(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Inferring the extinction risk of marine fish to inform global conservation priorities [1]

['Nicolas Loiseau', 'Marbec', 'Univ Montpellier', 'Cnrs', 'Ifremer', 'Ird', 'Montpellier', 'David Mouillot', 'Institut Universitaire De France', 'Paris']

Date: 2024-09

While extinction risk categorization is fundamental for building robust conservation planning for marine fishes, empirical data on occurrence and vulnerability to disturbances are still lacking for most marine teleost fish species, preventing the assessment of their International Union for the Conservation of Nature (IUCN) status. In this article, we predicted the IUCN status of marine fishes based on two machine learning algorithms, trained with available species occurrences, biological traits, taxonomy, and human uses. We found that extinction risk for marine fish species is higher than initially estimated by the IUCN, increasing from 2.5% to 12.7%. Species predicted as Threatened were mainly characterized by a small geographic range, a relatively large body size, and a low growth rate. Hotspots of predicted Threatened species peaked mainly in the South China Sea, the Philippine Sea, the Celebes Sea, the west coast Australia and North America. We also explored the consequences of including these predicted species’ IUCN status in the prioritization of marine protected areas through conservation planning. We found a marked increase in prioritization ranks for subpolar and polar regions despite their low species richness. We suggest to integrate multifactorial ensemble learning to assess species extinction risk and offer a more complete view of endangered taxonomic groups to ultimately reach global conservation targets like the extending coverage of protected areas where species are the most vulnerable.

Funding: NL is supported by the Fondation pour la Recherche sur la Biodiversité (FRB) and La region Occitanie (BiodivOc) in the context of the CESAB project ‘Creating a global database of fish functional traits: integrating physiology and ecology across aquatic ecosystems’ (PHENOFISH). DM was partly funded through the 2017–2018 Belmont Forum and BiodivERsA REEF-FUTURES project under the BiodivScen ERA-Net COFUND programme and with funding from ANR, DFG, NSF, Royal Society, ERC and NSERC. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Using available occurrence data, species biological traits, taxonomy, and human uses (A), we built an ensemble learning model using RF and ANN (B) to predict the IUCN status of marine fishes using complementary decisions between ANN and RF outputs (C). Then, we explored the consequences of including the predicted threatened species on the areas currently prioritized by conservation planning (D). See methods for a complete description of these steps. Map was created using R package rnaturalearth ( https://www.naturalearthdata.com/ ). ANN, artificial neural network; IUCN, International Union for the Conservation of Nature; RF, random forest.

Here, we used a combination of random forest model (RF) and artificial neural network algorithm (ANN) to predict extinction risk of 4,992 DDNE marine fish species ( Fig 2 ) based on their occurrence data, traits (i.e., body size, trophic position), taxonomy, and human uses. We then addressed 4 principal questions for the conservation of marine fishes: 1. Which attributes of a species are the best predictors of their extinction risk? 2. How does the addition of species predicted as Threatened (Critically Endangered, Endangered, and Vulnerable) change the distribution of hotspots of extinction risk? 3. Does the current network of marine protected areas (MPAs) cover this threatened marine fish diversity? 4. To which extent does the classification of marine fishes with no IUCN status modify conservation priorities to meet the 30 × 30 target?

IUCN-assessed species were classified into 3 categories: “Threatened” gathering the Critically Endangered, Endangered, and Vulnerable IUCN categories; “Non-Threatened” gathering the Least Concern and Near Threatened IUCN categories; and “DDNE” gathering the Data Deficient and Not Evaluated IUCN categories. Because some species were not present in the IUCN Red List, we updated the species list with https://www.birdlife.org/ , http://www.reptile-database.org/db-info/SpeciesStat.html , https://www.mammaldiversity.org/ , https://amphibiansoftheworld.amnh.org/ . Icons were generated using R (rphylopic package) and are under the Creative Commons Attribution 4.0 International (CC BY 4.0) License. The data underlying this figure can be found in https://zenodo.org/records/12783687 . IUCN, International Union for the Conservation of Nature.

