, and abyssal zones • Tidal range extension • Coral density • Estuary and seamount presence • Carbon dioxide flux at soil • Air surface temperature • Precipitation • Difference between air surface temperature and sea surface temperature • World population density • Sediment thickness • Atmospheric concentration of methane and nitrous oxide • Earth heat flow • Distance from crust plates • Earthquake density, depth, magnitude • Groundwater resources Sample Characteristic - Environment climate system Sample Characteristic - Location Oceans and Seas • Adriatic Sea • Aegean Sea • Baltic Sea • Bay of Biscay • Black Sea • Mediterranean Sea, Eastern Basin • North Sea • Mediterranean Sea, Western Basin
... Environmental drivers that could influence a population increase in 2020 were scarcely effective for common sole, since its habitat distribution is not sensible enough to the magnitude of alteration of chlorophyll, dissolved oxygen, and temperature that occurred in 2020 compared to the previous years (Coro et al., 2022b). Environmental change was also unlikely to be strong enough to change the habitat distribution of spottail mantis shrimp and anchovy. ...
... However, temperature increase and dissolved oxygen decrease in 2020-which were climate changerelated trends-might have fostered the habitat distribution of common cuttlefish and penalised that of pink shrimp with a resultant influence on biomass. Habitat unsuitability might indeed be one additional reason for the particular condition of deepwater pink shrimp shown in Figure 4. Instead, COVIDspecific environmental changes, like chlorophyll-a decrease, slightly penalised sardine (Coro et al., 2022b). A substantial water-column chlorophyll-a decrease from 2019 to 2020 was indeed measured in the Adriatic and was likely a consequence of the COVID-19 pandemic. ...
... A substantial water-column chlorophyll-a decrease from 2019 to 2020 was indeed measured in the Adriatic and was likely a consequence of the COVID-19 pandemic. This reduction was observed worldwide and potentially corresponded to CO 2 emission dropping in several areas as a result of human activity reduction (Coro et al., 2022b). One logical explanation is that chlorophyll-a is an integral part of the carbon cycle, because this cycle strongly depends on CO 2 consumption during photosynthesis. ...
Full-text available
The COVID-19 pandemic had major impacts on the seafood supply chain, also reducing fishing activity. It is worth asking if the fish stocks in the Mediterranean Sea, which in most cases have been in overfishing conditions for many years, may have benefitted from the reduction in the fishing pressure. The present work is the first attempt to make a quantitative evaluation of the fishing effort reduction due to the COVID-19 pandemic and, consequently, its impact on Mediterranean fish stocks, focusing on Adriatic Sea subareas. Eight commercially exploited target stocks (common sole, common cuttlefish, spottail mantis shrimp, European hake, red mullet, anchovy, sardine, and deepwater pink shrimp) were evaluated with a surplus production model, separately fitting the data for each stock until 2019 and until 2020. Results for the 2019 and 2020 models in terms of biomass and fishing mortality were statistically compared with a bootstrap resampling technique to assess their statistical difference. Most of the stocks showed a small but significant improvement in terms of both biomass at sea and reduction in fishing mortality, except cuttlefish and pink shrimp, which showed a reduction in biomass at sea and an increase in fishing mortality (only for common cuttlefish). After reviewing the potential co-occurrence of environmental and management-related factors, we concluded that only in the case of the common sole can an effective biomass improvement related to the pandemic restrictions be detected, because it is the target of the only fishing fleet whose activity remained far lower than expectations for the entire 2020.