sustainability

Article
Hybrid Intelligence for Marine Biodiversity: Integrating Citizen
Science with AI for Enhanced Intertidal Conservation Efforts at
Cape Santiago, Taiwan
Vincent Y. Chen 1, *, Day-Jye Lu 2 and Yu-San Han 1, *

1 Institute of Fisheries Science, College of Life Science, National Taiwan University, No. 1, Sec. 4, Roosevelt Rd.,
Taipei 106216, Taiwan
2 School of Forestry & Resource Conservation, National Taiwan University, Taipei 106319, Taiwan;
[email protected]
* Correspondence: [email protected] (V.Y.C.); [email protected] (Y.-S.H.); Tel.: +886-3-33663726 (V.Y.C.);
+886-2-33663726 (Y.-S.H.); Fax: +886-3-33669449 (V.Y.C.); +886-2-33669449 (Y.-S.H.)

Abstract: Marine biodiversity underpins the formation of marine protected areas (MPAs), neces-
sitating detailed surveys to account for the dynamic temporal and spatial distribution of species
influenced by tidal patterns and microhabitats. The reef rock intertidal zones adjacent to urban centers,
such as Taiwan’s Cape Santiago, exhibit significant biodiversity, yet they are increasingly threatened
by tourism-related activities. This study introduces an artificial intelligence (AI)-empowered citizen
science (CS) approach within the local community to address these challenges. By integrating CS
with AI, we establish a hybrid intelligence (HI) system that conducts in situ biological surveys and
educational programs focused on reef ecological conservation. This initiative not only facilitates the
collective gathering and AI-assisted analysis of critical data but also uses machine-learning outputs to
gauge data quality, thus informing subsequent data collection and refinement strategies. The resulting
collectivity and iterative enhancement foster a mutual and continuous HI learning environment. Our
HI model proves instrumental in fostering community engagement and public involvement in CS
endeavors, cultivating the skills necessary for documenting rocky intertidal biodiversity shifts. These
efforts are pivotal for informing the design and governance of future MPAs, ensuring their efficacy
Citation: Chen, V.Y.; Lu, D.-J.; Han,
and sustainability in marine conservation.
Y.-S. Hybrid Intelligence for Marine
Biodiversity: Integrating Citizen
Keywords: marine protect areas; rocky intertidal ecosystem; hybrid intelligence; citizen science;
Science with AI for Enhanced
artificial intelligence
Intertidal Conservation Efforts at
Cape Santiago, Taiwan. Sustainability
2024, 16, 454. https://doi.org/
10.3390/su16010454
1. Introduction
Academic Editor: Tim Gray
Coastal regions, including the marine littoral zone, are biological powerhouses [1].
Received: 13 November 2023 They offer refuge to an expansive array of marine life and directly benefit humanity in
Revised: 18 December 2023 multifarious ways [2]. From food sources and economic avenues to recreational pursuits
Accepted: 30 December 2023 and the conservation of cultural heritages, these ecosystems are indispensable [3]. In
Published: 4 January 2024 parallel, rocky intertidal systems, especially those adjacent to urban locales, are paramount
to sustaining ecological balances while also being focal points for economic, social, and
recreational engagements [4].
However, the allure of these regions is both their strength and vulnerability. The
Copyright: © 2024 by the authors.
recent upsurge in coastal tourism and its resultant economic windfall for local regions [5,6]
Licensee MDPI, Basel, Switzerland.
draw legions of tourists. Urban inhabitants, especially, are enchanted by the serene coastal
This article is an open access article
landscapes [7] and the prospect of close encounters with marine fauna. Yet, this very
distributed under the terms and
conditions of the Creative Commons
attraction coupled with activities like marine organism harvesting and heedless trampling
Attribution (CC BY) license (https://
cast shadows of environmental stress on these habitats [8–10]. The rich biodiversity, a
creativecommons.org/licenses/by/ hallmark of these zones, faces relentless challenges. Rocky intertidal areas, rife with
4.0/). microhabitats, stand as evidence of this. The seemingly benign acts of visitors, such as

Sustainability 2024, 16, 454. https://doi.org/10.3390/su16010454 https://www.mdpi.com/journal/sustainability

Sustainability 2024, 16, 454 2 of 20

walking over tide pools or curiously overturning rocks, inflict lasting ecological scars. The
cascading effects of such disturbances manifest in dwindling algal populations, impacting
both fleshy and coralline types [11], and in the perceptible strain on marine entities, like
echinoderms [12], mollusks [13,14], and crustaceans [15].
The Anthropocene epoch, marked by significant human impact on the Earth’s biophys-
ical systems, has led to intensified pressures on coastal regions due to population growth,
urbanization, and climate change [16]. This has notably affected rocky intertidal ecosys-
tems with human interaction evolving from fishery exploitation to recreational activities
like adventure tourism [17]. A multifaceted coastal management approach, incorporating
diverse strategies from legal to educational, is essential to mitigate environmental impacts
and encourage societal participation [18]. Additionally, the largely unexplored marine
domain presents opportunities for marine citizen science (MCS). MCS plays a crucial role
in filling research gaps and enhancing global marine conservation efforts [19,20]. It pro-
vides cost-effective, robust data that inform policy decisions [21,22] while increasing public
science literacy and community engagement in marine issues [23–25] from cetacean con-
servation [26,27] to addressing plastic pollution [28], thereby enabling community-driven
initiatives to transform research into effective policies [29,30].
At the heart of this movement is the citizen science (CS) research methodology that
enlists the general public in data collection, categorization, or scientific analysis [31]. Histor-
ically, it has been instrumental in various marine research undertakings. Notable projects
include monitoring reef fish ecosystems [32], tracking queen conch species populations [33],
discerning seagrass bed dynamics [34], and mapping the distribution patterns of marine
litter [35,36]. The unique advantage of CS in marine projects lies in its capacity to amplify
the spatial and temporal scope of studies [37,38].
Yet, the integration of projects/research with CS is not without challenges. Concerns
often center on the data quality when gathered by non-professionals, especially for intricate
biodiversity datasets, which could potentially impede their application [39]. Nevertheless,
recent innovations offer solutions. As highlighted by Earp et al. [40], implementing strin-
gent protocols, ensuring thorough training, and adhering to meticulous data verification
processes can elevate the data’s quality. With such measures in place, data sourced from
citizen scientists can rival if not match those procured by their professional counterparts,
as evidenced in their intertidal algae ecology experiment.
AI has been employed to augment and enhance human understanding of the environ-
ment, including perceptions of citizen scientists [41]. While CS and AI are often viewed
as separate tools for ecological monitoring, recent studies indicate that a symbiotic rela-
tionship between human intelligence and AI, termed hybrid intelligence (HI) [42,43], can
strategically unite the two, enhancing outcomes for conservation activities. By pairing the
public engagement benefits of CS projects with the sophisticated analytical prowess of AI,
there is the potential to foster greater multi-stakeholder consensus on matters of public [44]
and scientific importance. Moreover, the integration of both methodologies can expedite
data collection and processing relative to traditional scientific approaches, indicating a
promising avenue for accelerated monitoring and conservation efforts [45].
Cape Santiago, located on Taiwan Island’s easternmost point near Taipei, is at the
intersection of the Kuroshio and longshore currents. Named by Spanish explorers 400 years
ago, its 4.86 km coastline features wave-cut benches and rocky shores and is home to the
small Magang fishery harbor community. With declining fishery resources driving youth
to the cities, the local government, aiming to boost tourism, built a bicycle route to the
coast in 2011. However, the influx of tourists engaging in harmful activities like capturing
marine specimens has raised concerns about the impact on biodiversity and the coastal
environment. The Cape Santiago Culture Development Association (SDCDA) was formed
to protect the fishery village culture and marine ecosystem. Legally, Cape Santiago’s coast
is an “Ocean Resource Protected Area” under the Urban Planning Law [46], restricting
construction but not specifically addressing environmental harm.
Sustainability 2024, 16, 454 3 of 20

Acknowledging the existing shortcomings in marine ecological conservation, the

Ocean Affairs Council introduced the Marine Conservation Act proposal in 2019 [47]. Draft
article 7 of this act empowers management authorities to designate coastal regions with
significant biodiversity that require special protection as marine sanctuaries. They are also
mandated to implement management measures such as ecological monitoring, as detailed
in draft article 6. Thus, the SDCDA is advocating for the wave-cut bench on the western
side of the fishery harbor to be recognized as a strictly regulated marine protected area.
This designation would facilitate the management of environmentally detrimental actions
under the Marine Conservation Act. The scientific assessment of biodiversity along this
coast will be a crucial factor in determining if the area qualifies as a marine protected
zone. The community members are unified in their stance and are mobilizing volunteers
for coastal biodiversity CS surveys. Before furnishing the necessary scientific data to
legislative bodies for the inclusion of this area as a marine protected zone, it is imperative to
proactively educate tourists on the coastal ecosystem. This initiative aims to teach visitors
the importance of marine conservation, discouraging activities that can harm marine life
and the environment.
This study aims to assist the SDCDA in employing a CS project and AI for the conser-
vation of Cape Santiago’s intertidal biodiversity. This study has set the following objectives:
1. Create a spatiotemporal biodiversity database for the benches of Cape Santiago;
2. Assess the performance of the AI model, the data contributions from CS, and the
overall impact of the collaboration;
3. Formulate an HI framework that merges CS and AI, targeting continuous monitoring
and environmental education.