This IUCN categorization is overdue and urgently needed for marine fishes which are highly diverse (N > 15,000 species), among which many are facing multiple threats [ 24 – 26 ] and support key contributions to nature and people like nutrient cycling, carbon sequestration, ecosystem resilience, productivity, as well as nutritional and cultural values [ 27 – 29 ]. Among vertebrates, marine teleost fishes have the highest proportion of DDNE species (38%, n = 4,992 of 13,195, Fig 1 ). Ultimately, a more extensive and accurate IUCN classification of marine fishes could reevaluate the prioritization of new protected areas. This is particularly relevant concerning the new agenda to protect 30% of marine waters before 2030 (30 × 30 target, Conference of the Parties to the Convention on Biological Diversity, COP 15).

Some methods have been proposed to infer the IUCN status of unassessed (DD and NE) species [ 11 , 5 ]. Species distribution models (SDMs) can predict the spatial distribution or the temporal dynamics of populations according to their environmental or ecological niches [ 12 ]. However, these models require a large spectrum of variables (such as climatic niche, habitat, human footprint index, geographic distribution, phylogeny, and traits) to eventually predict the loss of suitable habitat and then extinction risk [ 13 ]. Modeling the distribution of rare, and therefore the most threatened species [ 14 , 15 ] is also associated with a high level of uncertainty due to the low number of occurrences, which can lead to model overfitting and inaccuracy [ 16 ]. SDMs also require climate projections from global circulation models (GCMs) that in turn rely on socioeconomic scenarios (RCPs) making the overall process challenging to achieve for many taxonomic groups [ 17 ]. Meanwhile, the prediction of species IUCN status has benefited from the development of machine learning models with the underlying assumption that unassessed species are likely to share a similar IUCN status to those having similar biological traits, geographic distribution, or evolutionary history [ 18 ]. For example, machine learning models could predict IUCN categories for mammals, amphibians, reptiles [ 19 – 22 ], sharks, rays [ 23 ], and orchids [ 18 ] with high accuracy (e.g., up to 92% for terrestrial mammals [ 19 ]).

In 2023, the IUCN Red List contains 150,388 species (including mammals, birds, reptiles, amphibians, fishes, insects, and plants) classified in 3 categories: (1) “Threatened,” which encompasses the Critically Endangered, Endangered, and Vulnerable IUCN categories; (2) “Non-Threatened,” which includes the Least Concern and Near Threatened IUCN categories; and (3) Data Deficient (DD), which contains the largest number of species. In addition, most animal biodiversity (estimated to >1.8 million species of metazoans) has not been evaluated (NE). Even for the most studied vertebrate taxa such as mammals and reptiles, the proportion of DD or NE species, hereafter grouped under the DDNE category, is still high (respectively 22.9% and 27.8%, Fig 1 ). This knowledge gap may leave threatened species out of conservation priorities and bias conservation prioritization based on extinction risk. For example, in 2014 nearly half (454 species) of sharks and rays (chondrichthyan) were still NE, and after a new assessment of this group in 2021, 37.5% (against 17%) were classified as threatened by extinction [ 5 , 6 ]. Given the global biodiversity crisis, the ever-increasing number of threatened species urges for a greater collective effort to fill these IUCN classification gaps to better guide conservation planning. Yet, this ambitious goal is far from being reachable given the millions of species on Earth and the inherent difficulty to obtain accurate information for most of them due to their remote or hardly accessible habitat (e.g., deep sea, high mountain), behavior (e.g., elusive, nocturnal), body size (e.g., <1 cm), or rarity (e.g., endemic). Alternative methods are thus needed to predict species extinction risk status and to ultimately fuel global conservation prioritization algorithms [ 7 – 10 ].

