Introduction
Biodiversity is essential for human health, well-being, and a stable
environment. Although significant efforts have been devoted toward
conservation, biodiversity loss remains a global challenge (Johnson et
al. 2017). Anthropogenic activities such as urbanization, agricultural
intensification, and species exploitation reduce biodiversity, and
evidence indicates that species extinction rates are progressing much
faster than in the past (Ceballos et al. 2015). In addition,
globalization has led to the introduction of various organisms from
their native habitats into new environments and the establishment of
non-native populations in new areas. These invasive species cause
ecosystem impacts such as predation, niche displacement, and
introduction of diseases (Doherty et al. 2016; Haubrock et al. 2021;
Kortz and Magurran 2019; Scheele et al. 2019). Furthermore, non-native
species are recognized as a driver of recent extinctions (Bellard et al.
2016). The impact of non-native species on biodiversity and ecosystems
is accelerating, and this trend is expected to continue (Pyšek et al.
2020). Therefore, the mitigation of biological invasions is essential
for biodiversity conservation.
When non-native species are introduced into a new habitat, newcomers
sometimes encounter close relatives. In such cases, hybridization occurs
due to incomplete reproductive isolation from closely related species.
Hybridization in non-native species is frequently observed and
considered an evolutionary mechanism that determines invasion success
(Bock et al. 2021). For example, hybrid fitness is occasionally superior
to the parental species; (i.e., hybrid vigor). In addition, hybrids also
have intermediate traits or different traits from the parent species,
and some traits may determine the establishment success of invasive
species (Coulter et al. 2020). For instance, a meta-analysis of plants,
animals, and fungi demonstrated that invasive hybrids have a larger body
size and are more fecund than their parent species (Hovick and Whitney
2014). Furthermore, early invasive populations are affected by
density-dependent processes such as the Allee effect. However,
hybridization provides mating partners for invasive species, which could
reduce the Allee effect and promote invasions (Yamaguchi et al. 2019).
Hybrids of similar species pose a threat to genetic diversity because
introduced alleles may eventually replace the native alleles
(Fitzpatrick and Shaffer 2007). Therefore, controlling hybrid species is
necessary to conserve biodiversity. However, difficulties in
distinguishing between native and hybrid species is a critical issue
when trapping hybrids. Hybrids were detected using morphological
characteristics until the mid-1960s (Allendorf et al. 2001). This
approach assumes that hybrids exhibit intermediate characteristics of
their parent species; however, this assumption does not generalize to
all cases because they often show a mosaic of parental phenotypes. In
addition, morphological characteristics cannot be determined whether an
individual is a first-generation or a backcross-generation hybrid.
Misidentification of species can also cause conservation problems. For
example, inadequate identification of target species could negatively
impact native species. In fact, native frogs have been killed in
Australia due to misjudgments while removing the non-native cane toad
(Rhinella marina ) (Somaweera et al. 2010).
The development of molecular genetic techniques, such as allozyme
electrophoresis and PCR, has overcome these challenges (Allendorf et al.
2001). DNA analysis allows accurate species identification and can
reveal individuals’ degree of hybridization, which would be difficult to
determine using morphological traits. However, these analyses are
time-consuming and costly, limiting the quick identification of hybrids
and large-scale surveys.
In recent years, deep learning image recognition technology, a novel
group of artificial intelligence approaches, has begun to be utilized in
both species and individual identification in ecology. Identifying and
counting animal species in images provides basic but essential
information (Tuia et al. 2022). Many previous studies have combined
camera traps and deep learning to identify species. For instance,
Norouzzadeh et al. (2018) used 3.2 million images from camera traps in
the Serengeti National Park to successfully identify 48 species. In
addition, these techniques have been applied to individual
identification, such as green turtles (Carter et al. 2014), chimpanzees
(Schofield et al. 2019), and brown bears (Clapham et al. 2020).
Furthermore, this technology has already been used to detect non-native
species (Ashqar and Abu-Naser 2019; Guo et al. 2022; Takaya et al.
2022). Although a similar approach may provide a new method for
identifying hybrids in the field, studies have yet to apply deep
learning models to identify hybrids.
The Japanese giant salamander (Andrias japonicus ) is an amphibian
endemic to Japan and is threatened with extinction, as its population
has declined due to habitat degradation and fragmentation (Tochimoto et
al. 2007; Taguchi and Natuhara 2009; Yamasaki et al. 2013). In the 2022
IUCN Red List, the conservation status rank of this species was changed
from Near Threatened to Vulnerable (IUCN 2022). One reason for this
change is its hybridization with the non-native Chinese giant salamander
(Andrias davidianus ), which is the same genus as A.
japonicus . The Chinese giant salamander is also threatened with
extinction in their original habitat, but individuals introduced to
Japan in the early 1970s have become wild, and hybridization with
Japanese giant salamanders is an issue. For example, the Kyoto City
government survey revealed that only 4 (2%) out of 244 captured
individuals were native species, and the remaining were 240 (98%)
hybrids and non-native species in the Kamo river basin in Kyoto,
requiring rapid action (The Kyoto City Government 2015). However, the
number of areas where hybrids were caught is increasing and has already
been confirmed in six prefectures in western Japan (Kyoto, Mie, Nara,
Shiga, Okayama, and Hiroshima). Since hybrid species have a spot pattern
that inherits the characteristics of both native and non-native species,
individuals with the potential for hybridization are captured by visual
screening and DNA analysis is also applied for accurate identification.
While this approach is reliable, identifying hybrids by their spot
patterns requires specialized knowledge, and DNA analysis is
time-consuming and expensive. If identifying hybrid salamanders from
images could work well, it does not need time and cost as DNA analysis.
It also facilitates early detection and effective capture of suspected
hybrid individuals via citizen science, thereby contributing to the
effective conservation of native Japanese giant salamanders.
The first objective of this study is to identify hybrids between
Japanese giant salamanders and Chinese giant salamanders from images
based on deep learning. Our approach allows the public to photograph and
detect hybrid individuals without specialized knowledge. In recent
years, citizen science has been adopted to manage invasive species
(Larson et al. 2020), and a similar method could be applied to hybrids.
The second objective is to clarify which features the AI model uses as
criteria to determine hybrids. Spot patterns are difficult to quantify
compared to measurable morphological features. However, techniques such
as Grad-CAM allow visualization of the important region for the AI
model’s prediction. If specific essential areas in identifying hybrids
can be clarified, that information is valuable for the general public to
identify hybrids. Although there is a proposal to divide the Chinese
giant salamander into three species (Turvey et al. 2019), our study usesAndrias davidianus instead of making this distinction.