2. Methods
2.1. Study Area and the Citizen Scientist Project
The study area of this paper is located on the northwest side of the Cape Santiago
coastline, specifically focusing on a wave-cut bench, termed the Santiago bench. This bench
spans approximately 160 m in length and 70 m in width (Figure 1). The primary aim of the
survey is to identify and document target marine organisms within this predefined area.
To evaluate the potential marine protected areas (MPAs), the Ocean Consecration
Administration (OCA) conducted biodiversity assessments at 67 rocky shores spanning
Taiwan’s main island and its surrounding islands [48]. They utilized transect line and
quadrat methodologies to assess marine species richness, biomass, and a range of envi-
ronmental factors, including the impacts of sewage outfalls, man-made facilities, tourist
activity, siltation levels, and the abundance of rock pools. By integrating both biodiversity
and environmental indicators, the OCA determined the conservation priorities of these
67 rocky shores. Among the surveyed locations were four sites at Cape Santiago, including
Santiago bench, Lai-Lai-1, Lai-Lai-2, and Lai-Lai-3. Notably, the Lai-Lai-2 site, located at
the southeastern edge of Cape Santiago, showcased the greatest marine species diversity.
Along with the neighboring Lai-Lai-3 site, both were identified as having the highest
conservation value, as reported in the OCA’s pilot survey [48].
The participants in the CS project were recruited by the SDCDA and underwent a
comprehensive 16 h training session. This training workshop, conducted by the authors
in collaboration with marine experts, encompassed topics such as rocky shore ecology,
macrobenthos species identification (Table 1), AI training procedures, survey assistance
methods, and environmental education. To enhance their training, materials from a prior
survey compiled from the author’s dataset [49] that identified 13 phyla and cataloged
234 marine species, including a new marine flatworm species [50], were provided, com-
plemented by images of each species previously documented in the region. The main
responsibilities assigned to the CS participants were twofold: data collection through field
surveys and public education regarding the intertidal environment.
Sustainability 2024, 16, x FOR PEER REVIEW 4 of 20

Sustainability 2024, 16, 454 responsibilities assigned to the CS participants were twofold: data collection through 4field
of 20
surveys and public education regarding the intertidal environment.

Figure 1. Location of the study site at Cape Santiago with its position within Taiwan highlighted
Figure 1. Location of the study site at Cape Santiago with its position within Taiwan highlighted
(inset) and the three survey lines conducted by the OCA.
(inset) and the three survey lines conducted by the OCA.

Table 1. The structure of the macrobenthos in the study area of marine species of Santiago bench.
Table 1. The structure of the macrobenthos in the study area of marine species of Santiago bench.
Kingdom
Kingdom Phylum Phylum Common Common NamesNames Species Numbers
Species Numbers
Porifera
Porifera Sponges
Sponges 88
Bryozoans, Moss
BryozoaBryozoa Bryozoans, Moss Animals 4
Animals
Cowries, Chitons, Sea Slugs,
Cowries, Chitons, Sea Oc-
Mollusca
Mollusca 72
72
topuses, Oysters Oysters
Slugs, Octopuses,
Animalia
Arthropoda Crabs, Shrimps,
Arthropoda Crabs, Shrimps, Barnacles 22
22
Barnacles
Annelida Segmented Worms 8
Annelida Segmented Worms 8
Corals, Sea Anemones,
Corals, Sea Jelly-
Anemones,
Animalia Coelenterate
Coelenterate 10
fishes
Jellyfishes
Echinoderms Sea Stars,Sea
SeaStars, Sea Urchins,
Urchins, Sea Cu- 27
Echinoderms Sea Cucumbers 27
Nemertina
cumbers
Ribbon Worms 2
Nemertina
Platyhelminthes RibbonFlatworms
Worms 28
Platyhelminthes Tunicates,
Flatworms Vertebrates, 8
Chordata 39
Fishes, Sea Snakes
Tunicates, Vertebrates, Fishes,
Chordata
Chlorophyta
39
Sea Green
Snakes Algae 9
Plantae Rhodophyta Red Algae 19
Chlorophyta Green Algae 9
Plantae Chromophyta Brown Algae 6
Rhodophyta Red Algae 19

The survey protocol, derived from the Ocean Consecration Administration’s (OCA)
transect line method, required two individuals to methodically search and document
marine species along a designated transect line, spanning from the high tide zone to the
Sustainability 2024, 16, 454 5 of 20

low tide area of the shore. The survey procedure was conducted by volunteers who had
previously received a CS training workshop. During the survey, the participants followed a
pre-set route marked by the transect line, photographing all marine specimens encountered
and recording the quantities of specific target species on survey forms. Following the
survey, the collected data and images were systematically compiled and uploaded to create
a comprehensive training dataset for future artificial intelligence (AI) training sessions.

2.2. Training AI Model and AI Performance Evaluations

The convolution neural network (CNN) is a cornerstone in AI models for image pattern
extraction and recognition [51]. R-CNN and YOLO, both built on the CNN framework,
can distinguish various objects within an image [52,53]. The YOLOv5s [54] model was
chosen for this study because of its high calculating speed and accuracy in detecting marine
life in natural settings. YOLOv5s, being a deep-learning approach in machine vision,
requires comprehensive training on an extensive image dataset to perform optimally [55,56].
The images of organisms from the participants’ field survey dataset were meticulously
identified and annotated as ‘instances’ using the open-source Python program labelImage
(https://github.com/HumanSignal/labelImg, accessed on 10 July 2022). Each instance
was marked as a bounding box with its location, size, and common name within the image.
Effective AI recognition of a particular specimen necessitates a substantial image database
for that specimen, typically comprising 200 or more instances, as per our experience.
Consequently, species that are commonly observed at the research site are usually selected
for AI training. To keep participants engaged, we also include a selection of rare target
species, like sea hares and blue-ringed octopuses, encouraging participants to search for
and document these less frequently encountered organisms. These annotated images were
then compiled into a training dataset for YOLOv5s. All datasets were randomly split
with 80% allocated for training and 20% for validation. Training was conducted on Linux
desktop computers equipped with an RTX3090-24G GPU, using hyperparameters that
encompassed a pre-trained weight, a training batch size of 32, an image resolution of
800 × 800, and a learning epoch of 200.
Two primary tasks need to be executed by a trained AI model. The first involves
identifying target species instances from an image, while the second requires the accurate
prediction of the specific species’ common name, distinguishing it from other species.
The AI model’s predictions for any given specimen can be categorized into four possible
outcomes:
4. Correctly identifying the actual species—true positive (TP).
5. Accurately recognizing a species other than the target—true negative (TN).
6. Missing the identification of the actual species—false negative (FN).
7. Incorrectly labeling a different species as the target species—false positive (FP).
The outcomes of distinguishing between various species can be depicted in a two-
dimensional table, commonly referred to as a confusion matrix. In this matrix, the actual
species are represented on the x-axis, while the predicted species appear on the y-axis
(refer to Figure 2). The overall model’s performance was assessed using Equations (1), (2),
and (4):
TP
precision = (1)
TP + FP
TP
recall = (2)
TP + FN
To calculate the ratio of the area in the predicted instance bounding box Bp overlapping
the area of the actual ground truth box Bg , the intersection of union, IoU, formula is used
as follows:
aera | Bp ∩ Bgt|
IOU = (3)
aera | Bp ∪ Bgt|
 𝑟𝑒𝑐𝑎𝑙𝑙 = (2)
𝑇𝑃 + 𝐹𝑁
To calculate the ratio of the area in the predicted instance bounding box Bp overlap-
ping the area of the actual ground truth box Bg, the intersection of union, IoU, formula is
Sustainability 2024, 16, 454 used as follows:
6 of 20
𝑎𝑒𝑟𝑎 |𝐵𝑝 ∩ 𝐵𝑔𝑡|
𝐼𝑂𝑈 = (3)
𝑎𝑒𝑟𝑎 |𝐵𝑝 ∪ 𝐵𝑔𝑡|
The mAP Equation (3) measures the averaged precision value of all the cases when
The mAP0.50.5
Equation (3) measures the averaged precision value of all the cases when
the IOU is 50% threshold area overlap.
the IOU is 50% threshold area overlap.
1 n
n ∑ i =1
mAP = ∑
𝑚𝐴𝑃0.5. = 𝐴𝑃APcategory i 𝑤ℎ𝑒𝑛
when𝐼𝑜𝑈_𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
IoU_threshold==0.5
0.5 (4)(4)

Figure 2. An example of a 3 × 3 multi-class confusion matrix with actual species as x-axis/row grids
Figure 2. An example of a 3 × 3 multi-class confusion matrix with actual species as x-axis/row grids
and predicted species as y-axis/column grids. While focusing on class A, the AA grid is true positive,
and predicted species as y-axis/column grids. While focusing on class A, the AA grid is true positive,
TP, therest
TP, the restof
of the
the row
row grids
grids are
are false
false positive,
positive, FP,
FP,the
therest
restofofthe
thecolumn
columngrids areare
grids false negative,
false FN,
negative,
and the rest of the grids except all above grids are all true negative,
FN, and the rest of the grids except all above grids are all true negative, TN. TN.