Target 3 of the Convention on Biological Diversity’s Kunming-Montreal Global Biodiversity Framework—adopted in December 2022—aims to increase the global coverage of protected areas (PAs) to at least 30% by 2030 (hereafter referred to as 30 × 30), with the ultimate goal to deliver benefits for nature and people where the needs are the most pressing. Consequently, prioritizing the establishment of new PAs and the strategic use of limited conservation resources are crucial to mitigate the ongoing global biodiversity crisis [ 1 , 2 ]. However, a strategy that protects as many species as possible—regardless of the risk of extinction—may lead to a different prioritization of new protected areas than a strategy that emphasizes the protection of the most threatened species. To address this issue, assessing species extinction risk is of primary importance despite persistent challenges [ 3 , 4 ]. In this regard, the International Union for the Conservation of Nature (IUCN) regularly updates the global Red List ( www.iucnredlist.org ), which classifies species by their increasing extinction risk (Vulnerable, Endangered, and Critically Endangered) mainly based on their population and geographic range size. Yet, this classification requires extensive knowledge and many species with limited information are considered as Data Deficient.

( A ) Relationship between ranks before and after predictions; dots color gradient indicates cell species richness (log10). The dashed line represents x = y (i.e., cells above this line have seen their priority rank increasing). ( B ) Relationship between the change in rank (rank after minus rank before) of each cell and its latitude; dots color gradient indicates cell species richness (log10). Lines show for both, negative and positive delta rank values, the relationships with latitude were obtained with a generalized additive model. The data underlying this figure can be found in https://zenodo.org/records/12783687 . IUCN, International Union for the Conservation of Nature.

Overall, we found a marked change in conservation priority ranking after species IUCN status predictions. This is particularly true at low and intermediate values of species richness where the ranking is more likely to shift ( Fig 7A and Fig H in S1 Text ). By plotting the delta ranks (rank after minus rank before) on the latitudinal gradient, we found that the major changes in high ranking were at low (<30°) and high latitudes (>50°) corresponding to temperate and polar climatic zones for which species richness is the lowest, as well as in Pacific islands (Figs 5D , 7B , and Fig H in S1 Text ).

To test whether the predictions of the IUCN status for currently DDNE species could disrupt conservation planning, we compared conservation priorities based on assessed species IUCN status (Scenario 1: 7,869 Non-Threatened, 4,992 DDNE, and 334 Threatened species) with conservation priorities accounting for new species predicted IUCN status (Scenario 2: 10,451 Non-Threatened, 1,073 DDNE, and 1,671 Threatened species). We used the Zonation algorithm that identifies which locations in a seascape are most important for protecting threatened biodiversity (see Methods ). Zonation ranks locations (hereafter “cells”) in function of their importance for conservation. The least valuable cells received the lowest ranks (0), and those having the highest priority reached the highest ranks. We fixed species priority weights to 1 for Non-Threatened species, 6 for Threatened species, and 2 for DDNE species because there is accumulating evidence that DD and also NE species are more at risk than Non-Threatened [ 30 ] (see Methods for details on the weighting and Fig G in S1 Text ). We then compared the ranking of each cell between both scenarios (with and without predicted IUCN status).

Furthermore, we assessed species-specific coverage by the global network of MPAs and target achievement, defined as the proportion of a species’ geographic range covered by these protected areas, before and after predictions of species IUCN status. These specific targets were related to species range sizes with the most restricted species needing more coverage (e.g., 100%) than widespread one (e.g., 10%) to avoid extinction. First, we found that regardless of gap-filling status predictions, target achievement, and MPA coverage of Threatened and DDNE species were significantly smaller than for Non-Threatened species ( Fig 6 , Kruskal–Wallis chi-squared = 617, df = 2, P < 0.001). Threatened species were significantly less protected than DDNE species, but to a lower extent ( Fig 6 , Kruskal–Wallis multiple comparison, Z = 2.87, P < 0.05). Second, we observed a decrease in the attainment of target protection for DDNE (Wilcoxon test, W = 2785657, P < 0.05) and Threatened species, albeit not significantly (Wilcoxon test, W = 271429, P = 0.4), indicating that these species were not as protected as they should be. Average target achievement for Threatened and DDNE species were both equal to 3.6% before IUCN status predictions compared to 2.3% and 2.5%, respectively after predictions. However, we did not observe a significant difference in the percentage of species-range covered by MPAs for Threatened species before and after IUCN status predictions ( Fig 6 , Wilcoxon test, W = 271429, P = 0.4).