2.3. Hybrid Intelligence Criteria for CS and AI Synergy

2.3. Hybrid Intelligence Criteria for CS and AI Synergy
To investigate the potential synergy between CS and AI in forming HI, we adopted the
To investigate the potential synergy between CS and AI in forming HI, we adopted
operational definition of hybrid intelligence (HI) from [57], which is based on three criteria:
the operational definition of hybrid intelligence (HI) from [57], which is based on three
8. Collectiveness: This emphasizes the collaboration between humans and AI with the
criteria:
aim of collectively
8. Collectiveness: Thisaddressing
emphasizes a task to achieve a system-level
the collaboration goal. Itand
between humans is acknowledged
AI with the
that individual agents might have sub-goals that deviate from the overarching
aim of collectively addressing a task to achieve a system-level goal. It is acknowl- system
objective.
edged that individual agents might have sub-goals that deviate from the overarching
9. Solution Superiority: This asserts that the combined sociotechnical system, which
system objective.
includes both humans and AI, produces results surpassing what individual agents,
9. Solution Superiority: This asserts that the combined sociotechnical system, which
whether human or AI, could achieve independently.
includes both humans and AI, produces results surpassing what individual agents,
10. Mutual and Continuous Learning: The system shows consistent improvement, both
whether human or AI, could achieve independently.
collectively and at the individual component level (human and AI). This continuous
enhancement signifies persistent learning and development from both parties.

3. Results
3.1. Survey Protocols
Initially, the survey protocol was adapted from the OCA’s methodology [48]. This
approach involved setting up 10 transect lines on the Santiago bench within the determined
survey zone. The survey groups, each comprising two members, aligned with these lines.
Therefore, a total of 10 groups worked in tandem, collecting data and photographs for AI
 3.1. Survey Protocols
Initially, the survey protocol was adapted from the OCA’s methodology [48]. This
approach involved setting up 10 transect lines on the Santiago bench within the deter-
Sustainability 2024, 16, 454 mined survey zone. The survey groups, each comprising two members, aligned with these 7 of 20
lines. Therefore, a total of 10 groups worked in tandem, collecting data and photographs
for AI training. The primary objective was to document the time and spatial occurrence of
marine
training.species. Upon implementation,
The primary some challenges
objective was to document the timewith
andthe original
spatial proposal
occurrence survey
of marine
protocol (PSP) (Figure 3a) became apparent. Some species, like the sea
species. Upon implementation, some challenges with the original proposal survey protocolhare, were absent
during the summer
(PSP) (Figure surveys.
3a) became Considering
apparent. Somethe route’s
species, dual
like thefunction
sea hare,aswere
a prospective edu-
absent during
cational
the summerpathway, a looped
surveys. configuration
Considering was determined
the route’s dual function to beas more suitable than
a prospective a linear
educational
design.
pathway, This feedback
a looped led to the development
configuration was determined of an adaptive
to be survey protocol
more suitable (ASP).
than a linear The
design.
pivotal changeled
This feedback in the adaptive
to the survey protocol
development involved
of an adaptive the survey
survey route.
protocol Instead
(ASP). The of trav-
pivotal
ersing
changedirectly from thesurvey
in the adaptive shore to the sea,involved
protocol the routethenow followed
survey grooves
route. Instead(GRV) parallel
of traversing
to the coastline.
directly from theThese
shoregrooves acted
to the sea, the as paths
route now that led to four
followed tidal(GRV)
grooves pools parallel
(TP) andtofour
the
grooves
coastline.(GRV).
These This revised
grooves actedprotocol
as pathsoffered
that ledflexibility in pools
to four tidal group(TP)sizes,
and and thegrooves
four teams
could
(GRV).nowThisconduct
revised their surveys
protocol asynchronously,
offered adhering
flexibility in group sizes,toand
the the
predetermined
teams couldroutenow
(Figure
conduct3b).
their surveys asynchronously, adhering to the predetermined route (Figure 3b).

Figure 3. The survey transect routes of different survey protocols. (a) The routes of ten transect lines
Figure 3. The survey transect routes of different survey protocols. (a) The routes of ten transect lines
(TSL) of the proposed survey protocol (PSP), and (b) a loop comprised of 4 grooves (GRV) and 4
(TSL) of the proposed survey protocol (PSP), and (b) a loop comprised of 4 grooves (GRV) and 4 rock
rock pools (RP) of the adaptive survey protocol (ASP).
pools (RP) of the adaptive survey protocol (ASP).
The survey data are categorized based on both temporal and spatial factors. Tem-
The survey data are categorized based on both temporal and spatial factors. Temporal
poral data capture the season during which the survey was conducted, categorized into
data capture the season during which the survey was conducted, categorized into spring
spring (SP), summer (SU), autumn (F), and winter (W). The surveys were uniquely timed
(SP), summer (SU), autumn (F), and winter (W). The surveys were uniquely timed to
to coincide with low tide moments, allowing for the CS participants to access the exposed
coincide with low tide moments, allowing for the CS participants to access the exposed
bench. Spatial data are differentiated based on tide levels: high tide level (HTL) and low
bench. Spatial data are differentiated based on tide levels: high tide level (HTL) and low
tide
tide level
level (LTL).
(LTL). As
As the
the tide
tide recedes
recedes and
and the Santiago bench
the Santiago bench isis unveiled,
unveiled, areas retaining
areas retaining
water in depressions are labeled as LTL, while areas showcasing organisms above
water in depressions are labeled as LTL, while areas showcasing organisms above the the wa-
water
ter surface are designated as high water level (HWL). These designations aid in recording
surface are designated as high water level (HWL). These designations aid in recording the
the spatial
spatial distribution
distribution of marine
of marine organisms
organisms (Figures
(Figures 4a and
4a and 5a). 5a).

3.2. Proposal Survey Protocol and AI Training

From the proposed survey protocol, we garnered a collection of 1301 photographs
that showcased nine distinct marine species, complete with their spatial and tempo-
ral distribution. The total annotated instances across all specimens amounted to 2643
(Figure 4a). Utilizing this dataset, we trained the PSP YOLOv5s neural network, achieving
a performance marked by a precision of 0.96175, recall of 0.93533, and mAP_0.5 of 0.95485.
An examination of the confusion matrix indicated high true positive (TP) values for almost
all species. Notably, barnacles and oysters registered TP values of 0.83 and 0.87, respectively.
All other species achieved impressive TP values exceeding 0.97 with some even reaching a
perfect score of 1.0 (100%) (Figure 4b).
 Sustainability 2024,
Sustainability 2024, 16,
16, xx FOR
FOR PEER
PEER REVIEW
REVIEW 88 of
of 20
20
Sustainability 2024, 16, 454 8 of 20

Figure
Figure 4.
Figure 4. (a)
4. (a) Displays
(a) Displays the
Displays the species
the species common
species common name,
common name, scientific
name, scientific name,
scientific name, spatial
name, spatial distribution
spatial distribution (high
distribution (high water
(high water
water
level
level (HWL)
(HWL) or
or low
low water
water level
level (LWL)),
(LWL)), temporal
temporal distribution
distribution (spring
(spring (SP),
(SP),
level (HWL) or low water level (LWL)), temporal distribution (spring (SP), summer (SU), autumn summer
summer (SU),
(SU), autumn
autumn
(F),winter
(F),
(F), winter(W)),
winter (W)),and
(W)), andinstance
and instancenumber
instance numberof
number of999 species
of speciescollected
species collected and
collected and used
and used for
used for training
for training YOLOv3
training YOLOv3 following
YOLOv3 following
following
adaptive
adaptive protocol
protocol surveys
surveys conducted
conducted bybyCS CS participants.
participants. (b)
(b) Shows
Shows thethe confusion
confusion
adaptive protocol surveys conducted by CS participants. (b) Shows the confusion matrix true matrix
matrix truetrue posi-
positive
posi-
tive
(TP) (TP)
tive (TP) results
results
results for different
for different
for different species
species
species classes.
classes.
classes.

Figure 5. Cont.
Sustainability 2024, 16, x FOR PEER REVIEW 9 of 20
Sustainability 2024, 16, 454 9 of 20

Figure 5. (a) Displays the species common name, scientific name, spatial distribution (high water
Figure 5. (a) Displays the species common name, scientific name, spatial distribution (high water
level (HWL) or low water level (LWL)), temporal distribution (spring (SP), summer (SU), autumn (F),
level (HWL) or low water level (LWL)), temporal distribution (spring (SP), summer (SU), autumn
winter (W)),
(F), winter and
(W)), andinstance number
instance numberofof2727species
speciescollected
collectedand
andused
usedfor
fortraining
trainingYOLOv3
YOLOv3following
follow-
adaptive protocol surveys conducted by CS participants. (b) Shows the confusion
ing adaptive protocol surveys conducted by CS participants. (b) Shows the confusion matrixmatrix true positive
true
(TP) results
positive for different
(TP) results speciesspecies
for different classes.classes.
The darker color shows
The darker color the higher
shows TP accuracy.
the higher TP accuracy.

3.3. Adaptive Survey Protocol and AI Training

3.2. Proposal Survey Protocol and AI Training
Based on the integration of the ASP and the earlier PSP, we obtained a comprehensive
From the proposed survey protocol, we garnered a collection of 1301 photographs
collection of 5461 photographs that depict 27 distinct marine species. This dataset not only
that showcased nine distinct marine species, complete with their spatial and temporal dis-
showcases the spatial and temporal distribution of these species but also includes a total
tribution. The total annotated instances across all specimens amounted to 2643 (Figure
of 7729 annotated instances (Figure 5a). Leveraging this enriched dataset, we trained the
4a). Utilizing this dataset, we trained the PSP YOLOv5s neural network, achieving a per-
ASP YOLOv5s model, which yielded a performance evaluation with a precision of 0.87412,
formance marked by a precision of 0.96175, recall of 0.93533, and mAP_0.5 of 0.95485. An
recall of 0.85843, and mAP_0.5: of 0.8707. A deeper dive into the confusion matrix, post
examination of the confusion matrix indicated high true positive (TP) values for almost
enhancement of instance and species counts, revealed that the true positive (TP) values
allfor
species. Notably, barnacles
species like red coral algaeand oysterssp.)
(Corallina registered TP values
and tunicates wereofthe
0.83 and 0.87,
lowest, respec-at
registering
tively. All other species achieved impressive TP values exceeding 0.97 with
0.56 and 0.57, respectively. Notably, most of the species TP values from the PSP remained some even
reaching
unchanged.a perfect
Thescore of 1.0
earlier low(100%) (Figure
TP values 4b).
observed in the PSP dataset for barnacles and
oysters showed slight improvements, rising from 0.83 to 0.84 and 0.87 to 0.90, respectively
3.3. Adaptive
(Figure 5b).Survey Protocol and AI Training
Based on the integration of the ASP and the earlier PSP, we obtained a comprehensive
3.4. Deployment
collection of 5461 of AI for In Situ
photographs Identification
that depict 27 distinct marine species. This dataset not only
showcases the spatialprogram
An application and temporal distribution
interface (API) was of these species
developed usingbut
thealso includes
Python a total
programming
oflanguage
7729 annotated instances (Figure 5a). Leveraging this enriched dataset, we
to facilitate a web-based interface (Figure 6). This interface hosts the trainedtrained theAI
ASP
model YOLOv5s
dedicated model, whichspecimen
to marine yielded recognition
a performance and evaluation
is stationedwith a precision
on a cloud serveroffor
0.87412,
efficientrecall of 0.85843,
specimen and mAP_0.5:
identification tasks.ofCrafted
0.8707. A deeper
with dive into
the intent the the
to aid confusion matrix,
CS participants
 3.4. Deployment of AI for In Situ Identification
An application program interface (API) was developed using the Python program-
Sustainability 2024, 16, 454 ming language to facilitate a web-based interface (Figure 6). This interface hosts the
10 of 20
trained AI model dedicated to marine specimen recognition and is stationed on a cloud
server for efficient specimen identification tasks. Crafted with the intent to aid the CS par-
ticipants during surveys or environmental educational endeavors, this tool proves inval-
during surveys or environmental educational endeavors, this tool proves invaluable in
uable in real-time
real-time specimenspecimen identification.
identification. On encountering
On encountering marine organisms
marine organisms in their in their
natural
natural habitat, users can snap photos using their mobile devices and promptly
habitat, users can snap photos using their mobile devices and promptly upload them for upload
them for
instant instant AI-based
AI-based identification.
identification. Upon receiving
Upon receiving the image,the
theimage, the API communicates
API communicates with the
AI model to carry out the recognition process, subsequently relaying therelaying
with the AI model to carry out the recognition process, subsequently the results
results back to the
back (Figure
user to the user (Figure
7). This 7). This streamlined
streamlined process
process greatly greatly
enhances theenhances
efficiencythe efficiency
and accuracy andof
accuracy of specimen surveys and educational
specimen surveys and educational activities. activities.

Figure 6. The program interface API incorporates an AI recognition service. Users can utilize the
provided AI recognition capabilities to search for organism in the field. When they encounter a
suspected specimen (pictures listed below as a guide), they can capture a photo and upload it from
their mobile phones to the server to request AI identification.
Sustainability 2024, 16, x FOR PEER REVIEW 11 of 20

Figure 6. The program interface API incorporates an AI recognition service. Users can utilize the
Sustainability 2024, 16, 454 provided AI recognition capabilities to search for organism in the field. When they encounter11a of
sus-
20
pected specimen (pictures listed below as a guide), they can capture a photo and upload it from
their mobile phones to the server to request AI identification.

Figure 7. Examples of AI-identified specimen instances: (a) 4 blotched-nrites, 1 cowrie, and 1 rocky
Figure 7. Examples of AI-identified specimen instances: (a) 4 blotched-nrites, 1 cowrie, and 1 rocky
oyster, (b) 1 eye spot sea hare, (c) 2 flat rock crabs, (d) 2 sea lettuce, (e) 10 barnacle shells, (f) 1 blue-
oyster, octopus.
ringed (b) 1 eye spot sea hare, (c) 2 flat rock crabs, (d) 2 sea lettuce, (e) 10 barnacle shells, (f) 1 blue-
ringed octopus.

4. Discussions
4.1. Adaptive Survey Protocol for Rocky Intertidal Biodiversity
The Ocean Conservation Administration’s proactive survey [48] initiative aligns with
the prescriptive Marine Conservation Act draft, aiming to foster a deeper understanding
of Taiwan’s extensive coastline, spanning approximately 1600 km, including the rocky
intertidal ecosystems. Among the 67 identified survey sites, each was meticulously assessed
Sustainability 2024, 16, 454 12 of 20

for environmental and biological diversity. This assessment entailed a 60 m transect-line

study from the high tide to the low tide mark, where researchers diligently documented
observable marine species within a 60 min timeframe. Complementary to this, quadrat
photographs were used to further analyze species presence and abundance.
However, the photographic method within the quadrat has its limitations, primar-
ily disturbing mobile species, which may lead to an underrepresentation of non-sessile
organisms. This method also struggles with capturing seasonal biodiversity variations,
as demonstrated with the surveys conducted in early July at Cape Santiago. Seasonal
dynamics, such as the absence of green algae due to summer conditions, could significantly
skew the diversity index, omitting critical faunal dependencies.
The geological complexity of Cape Santiago’s intertidal zones with its pits, grooves,
crevices, and rock pools inherently impacts marine biodiversity. The intricate structures
provide a multitude of microhabitats that support a rich array of life [58]. Davidson
et al. [59] found that the diversity in mobile taxa can be twice as high as that in sessile
taxa with sessile organisms presenting far greater abundance. This indicates that Cape
Santiago’s biodiversity is likely to exhibit considerable spatial and temporal variations that
are influenced by monsoons, tides, water temperature, and sunlight exposure.
Addressing the spatial–temporal complexities and survey feedback from the CS par-
ticipants, the adaptive survey protocol (ASP) was designed to overcome the limitations of
the OCA transect-line method. It aims to prevent trampling damage to microhabitats by
modifying survey routes to a single track along bench crevices (Figure 3), allowing for the
CS participants to survey at their convenience, including nighttime. This flexible approach
expands the spatial–temporal scope of data collection, which is essential for monitoring
biodiversity changes.