Spatial distribution of the difference in the number of Threatened (A), Non-Threatened (B), and DDNE (C) species before and after prediction, and in the prioritization rank after the prediction (rank after—rank before, D). The color gradient indicates the difference in the number of species or prioritization ranking from red (value of the cell is higher after prediction) to blue (value of the cell is lower after prediction). Green indicates already protected cells. Maps were created using the R package rnaturalearth ( https://www.naturalearthdata.com/ ). The data underlying this figure can be found in https://zenodo.org/records/12783687 .

Before gap-filling prediction, Threatened species were mainly aggregated in the Caribbean, the South China Sea, the Philippine Sea and the Celebes Sea. After prediction, new hotspots of Threatened species emerged in western Australia and on the west coast of North America ( Fig 5A ). Overall, the China Sea, Philippine Sea, and south Japan aggregated the highest number of species predicted as Threatened ( Fig 5A ). The distribution of Non-Threatened species before and after prediction followed the gradient of marine fish richness, with Non-Threatened species peaking in the Indian Ocean and Coral Triangle ( Fig 5B ). Finally, the remaining DDNE species were mainly aggregated in the China Sea, the Philippine Sea, and in southern Japan ( Fig 5C ).

RF models provided information about which features were the best to predict species IUCN status ( Fig 4 ). Species predicted as Threatened were mainly characterized by a small geographic range, a relatively large body size, and a low growth rate ( Fig 4 ). The likelihood of species being Threatened increased with their preference for very shallow habitats ( Fig 4 ). We also found that the Family variable contributed to the prediction of the IUCN status probably due to the strong phylogenetic conservatism in size, growth rate, and vertical position. Closely phylogenetically related species were indeed significantly more likely to share the same IUCN status than distantly related species (Fig D in S1 Text ). Some families gathered a high proportion of species predicted as Threatened (phylogenetic signal D index = 0.68 ± 0.01). For example, we predicted 19 species of Bythitidae out of 28 (67.8%) and 56 species of Serranidae out of 131 (42.7%), as Threatened, thus changing their previous DDNE classification (Fig E in S1 Text ). Some cryptobenthic fish families like Gobiidae, Gobiesocidae, Blennidae had also an important proportion of species predicted as Threatened. Conversely, some families hosted a low proportion of species predicted as Threatened. For example, none of the 17 DDNE Myctophidae were predicted as Threatened. Overall, we found that species predicted as Non-Threatened or remaining DDNE were also clustered across the phylogeny (D NonT index 0.59 ± 0.01; D DDNE 0.90 ± 0.01, Fig D in S1 Text ).

Distribution of species among the extinction risk IUCN categories before (i.e., based on current Red List) and after predictions. Species are grouped into 3 broad categories following the IUCN status: “Threatened” (red) including Critically Endangered, Endangered, and Vulnerable species; “Non-Threatened” (blue) including Least Concern and Near Threatened species; DDNE (gold) merging Data Deficient and Not Evaluated species. The DDNE category after prediction refers to the species for which the 2 algorithms disagreed, no prediction could be made. The data underlying this figure can be found in https://zenodo.org/records/12783687 . IUCN, International Union for the Conservation of Nature.

After the within-consensus framework (i.e., cross-check within each algorithm, see Methods ), we combined ANN and RF outputs using a complementary decision tree: A status (Threatened or Non-Threatened) was attributed to a given species when both methods predicted the same status or when only one method was able to accurately predict a status while the DDNE status was kept when the predictions of the two algorithms differed ( S1 Table ). Predictions differed for 573 species that remained DDNE. Overall, out of the 4,640 DDNE species (4,992 minus the 352 unpredictable species with too many missing trait values and that remained DDNE, see Methods ), 1,337 were categorized as Threatened and 2,582 as Non-Threatened, resulting in a much higher proportion of threatened species than expected from the current IUCN categorization ( Fig 3 ). Overall, the number of DDNE species was reduced by 78.5% (1,073 out of 4,992 species remained DDNE), the number of Threatened species increased by 400% (from 334 to 1,671) while the number of Non-Threatened species increased only by 34.8% (from 7,750 to 10,451). We also applied a consensus decision from ANN and RF outputs and found that even if the number of predicted species decreased, the number of Threatened species increased (824) disproportionately compared to the number of Non-Threatened species (1,846 see Fig C in S1 Text ).