4.2. Synergic Learning between CS and AI

In this study, a collaborative learning system that integrates CS and AI has been
established. This system enables a symbiotic learning relationship where AI benefits from
human-generated data and humans refine their knowledge through AI feedback, promoting
mutual learning over extended periods. The flow of information and learning is illustrated
in Figure 8. The CS participants, through workshops focusing on specimen identification
and AI training, improve their skills and are, thus, able to contribute more effectively. They
collectively gather extensive image datasets, 1301 images from PSP and 5461 images from
ASP (shown in Figure 8c), which are crucial for effective PSP and ASP YOLOv5s model
training. The outcomes of the training then provide insights, 0.96175 precision for PSP
and 0.87412 precision for ASP, into the collective data quality of different survey protocols
(Figure 8d).
Utilizing an online AI tool, the CS participants enhance the AI’s identification capa-
bilities through real-time data provision. This iterative process may refine the YOLOv5s
model, which is our offline solution algorithm (offSA) that, after rigorous training using
images from CS activities, is now operational on a server (Figure 8e). A user-friendly
interface (Figure 6) has been established to streamline access to the online solution algo-
rithm (onSA), serving both the CS participants (CSPs) and tourists engaged in educational
programs (Figure 8a). The interface offers instant identification results (Figure 8b) and is
integral to data collection for continuous AI training. This strategic implementation forms
a dynamic feedback loop with data from queries contributing to successive training cycles,
progressively enhancing the AI model’s accuracy (Figure 8a,f).
Sustainability 2024, 16, x FOR PEER REVIEW 13 of 20
Sustainability 2024, 16, 454 13 of 20

information flow diagram (IFD) illustrating “Tier 3: hybrid intelligence (HI) at the
Figure 8. The information
collective level” adapted from Ref. [43]. Please refer to the accompanying text for details.

A crucial aspect of citizen science (CS) projects is maintaining data quality, a task
complicated by the varied experience levels of the contributors [60–62]. This study shifts
from individual to collective
collective intelligence evaluations using AI model performance as a
proxy
proxy for
fordata
dataquality
qualityassessment.
assessment. Participants
Participantsengaged
engagedin PSP maymay
in PSP question whether
question their
whether
dataset of 1301 images with 2643 annotations sufficiently trains the
their dataset of 1301 images with 2643 annotations sufficiently trains the PSP YOLOv5s PSP YOLOv5s model.
The AI’sThe
model. performance, precisionprecision
AI’s performance, of 0.96175,of recall
0.96175,of 0.93533,
recall ofand meanand
0.93533, average
meanprecision
average
(mAP_0.5) of 0.95485, suggests robust training efficacy. Similarly, those
precision (mAP_0.5) of 0.95485, suggests robust training efficacy. Similarly, those conduct-conducting ASP
with 5461 images and 7729 annotations must consider if their dataset
ing ASP with 5461 images and 7729 annotations must consider if their dataset is adequate.is adequate. The
ASP model’s
The ASP results,
model’s precision
results, of 0.87412,
precision recallrecall
of 0.87412, of 0.85843, and mAP_0.5
of 0.85843, and mAP_0.5 of 0.8707, pose
of 0.8707,
questions about disparities in performance. An inquiry arises: What factors
pose questions about disparities in performance. An inquiry arises: What factors contrib- contribute to
these differences, and can the ASP model achieve the PSP model’s level
ute to these differences, and can the ASP model achieve the PSP model’s level of accuracy?of accuracy?
Through confusionmatrix
Through confusion matrixanalysis,
analysis,thethe research
research evaluates
evaluates thethe dataset
dataset quality
quality for
for dif-
different models. Table 2, derived from the confusion matrices (Figures
ferent models. Table 2, derived from the confusion matrices (Figures 4 and 5), details in- 4 and 5), details
instances
stances and andtheir
theircorresponding
correspondingtrue truepositives
positives(TP)
(TP)recognized
recognized by by the
the models.
models. Certain
Certain
species,
species, such as the red coral algae (Corallina sp.) and tunicates, exhibit TP values thatthat
such as the red coral algae (Corallina sp.) and tunicates, exhibit TP values do
do not meet the desired standard. Despite their substantial representation in the training
not meet the desired standard. Despite their substantial representation in the training da-
dataset with 301 and 512 instances, respectively, they achieved TP values of only 0.56 and
taset with 301 and 512 instances, respectively, they achieved TP values of only 0.56 and
0.57. This stands in stark contrast to the Kurodai Sea Hare that, with merely 112 instances
0.57. This stands in stark contrast to the Kurodai Sea Hare that, with merely 112 instances
in the ASP dataset, secured a TP value of 0.98 (as shown in Table 2.). This discrepancy
in the ASP dataset, secured a TP value of 0.98 (as shown in Table 2.). This discrepancy
signals a challenge for the AI model in recognizing these specimens accurately, highlighting
signals a challenge for the AI model in recognizing these specimens accurately, highlight-
a need for improved data quality in both collection and annotation by the CS participants.
ing a need for improved data quality in both collection and annotation by the CS partici-
Thus, the CS participants learn that data collection and annotation processes require further
pants. Thus, the CS participants learn that data collection and annotation processes re-
enhancement for better AI training outcomes.
quire further enhancement for better AI training outcomes.
Sustainability 2024, 16, 454 14 of 20

Table 2. The instance amount and true positive (TP) values of different species data collected with
the proposal survey protocol (PSP) and adaptive survey protocol (ASP).

PSP ASP
Common Name
Instance Species TP Instance Species TP
Eye Spot Sea Hare 499 0.97 489 0.97
Julian Sea Hare 210 0.98 210 0.98
Kurodai Sea Hare 112 1.00 112 0.98
Barnacle 812 0.83 804 0.84
Rock Oyster 205 0.87 205 0.90
Cowrie 276 1.00 276 0.96
Blotched Nerite 312 0.99 312 0.99
Blue-Ringed Octopus 145 1.00 145 0.94
Flat Rock Crab 72 1.00 72 1.00
Goby NA 632 0.71
Chition NA 195 0.93
Potamidids NA 1266 0.87
Sargassum NA 781 0.79
Sea Roach NA 478 0.75
Crescent Grunter NA 298 0.77
Sea Lettuce NA 640 0.85
Swimming Crab NA 275 0.96
Pebble Crab NA 202 1.00
Xantho Crab NA 192 0.98
Red Coral Algae (Corallina) NA 301 0.56
Ballweed NA 280 0.96
Sponge NA 267 0.89
Tunicates NA 512 0.57
Pyramid Periwinkle NA 737 0.97
Brittle Star NA 213 1.00
Damsel Fish NA 221 0.78

4.3. The Criteria for Developing Hybrid Intelligence

Ref. [57] suggests three criteria for the development of a hybrid intelligence system.
The first criterion, “collectiveness” within the context of HI pertains to the concerted effort
of CS and AI to fulfill a system-level objective. In this study, the YOLO object-detection
model necessitates an extensive array of specimen photographs with annotated instances
for effective model training, a process reliant on the gradual and cumulative data collection
by various CS participants over multiple surveys. Since individual contributions may be
limited due to time constraints and unpredictable specimen appearances, aggregating data
across different times is crucial for capturing species diversity. AI, thus, emerges as a cata-
lyst, unifying the efforts of the CS participants toward a common goal, the comprehensive
documentation of rocky intertidal ecosystem biodiversity.
The second criterion for effective HI is the achievement of “superior outcomes” com-
pared to those possible with CS or AI in isolation. Repetitive surveys and environmental
education tasks are often taxing, potentially diminishing volunteer enthusiasm. However,
the assurance that their diligent survey efforts contribute to AI databases for educational
purposes coupled with the positive feedback received from leveraging AI’s novel technol-
ogy sustain participant motivation and ensure ongoing involvement, a testament to the
project’s operational success.
Lastly, the establishment of “perpetual learning” constitutes the third criterion of HI.
The AI’s confusion matrix serves as a tool for assessing the quality of data for specific
species, encouraging CS participants to persist in their data collection efforts. This ongoing
collection process is not just during formal surveys but also through interactive specimen
identification using AI during educational activities. As data accumulate, they refine
the training dataset, enhancing AI’s recognition capabilities. Moreover, AI becomes an
 collection process is not just during formal surveys but also through interactive specimen
Sustainability 2024, 16, 454 identification using AI during educational activities. As data accumulate, they refine the
15 of 20
training dataset, enhancing AI’s recognition capabilities. Moreover, AI becomes an inval-
uable aid for newcomers in CS, facilitating rapid species identification skill acquisition
and perpetuating
invaluable the educational
aid for newcomers in CS,growth of the
facilitating entire
rapid CS community.
species identification skill acquisition
and perpetuating the educational growth of the entire CS community.
4.4. AI and Citizen Science: A Socio-Technological Synthesis
4.4. AIOur
andstudy
Citizenunderscores
Science: A Socio-Technological
the importance of Synthesis
not only building an AI model but also
managing the socio-technological
Our study integrationofbetween
underscores the importance not onlyAIbuilding
and CS anto ensure
AI modelthe but
effective
also
functioning
managing theofsocio-technological
HI. Inspired by McClure et al.between
integration [45], whoAI delineated eight key
and CS to ensure the attributes
effective
critical to CSofand
functioning HI.AI integration,
Inspired we delved
by McClure into
et al. these
[45], who attributes,
delineatedas eight
they provide a com-
key attributes
prehensive
critical to CSframework for managing
and AI integration, HI systems
we delved into thesewithin our research
attributes, domain.aThese
as they provide compre-at-
hensive
tributesframework for managing
and their complex HI systems within
interrelationships with AIourand
research domain. These
CS, alongside attributes
their associated
and their
tasks, arecomplex
depictedinterrelationships with AI9)and
in a mind map (Figure andCS,arealongside their associated
further dissected tasks, are
in the subsequent
depicted
sections.in a mind map (Figure 9) and are further dissected in the subsequent sections.

Figure9.9.The
Figure Theillustration
illustrationdepicts
depictsthethe integrative
integrative mind
mind mapmap of the
of the hybrid
hybrid intelligence
intelligence system,
system, en-
encom-
compassing
passing components
components of CS of CSAI,
and and AI, along
along withdistinct
with two two distinct tasks.
tasks. The Theismap
map is further
further character-
characterized by
ized by eight attributes crucial for effective system management. The directional arrows
eight attributes crucial for effective system management. The directional arrows within the diagram within the
diagram indicate the influences between components and attributes. Detailed explanations can be
indicate the influences between components and attributes. Detailed explanations can be found in
found in the accompanying text.
the accompanying text.
11. Subject
11. Subject Charisma:
Charisma:
The intertidal zone of reef rocks teems with diverse marine life and is subject to dy-
The intertidal zone of reef rocks teems with diverse marine life and is subject to
namic ecological changes, presenting urbanites from terrestrial settings with a vivid and
dynamic ecological changes, presenting urbanites from terrestrial settings with a vivid and
elusive tableau of the sea’s inhabitants. These marine creatures, often sporting distinctive
elusive tableau of the sea’s inhabitants. These marine creatures, often sporting distinctive
and vibrant appearances, make the zone an accessible, enigmatic, and captivating new
and vibrant appearances, make the zone an accessible, enigmatic, and captivating new
realm that beckons for repeated exploration. Notably, charismatic species like the Eye Spot
realm that beckons for repeated exploration. Notably, charismatic species like the Eye Spot
Sea Hare (Aplysia oculifera), which bears a resemblance to the popular cartoon character
Sea Hare (Aplysia oculifera), which bears a resemblance to the popular cartoon character
Pikachu, play a crucial role in heightening the involvement of CS participants and visitors.
Pikachu, play a crucial role in heightening the involvement of CS participants and visitors.
The combination of such appealing creatures and the deployment of smartphone-based
The combination of such appealing creatures and the deployment of smartphone-based
AI tools for specimen identification enhances the marine biodiversity exploration into an
AI tools for specimen identification enhances the marine biodiversity exploration into an
engaging and educational ecological quest. This synergy significantly boosts environmen-
engaging and educational ecological quest. This synergy significantly boosts environmental
tal education initiatives and fosters public engagement.
education initiatives and fosters public engagement.
12. Subject Identification:
12. Subject
Specimen Identification:
recognition represents a considerable challenge, particularly for tourists
withSpecimen
limited knowledge
recognitionofrepresents
marine biodiversity. Our challenge,
a considerable study’s protocol addresses
particularly this by
for tourists
with limited knowledge of marine biodiversity. Our study’s protocol addresses thissea-
providing a list of species that accounts for their varied appearances throughout the by
sons. Sucha measures
providing ensure
list of species thatthat tourists
accounts forcan identify
their varied marine life during
appearances their visit,
throughout enrich-
the seasons.
ing their
Such encounter
measures ensurewith
thatmarine
touristsbiodiversity
can identifyand making
marine the identification
life during process an
their visit, enriching ed-
their
encounter with marineinbiodiversity
ucational experience its own right.and making the identification process an educational
experience in its own right.
13. Public Engagement:
The hybrid intelligence-driven online AI service introduced in this study is a testament
to the power of engaging the public in environmental stewardship. By enabling visitors to
Sustainability 2024, 16, 454 16 of 20

identify organisms within the intertidal zone through an engaging process, we not only
educate but also discourage detrimental behaviors, such as the capture of marine life. This
strategy is key to garnering support for the conservation efforts outlined in 8. financial
considerations.
14. Volunteer Involvement:
The success of CS initiatives is heavily reliant on the dedication of volunteers. Despite
the logistical challenges posed by the remote location of our study site, the dynamic nature
of marine biodiversity here continually inspires local volunteers. The incorporation of AI
technology into their routine surveys has been instrumental in sustaining their enthusiasm,
as they often discover previously unrecorded species, underscoring the value of their
involvement and the importance of their continued participation.
15. Participant Training:
Training for CS participants is designed to be both accessible and practical, utilizing
localized survey protocols and tailored species lists to navigate the complexities of marine
biodiversity. Ongoing training workshops are scheduled to mitigate emerging challenges
and ensure participants are well-equipped to conduct AI-assisted field surveys, which are
vital for the AI’s continuous learning, as highlighted in 6. technical expertise.
16. Technical Expertise:
The gradual elevation of expertise among participants through hands-on surveys and
engagement with online resources or marine experts highlights the mutual learning process
central to HI. By contributing to image annotation, participants provide valuable data that
enhance the AI training process and the model’s subsequent performance.
17. Data Sensitivity:
Sensitive handling of the images uploaded by participants for AI identification is
paramount. Our protocol ensures that users are aware their contributions are confidential
and used solely for the purpose of enhancing the AI model. Such transparency is essential
in maintaining trust and encouraging the responsible sharing of data.
18. Financial Considerations:
The choice between establishing a dedicated AI computing infrastructure versus using
cloud AI services involves significant financial deliberations. Our decision to develop our
own system reflects our commitment to research flexibility. Yet, the financial sustainability
of both the AI infrastructure and the CS project remains a key concern, one that is necessary
for the unbroken operation of HI and the achievement of our conservation goals.
By dissecting and addressing these attributes, our research not only provides a tech-
nical blueprint for AI–CS integration but also offers insights into the socio-technological
considerations that underpin the successful deployment of HI systems for environmental
conservation.

4.5. The Role of Hybrid Intelligence in MPA Designation

Designing a marine protected area (MPA) is a complex task. As per the recent updates
to the US National Marine Sanctuaries Act, effective management and a comprehensive
understanding of the area over a decade are prerequisites [63]. Addressing these challenges,
particularly those related to personnel resources, the OCA report has indicated the valuable
role citizen scientists could play [48].
By collating specimen survey data to create training materials, CS participants can
perform year-round, consistent monitoring within specified areas. This method paves the
way for a deeper understanding of the seasonal patterns within marine communities and
yields more detailed data on the temporal and spatial diversity of marine ecosystems.
HI, which merges the collective efforts of citizen science and artificial intelligence (AI),
establishes a progressive framework for marine species education and motivates the public
to partake in regular spatiotemporal surveys. This process of collecting data at different
Sustainability 2024, 16, 454 17 of 20

times and locations is crucial to meeting the comprehensive training and data collection
demands essential for setting up an MPA.
Beyond the ecological considerations, MPAs situated near urban areas face a wider
spectrum of socio-ecological challenges. Not only must they address the transformation of
the economic activities of fishing communities, but MPAs in suburban locales also present
opportunities for urbanites to reconnect with nature, alleviating the stress of confined urban
living conditions [64]. Through participatory design involving the community and the
public, HI serves as a biophilic instrument that enhances urban livability.

5. Conclusions
Effective system management enhances the integration of CS and AI within the hybrid
intelligence system, providing robust solutions for intertidal biodiversity conservation
initiatives at Cape Santiago. Field surveys undertaken by trained CS participants generate
image datasets crucial for AI training. The training outcomes, representing a transformation
of multifaceted image data into numerical values, facilitate intertidal biodiversity analysis
and reflect the overarching quality of data procured by all CS participants. Once an AI
model is trained and available on the Web, it functions as a marine organism detection
tool, supporting CS participants in their environmental education endeavors, especially
for tourist instruction. Continuous feedback from user interactions precipitates holistic
improvements in the HI system, encompassing the fine-tuning of survey methodologies,
the prioritization of specific species, and the customization of AI tools to meet the demands
of in situ environmental education. Strategically managing this HI not only fosters sus-
tained engagement among CS participants but also bolsters broader public involvement,
thereby strengthening advocacy for the prospective designation of the region as a marine
protected area.