For both RF and ANN, we performed cross-validation, it order to determine their accuracy (to predict IUCN status) based on the proportion of false positives (proportion of Non-Threatened predicted as Threatened) and false negatives (proportion of Threatened predicted as Non-Threatened). We found that RF models better predicted species extinction risk (accuracy of 0.77) than ANN algorithms (accuracy of 0.70). RF achieved a high rate of true-positive predictions (77.3%, SD = 3.89%, Figs A and B in S1 Text ) in cross-validation tests and a low rate of false-positives (12%, SD = 3.02%, Figs A and B in S1 Text ) and false-negatives (10.7%, SD = 3.1, Figs A and B in S1 Text ). ANN achieved a lower rate of true-positives (70%, SD = 3.67%, Figs A and B in S1 Text ) with a higher rate of false-positives (13.8%, SD = 3.62%, Figs A and B in S1 Text ) and false-negatives (16.2%, SD = 3.51%, Figs A and B in S1 Text ). RF models were able to predict the IUCN status of fewer species (2,324 Non-Threatened and 1,440 Threatened species) than ANN algorithms (2,677 Non-Threatened and 1,294 Threatened species).

Discussion

Models will never replace a direct evaluation of species extinction risk based on empirical robust data, but coupling machine and deep learning methods offer a unique opportunity to provide a rapid, extensive, and cost-effective evaluation of extinction status [20] while also pointing out the species on which data collection and conservation efforts should be prioritized. Several studies have already proposed automated methods to conduct a preliminary assessment of species conservation status based on their attributes or remotely sensed predictors [3,11,18–23,31]. However, to our knowledge, they have not yet been incorporated in the official Red List assessment [4]. We believe that ensemble learning is relevant since it is accurate and conservative. The performance of machine learning algorithms is known to vary based on factors such as the dimensionality of the data set [31]. To address this variability, we suggest a multi-model strategy combining distinct algorithms to leverage their strengths and mitigate their weaknesses. With relatively small data sets, Random Forests can achieve a high level of accuracy, whereas Neural Networks typically require more data to reach a similar level of performance [32]. Conversely, Random Forests show minimal performance improvement beyond a certain data threshold, while Neural Networks generally benefit from larger data sets and continuously improve their accuracy [32]. Random Forests are also advantageous in terms of interpretability, because they highlight which features are the best to predict species IUCN status.

The accuracy of our models (0.77, 0.70 for the RF and ANN, respectively) was slightly lower compared to the binary classifier developed by Borgelt and colleagues [33] for amphibians (85%) or the IUC-NN classifier developed by Zizka and colleagues [18] for orchids (84%). This lower accuracy can be attributed to the limited number of Threatened species included in our training data set whereas, for example, Zizka and colleagues [18] had a significantly larger representation of Threatened species, accounting for 49.7% of their data set. Our data set covered almost all marine fishes, with a very high initial number of Non-Threatened species (7,869) and a relatively low number of Threatened species (334). Furthermore, both Borgelt and colleagues [33] and Zizka and colleagues [18] classifiers were trained using habitat data, which unfortunately were not available for the majority of marine fish species in our study. Adding information about marine fish habitats should be a priority to increase the accuracy of our models which can be offered by the recent developments in satellite or acoustic imagery [34]. The robustness of our predictions was increased by our decision tree relying on (1) the within-consensus framework (cross-check within each algorithm); and (2) the use of 2 models rather than 1, even if this tends to predict fewer species status than using only 1 model. Indeed, if the predictions between the 2 models differ, we do not provide a status—this applies to 573 species (12% of the 4,640 DDNE in the models). Altogether, the multi-model strategy we proposed appears like a good compromise between accuracy and conservatism to predict IUCN status and should be tested on more taxa to validate its utility as a companion tool for IUCN assessment.