Author Contributions: Conceptualization, V.Y.C., D.-J.L. and Y.-S.H.; Data curation, V.Y.C.; Formal
analysis, V.Y.C.; Funding acquisition, Y.-S.H.; Investigation, V.Y.C.; Methodology, V.Y.C.; Resources,
Y.-S.H.; Software, V.Y.C.; Supervision, Y.-S.H.; Validation, V.Y.C. and D.-J.L.; Writing—original draft,
V.Y.C.; Writing—review and editing, V.Y.C. and D.-J.L. All authors have read and agreed to the
published version of the manuscript.
Funding: The National Science and Technology Council, Executive Yuan, Taiwan (MOST 111-2313-B-
002-016-MY3).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data are contained within the article.
Acknowledgments: We appreciate the support of the CHENG CHEN foundation, SDCDA, and the
invaluable contributions of all participating volunteers in gathering AI training data and providing
feedback for this study. Additionally, our sincere thanks go to the anonymous reviewers, whose
insightful comments significantly enhanced the quality of this manuscript.
Conflicts of Interest: The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in this paper.

References
1. Lubchenco, J.; Menge, B.A. Community development and persistence in a low rocky intertidal Zone. Ecol. Monogr. 1978, 48,
67–94. [CrossRef]
2. Polunin, N.V.C.; Raffaelli, D.G.; Williams, P.J.L. Marine Ecology: Processes, Systems and Impacts; Oxford University Press: Oxford,
UK, 2005.
3. Barbier, E.B.; Hacker, S.D.; Kennedy, C.; Koch, E.W.; Stier, A.C.; Silliman, B.R. The value of estuarine and coastal ecosystem
services. Ecol. Monogr. 2011, 81, 169–193. [CrossRef]
4. Worm, B.; Barbier, E.B.; Beaumont, N.; Duffy, J.E.; Folke, C.; Halpern, B.S.; Jackson, J.B.C.; Lotze, H.K.; Micheli, F.; Palumbi, S.R.;
et al. Impacts of biodiversity loss on ocean ecosystem services. Science 2006, 314, 787–790. [CrossRef] [PubMed]
Sustainability 2024, 16, 454 18 of 20

5. Stoeckl, N.; Birtles, A.; Farr, M.; Mangott, A.; Curnock, M.; Valentine, P. Live-aboard dive boats in the Great Barrier Reef: Regional
economic impact and the relative values of their target marine species. Tour. Econ. 2010, 16, 995–1018. [CrossRef]
6. Tapsuwan, S.; Asafu-Adjaye, J. Estimating the economic benefit of SCUBA diving in the Similan Islands, Thailand. Coast. Manag.
2008, 36, 431–442. [CrossRef]
7. Ergin, A.; Williams, A.T.; Micallef, A. Coastal scenery: Appreciation and evaluation. J. Coast. Res. 2006, 22, 958–964. [CrossRef]
8. Orams, M.B. Feeding wildlife as a tourism attraction: A review of issues and impacts. Tour. Manag. 2002, 23, 281–293. [CrossRef]
9. Curtin, S. Whale-watching in Kaikoura: Sustainable destination development? J. Ecotourism 2003, 2, 173–195. [CrossRef]
10. Velando, A.; Munilla, I. Disturbance to a foraging seabird by sea-based tourism: Implications for reserve management in marine
protected areas. Biol. Conserv. 2011, 144, 1167–1174. [CrossRef]
11. Schiel, D.R.; Taylor, D.I. Effects of trampling on a rocky intertidal algal assemblage in southern New Zealand. J. Exp. Mar. Biol.
Ecol. 1999, 235, 213–235. [CrossRef]
12. Ghazanshahi, J.; Huchel, T.D.; Devinny, J.S. Alternation of Southern California rocky shore ecosystems by public recreational use.
J. Environ. Manag. 1983, 16, 379–394.
13. Roy, K.; Collins, A.G.; Becker, B.J.; Begovic, E.; Engle, J.M. Anthropogenic impacts and historical decline in body size of rocky
intertidal gastropods in southern California. Ecol. Lett. 2003, 6, 205–211. [CrossRef]
14. Kido, J.S.; Murray, S.N. Variation in owl limpet Lottia gigantea population structures, growth rates, and gonadal production on
southern California rocky shored. Mar. Ecol. Prog. Ser. 2003, 257, 111–124. [CrossRef]
15. Murray, S.N.; Denis, T.G.; Kido, J.S.; Smith, J.R. Human visitation and the frequency and potential effects of collecting on rocky
intertidal populations in southern California marine reserves. Rep. Calif. Coop. Ocean. Fish. Investig. 1999, 40, 100–106.
16. Waters, C.; Turner, S. Defining the onset of the Anthropocene. Science 2022, 378, 6621. [CrossRef] [PubMed]
17. Neumann, B.; Vafeidis, A.T.; Zimmermann, J.; Nicholls, R.J. Future coastal population growth and exposure to sea level rise and
coastal flooding—A global assessment. PLoS ONE 2015, 10, e0118571. [CrossRef]
18. Rangel-Buitrago, N. Human epoch—Human responsibility: Rethinking coastal zone management in the Anthropocene. Ocean
Coast Manag. 2023, 244, 106801. [CrossRef]
19. Cigliano, J.A.; Meyer, R.; Ballard, H.L.; Freitag, A.; Phillips, T.B.; Wasser, A. Making marine and coastal citizen science matter.
Ocean Coast. Manage. 2015, 115, 77–87. [CrossRef]
20. Hyder, K.; Townhill, B.; Anderson, L.G.; Delany, J.; Pinnegar, J.K. Can citizen science contribute to the evidence-base that
underpins marine policy? Mar. Policy 2015, 59, 112–120. [CrossRef]
21. McKinley, D.C.; Miller-Rushing, A.J.; Ballard, H.L.; Bonney, R.; Brown, H.; Cook-Patton, S.C.; Evans, D.M.; French, R.A.; Parrish,
J.K.; Phillips, T.B.; et al. Citizen science can improve conservation science, natural resource management and environmental
protection. Biol. Conserv. 2017, 208, 15–28. [CrossRef]
22. Warner, K.A.; Lowell, B.; Timme, W.; Shaftel, E.; Hanner, R.H. Seafood sleuthing: How citizen science contributed to the largest
market study of seafood mislabeling in the US and informed policy. Mar. Policy 2019, 99, 304–311. [CrossRef]
23. Turrini, T.; Dörler, D.; Richter, A.; Heigl, F.; Bonn, A. The threefold potential of environmental citizen science—Generating
knowledge, creating learning opportunities and enabling civic participation. Biol. Conserv. 2018, 225, 178–186. [CrossRef]
24. Toivonen, T.; Heikinheimo, V.; Fink, C.; Hausmann, A.; Hiippala, T.; Järv, O.; Tenkanen, H.; Di Minin, E. Social media data for
conservation science: A methodological overview. Biol. Conserv. 2019, 233, 298–315. [CrossRef]
25. Nursey-Bray, M.; Palmer, R.; Pecl, G.T. Spot, log, map: Assessing a marine virtual citizen science program against Reed’s best
practice for stakeholder participation in environmental management. Ocean Coast. Manage. 2018, 15, 1–9. [CrossRef]
26. Cranswick, A.S.; Constantine, R.; Hendriks, H.; Carroll, E. Social media and citizen science records are important for the
management of rarely sighted whales. Ocean Coast Manag. 2022, 226, 106271. [CrossRef]
27. Matear, L.; Robbins, J.R.; Hale, M.; Potts, J. Cetacean biodiversity in the Bay of Biscay: Suggestions for environmental protection
derived from citizen science data. Mar. Policy. 2019, 109, 103672. [CrossRef]
28. Zettler, E.R.; Takada, H.; Monteleone, B.; Mallos, N.; Eriksen, M.; Amaral-Zettler, L.A. Incorporating citizen science to study
plastics in the environment. Anal. Methods 2017, 9, 1392–1403. [CrossRef]
29. Jarvis, R.M.; Bollard Breen, B.; Krägeloh, C.U.; Billington, D.R. Citizen science and the power of public participation in marine
spatial planning. Mar. Policy 2015, 57, 21–26. [CrossRef]
30. Danielsen, F.; Burgess, N.D.; Jensen, P.M.; Pirhofer-Walzl, K. Environmental monitoring: The scale and speed of implementation
varies according to the degree of people’s involvement. J. Appl. Ecol. 2010, 47, 1166–1168. [CrossRef]
31. Bonney, R.; Shirk, J.L.; Phillips, T.B.; Wiggins, A.; Ballard, H.L.; Miller-Rushing, A.J.; Parrish, J.K. Next steps for citizen science.
Science 2014, 243, 1427–1436. [CrossRef]
32. Pattengill-Semmens, C.; Semmens, B. Conservation and management applications of the REEF volunteer fish monitoring program.
Environ. Monit. Assess. 2003, 81, 43–50. [CrossRef] [PubMed]
33. Cigliano, J.A.; Kliman, R.M. Density, age structure, and length of queen conch (Strombus gigas) in shallow-water aggregations in
the Sapodilla Cayes Marine Reserve, Belize. Caribb. J. Sci. 2014, 48, 18–30. [CrossRef]
34. Zhang, J.; Chen, S.; Cheng, C.; Liu, Y.; Jennerjahn, T.C. Citizen science to support coastal research and management: Insights from
a seagrass monitoring case study in Hainan, China. Ocean Coast Manag. 2023, 231, 106403. [CrossRef]
35. Chiu, C.-C.; Liao, C.-P.; Kuo, T.-C.; Huang, H.-W. Using citizen science to investigate the spatial-temporal distribution of floating
marine litter in the waters around Taiwan. Mar. Pollut. Bull. 2020, 157, 111301. [CrossRef] [PubMed]
Sustainability 2024, 16, 454 19 of 20

36. Owens, K.; Divakaran Sarasamma, J.; Conlon, K.; Kiruba, S.; Biju, A.; Vijay, N.; Subramanian, M.; Asok Vijayamma, S.; Jayadev,
A.; Hoon, V.; et al. Empowering Local Practitioners to Collect and Report on Anthropogenic Riverine and Marine Debris Using
Inexpensive Methods in India. Sustainability 2022, 14, 1928. [CrossRef]
37. Miller-Rushing, A.; Primack, R.; Bonney, R. The history of public participation in ecological research Front. Ecol. Environ. 2012, 10,
285–290. [CrossRef] [PubMed]
38. Ward, E.J.; Marshall, K.N.; Ross, T.; Sedgley, A.; Hass, T.; Pearson, S.F.; Joyce, G.; Hamel, N.J.; Hodum, P.J.; Faucett, R. Using
citizen-science data to identify local hotspots of seabird occurrence. PeerJ 2015, 3, 695–704. [CrossRef]
39. Burgess, H.K.; Debey, L.B.; Froehlich, H.E.; Schmidt, N.; Theobald, E.J.; Ettinger, A.K.; Hillerislambers, J.; Tewksbury, J.; Parrish,
J.K. The science of citizen science: Exploring barriers to use as a primary research tool. Biol. Conserv. 2017, 208, 113–120. [CrossRef]
40. Earp, H.; Vye, S.; Bohn, K.; Burrows, M.; Chenery, J.; Dickens, S.; Foster, C.; Grist, H.; Lamont, P.; Long, S.; et al. Do You See What
I See? Quantifying Inter-Observer Variability in an Intertidal Marine Citizen Science Experiment. Citiz. Sci. Theory Pract. 2022, 7,
12. [CrossRef]
41. Hsu, Y.-C.; Huang, T.-H.K.; Verma, H.; Mauri, A.; Nourbakhsh, I.; Bozzon, A. Empowering local communities using artificial
intelligence. Patterns 2022, 3, 100449. [CrossRef]
42. Dellermann, D.; Calma, A.; Lipusch, N.; Weber, T.; Weigel, S.; Ebel, P. The Future of Human-AI Collaboration: A Taxonomy of
Design Knowledge for Hybrid Intelligence Systems. arXiv 2021, arXiv:2105.03354. [CrossRef]
43. Rafner, J.; Gajdacz, M.; Kragh, G.; Hjorth, A.; Gander, A.; Palfi, B.; Berditchevskaia, A.; Grey, F.; Gale, K.; Segal, A.; et al. Revisiting
Citizen Science Through the Lens of Hybrid Intelligence. arXiv 2021, arXiv:2104.14961. [CrossRef]
44. Lotfian, M.; Ingensand, J.; Brovelli, M.A. The Partnership of Citizen Science and Machine Learning: Benefits, Risks, and Future
Challenges for Engagement, Data Collection, and Data Quality. Sustainability 2021, 13, 8087. [CrossRef]
45. McClure, E.C.; Sievers, M.; Brown, C.J.; Buelow, C.A.; Ditria, E.M.; Hayes, M.A.; Pearson, R.M.; Tulloch, V.J.D.; Unsworth, R.K.F.;
Connolly, R.M. Artificial Intelligence Meets Citizen Science to Supercharge Ecological Monitoring. Patterns 2020, 1, 100109.
[CrossRef] [PubMed]
46. Ocean Conservation Administration. Ocean Affairs Council, Taiwan. Ocean Resource Protected Area. Available online:
https://mpa.oca.gov.tw/ProtectedDetail.aspx?id=30 (accessed on 16 August 2023).
47. Ocean Affairs Council, Taiwan. Draft of Marine conservation Act. 2019. Available online: https://www.oac.gov.tw/ch/home.
jsp?id=67&parentpath=0,6&mcustomize=bulletin_view.jsp&dataserno=201912030004 (accessed on 12 July 2022).
48. Ocean Conservation Administration. Ocean Affairs Council, Taiwan. Survey Report. 2021. Available online: https://www.oca.
gov.tw/ch/home.jsp?id=220&parentpath=0&mcustomize=research_view.jsp&dataserno=202101280023 (accessed on 15 August
2023).
49. Chen, V.Y. The Precious Secrets of a Tidal Pool; Flâneur Culture Lab Publishers: Taipei, Taiwan, 2011.
50. Jie, W.B.; Gan, B.Q.; Chen, V.Y.; Kuo, S.H. Pseudoceros magangensis: A new species of pseudocerotid flatworm (Platyhelminthes:
Polycladida) from Taiwan. Platax 2016, 13, 3.
51. Fukushima, K. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift
in position. Biol. Cybern. 1980, 36, 193–202. [CrossRef]
52. Redmon, J.; Farhadi, A. YOLOv3: An Incremental Improvement. arXiv 2018, arXiv:1804.02767.
53. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards real-time object detection with region proposal networks. Adv. Neural
Inf. Process. Syst. arXiv 2016, arXiv:1506.01497. [CrossRef]
54. Glenn, J. YOLOv5 Release v6.1. 2022. Available online: https://github.com/ultralytics/yolov5/release/tag/v6.1 (accessed on 12
January 2023).
55. Song, K.; Jung, J.; Lee, S.; Park, S.; Yang, Y. Assessment of Marine Debris on Hard-to-Reach Places Using Unmanned Aerial
Vehicles and Segmentation Models Based on a Deep Learning Approach. Sustainability 2022, 14, 8311. [CrossRef]
56. Yang, X.; Samsudin, S.; Wang, Y.; Yuan, Y.; Kamalden, T.; Yaakob, S. Application of Target Detection Method Based on
Convolutional Neural Network in Sustainable Outdoor Education. Sustainability 2023, 15, 2542. [CrossRef]
57. Dellermann, D.; Ebel, P.; Söllner, M.; Leimeister, J.M. Hybrid Intelligence. Bus. Inf. Syst. Eng. 2019, 61, 637–643. [CrossRef]
58. Loke, L.H.; Todd, P.A. Structural complexity and component type increase intertidal biodiversity independently of area. Ecology
2016, 97, 383–393. [CrossRef] [PubMed]
59. Davidson, I.C.; Crook, A.C.; Barnes, D.K.A. Quantifying Spatial Patterns of Intertidal Biodiversity: Is Movement Important? Mar.
Ecol. 2004, 25, 15–34. [CrossRef]
60. Kosmala, M.; Wiggins, A.; Swanson, A.; Simmons, B. Assessing data quality in citizen science. Front. Ecol. Environ. 2016, 14,
551–560. [CrossRef]
61. Rosenthal, I.S.; Byrnes, J.E.; Cavanaugh, K.C.; Bell, T.W.; Harder, B.; Haupt, A.J.; Rassweiler, A.T.; Pérez-Matus, A.; Assis, J.;
Swanson, A.; et al. Floating forests: Quantitative validation of citizen science data generated from consensus classifications. arXiv
2018, arXiv:1801.08522.
62. van der Velde, T.; Milton, D.A.; Lawson, T.; Wilcox, C.; Lansdell, M.; Davis, G.; Perkins, G.; Hardesty, B.D. Comparison of marine
debris data collected by researchers and citizen scientists: Is citizen science data worth the effort? Biol. Conserv. 2017, 208, 127–138.
[CrossRef]
Sustainability 2024, 16, 454 20 of 20

63. National Oceanic and Atmospheric Administration, Legislative History of the National Marine Sanctuaries Act. Available online:
https://sanctuaries.noaa.gov/about/legislation/leg_history.html (accessed on 20 September 2023).
64. Lefosse, D.; van Timmeren, A.; Ratti, C. Biophilia Upscaling: A Systematic Literature Review Based on a Three-Metric Approach.
Sustainability 2023, 15, 15702. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.