Using only 3 categories—Threatened, Non-Threatened, and DDNE—comes with some limitations. For example, assigning the same weight to Vulnerable or Critically Endangered species despite their distinct status may not accurately reflect the varying levels of conservation effort required for their protection. This was again linked to the low number of species within these 2 categories (224 VU, 77 EN, and 33 CR) in the original data set. It suggests prioritizing the direct evaluation of species for which we have predicted Threatened status to be able to refine our predictive model. The 573 species for which we could not reach a consensus on both models should be also prioritized for future evaluation since one of the algorithms predicted them as Threatened. Another limitation of our approach stands from having predicted missing species traits with a Random Forest algorithm, which may ultimately lead to a misclassification of some species. However, only traits with a missForest performance exceeding 0.6 (R squared > 0.6 for regression or 60% for classification) were attributed, thus minimizing the probability of errors in trait inference. Moreover, we used coarse biological traits already available on Fishbase. Although we acknowledge that extinction probabilities are related to species responses to climate change [26], accounting for such effects would require gathering more ecophysiological-based traits [25,26] (i.e., metabolic rates, thermal optimum, reproduction). While these traits were not available for most marine species, the growing availability of fish traits will ultimately make categorical predictions of conservation status more effective in the near future. Finally, we used species range maps provided by Albouy and colleagues [35] which do not perfectly reflect the current distribution of marine fishes but are nevertheless based on a robust method to minimize errors in the original OBIS data set (see Methods). Since OBIS is continuously aggregating new observations, it cannot assess range contractions or regional extirpations due to environmental shifts or overexploitation [36,37], which could result in an underestimation of the number of Threatened species.

Altogether, the limitations of the in silico species-risk assessment open opportunities for improvements and inputs from the organization (IUCN) for which the predictions are made, which could trigger a positive virtuous loop and lead to an effective in situ/silico assessment of species extinction risks. Indeed, our prediction of the IUCN status for marine fish species shows a fivefold increase of fish species with a Threatened status (from 334 to 1,671 Threatened species), so from 2.5% to 12.7% of total species richness. Meanwhile, the number of species with Non-Threatened (NT) status only increased by 34.8% (from 7,869 to 10,451 Non-Threatened species). Overall, 1,073 species remained DDNE (8.1%), which suggests that there is still some potential to increase the accuracy of our predictive model. Even when we applied a consensus decision to the ANN and RF outputs, we found that the number of Threatened species increased disproportionately, from 2.5% to 8.8%.

Given the strong phylogenetic conservatism of environmental and trophic niches among marine fishes [38], we expected that the assessment of a species as Threatened would often coincide with the status of its closest relatives. Thus, a strong proportion of species from some families like Sebastidae, Bythitidae, or Serranidae (Fig C in S1 Text) has been predicted as Threatened. Additionally, some cryptobenthic fishes families like Gobiidae, Gobiesocidae, Blennidae also host an important number of species predicted as Threatened (S3 Fig C in S1 Text). Cryptobenthic fishes fulfill crucial ecological roles, particularly in the dynamics of trophic interactions and the overall functioning of reef ecosystems [39]. Due to their elusive behavior and reliance on specific habitats that restrict the assessment of their populations, certain species within these families may be undergoing a silent extinction process, underscoring the urgent need for increased evaluation efforts on these species.

Although we found that closely related species were more similar in their IUCN categories than distantly related species (for Threatened, Non-Threatened, and DDNE), we found that taxonomy (family and genus) was not the best predictor of IUCN categories. Rather, species attributes (being relatively common in vulnerability assessment), geographical range, body size [4], and growth rate [40] were much better predictors. Note that species traits indirectly include phylogenetic information, which might reduce the importance of taxonomy in our models. This result highlights which ecological species attributes should be assessed in priority to enhance our ability to accurately predict and detect threatened species. It can also determine which species should be assessed in priority by IUCN experts as a precautionary principle: fishes with small geographical range (already well used as a criterion in the IUCN assessments), large body size, and slow growth rate (known to be correlated [36]).

By mapping the distribution of the predicted species, we provide 2 crucial pieces of information for future evaluation: the hotspots of predicted Threatened species where conservation effort should increase, and the hotspots of DDNE species where research effort should increase. After IUCN status predictions, Threatened species predominantly occurred in the tropics, peaking in Indonesian islands, West-Australia, and in the China Sea, as well as in the west coast of America. For these regions, the establishment or reinforcement of effective marine protected areas should be prioritized, along with increased research effort. Conversely, the gain in Threatened species after the prediction was lower in the Caribbean Sea. This could be explained by a higher research effort [41] in this part of the world leading to better classification of IUCN fish status. Despite being recognized worldwide as a hotspot of diversity, we found that the coral triangle was also a hotspot of DDNE after IUCN status predictions. Since the most important changes in sea surface temperature are occurring in this part of the world [42], the risk of species extinction here is particularly high and the status of these remaining DDNE species should be prioritized. The China Sea also requires a particular effort to provide new information on species to assess their extinction risk.

Because the IUCN Red List is an instrument for conservation planning, management, monitoring, and decision-making [7], we expected that target achievement would be higher for Threatened species. Meanwhile, DDNE species are typically overlooked in conservation planning [19], with the implicit assumption that extinction risk for DDNE and Non-Threatened species is similar [43]. By reducing the number of DDNE species and increasing the number of Threatened species, we show that Threatened marine fish species generally reach low conservation target achievement and are poorly covered by current protected areas. This strongly contrasts with the higher level of target achievement observed for Non-Threatened species (Fig 5B).

We also examined the extent to which inclusion of predicted Threatened species affected the spatial prioritization to conserve worldwide marine fish diversity. Since the prioritization algorithm is strongly influenced by the number of species [44], we found that the ranking of the richest regions was marginally modified. However, we found that low- to middle-rank regions were increasing in conservation priority, revealing the importance of protecting subpolar, polar, and Pacific Island areas as well (Fig 7). Specifically, a strong shift in conservation priority was observed in the subpolar and polar regions of the Southern Hemisphere. Since the Antarctic region is typically not subjected to many global agreements (such as the Convention on Biological Diversity’s (CBD) Aichi Targets of the Strategic Plan for Biodiversity 2011–2020), our results advocate for a deeper evaluation of the conservation status of marine fish species in this region. The strong velocity of isotherm and species range shifts due to climate change observed in these cold waters [45] also poses a significant challenge to the success of ambitious conservation strategies. Some strong changes in prioritization were also observed close to the Pacific Islands. Given that only 13% of marine island areas are currently designated as protected, and that half of all islands lack any protected areas [46], it is likely that fish species in these areas face even greater threats than what our framework predicts. This highlights the urgent need for a significant risk assessment by the IUCN of fish fauna occurring close to islands.

IUCN will increase its efforts in the next decade to complete the extinction risk assessment for many taxa, but there will still be millions of other species to assess, which is simply not feasible given the IUCN standards. Also, paradoxically, the highly publicized annual update of the IUCN Red List brings to the public biased information on the state of biodiversity with a much greater emphasis on few taxa, such as vertebrates [47]. Consequently, whatever efforts the IUCN puts into assessing species from other taxa and communicating about the inherent biases of the Red List, it is now essential to develop a pragmatic approach to extend extinction risk assessments towards overlooked taxa. This means bringing some in silico assessments into the IUCN procedure. As illustrated here, combining large-scale data sets into a multi machine learning framework allows to at once provide reliable extinction risk status for species not evaluated by the IUCN, and point out which species attributes and geographic regions should be assessed in priority to increase the accuracy of the modeling approach and predict status for still unpredictable species. Understanding all the steps associated with this in silico assessment of extinction threats for many different taxa (see for example, Borgelt and colleagues) [33] will also provide a more comprehensive understanding of species conservation status [4]. Such an integrated strategy will improve prioritizing efforts as well as allocating resources effectively to mitigate extinction risk globally [4]. We also advocate for the IUCN to integrate recent developments in forecasting species extinction risks (including our approach) into a synthetic new index of “predicted IUCN status” that could complement the actual “measured IUCN status.” This change would help provide the scientific community with more data on species extinction risks. In addition, governments and the broader public would have their attention brought to a more balanced taxa perception of the ongoing biodiversity crisis.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002773

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.