Examining the co-occurrences of human threats within terrestrial protected areas

Abstract

Human threats to biodiversity are prevalent within protected areas (PAs), undermining their effectiveness in halting biodiversity loss. Certain threats tend to co-occur, resulting in amplified cumulative impact through synergistic effects. However, it remains unclear which threats are related the most. We analyzed a dataset of 71 human threats in 18 013 terrestrial PAs of the European Union's Natura 2000 network, using a Joint Species Distribution Modelling approach, to assess the threats' co-occurrence patterns and potential drivers. Overall, threats were more frequently correlated positively than negatively. Threats related to agriculture and urbanization were correlated strongly with most other threats. Approximately 70% of the variance in our model was explained by country-specific factors, indicating the importance of local drivers. Minimizing the negative impact of key threats can likely reduce the impact of related threats. However, more research is needed to understand better the relationships among threats and, importantly, their combined impact on biodiversity.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13280-023-01966-6.

Introduction

Human activities threaten the world's biodiversity, causing what has been described as the Sixth Mass Extinction (Cowie et al. 2022). Protected areas (PAs) represent one of the most important strategies for mitigating human threats (Watson et al. 2014; Rodrigues and Cazalis 2020). Currently, terrestrial PAs account for approximately 17% of the Earth's land area and inland waters, in line with the percentage goal of the Aichi Target 11 of the Convention on Biological Diversity (CBD) set in 2010 (CBD 2011). In December 2022, the Parties to the CBD agreed to increase the percentage of global protected land to 30% by 2030. However, increasing PA coverage alone is unlikely to be sufficient to halt global biodiversity loss. Recent studies have shown that many of the world's PAs remain ineffective (Geldmann et al. 2019; Elleason et al. 2021) and that over a third of the protected land globally is under intense human pressure (Jones et al. 2018). To improve the effectiveness of PAs, we need better management (Wauchope et al. 2022) and, subsequently, a greater understanding of the potential relationships between human threats within the PAs and their cumulative effects (Geary et al. 2019; Geldmann 2023).

A key reason PAs have been described as lacking effectiveness is the number of human threats occurring within their borders (Tsiafouli et al. 2013; Jones et al. 2018; Schulze et al. 2018). Globally, hunting and recreational activities are the most reported threats inside PAs (Schulze et al. 2018). Deforestation within PAs is also ubiquitous, albeit with important regional differences (Leberger et al. 2020; Wade et al. 2020). Light and noise pollution is widespread within PAs in Europe and North America (Buxton et al. 2017; Mu et al. 2021). Importantly, human threats do not occur in isolation, and the presence of some threats exacerbates the presence and impact of others, causing cumulative environmental effects (Côté et al. 2016). For example, roads have been shown to increase the hunting pressure on wildlife and the spread of invasive species (Mortensen et al. 2009; Espinosa et al. 2014).

To improve the management and effectiveness of PAs, we need to understand better the relationships between the human threats occurring within their borders (Schulze et al. 2018), and how those threats may interact with one another, generating cumulative environmental effects (Geary et al. 2019). Furthermore, we must distinguish between threats that co-occur due to common drivers and those that are closely associated, i.e., where the presence of one influences the likelihood of another. An example of the former would be the co-occurrence of hunting and human intrusions (e.g., outdoor sports) due to easier access to the site or higher nearby human population density. In this case, the two threats do not necessarily interact with one another but are rather influenced by a third factor. This differs from roads directly influencing threats such as hunting, as mentioned above.

Despite a growing literature on human threats within PAs (Schulze et al. 2018), important knowledge gaps remain regarding the relationships between threats. First, most studies examining the presence of human threats within PAs do not explore potential relationships but rather summarize, often using descriptive statistics, the presence of various threats (Tsiafouli et al. 2013; Mazaris et al. 2019). Second, the few studies that have explored potential relationships between threats were conducted at small spatial scales, such as specific PAs (Godínez-Gómez et al. 2020) or a single country (Watson et al. 2015; Williams et al. 2022). Although the above studies are valuable, they do not allow widely generalizable inferences on the possible relationships between threats. Studies conducted at larger spatial scales typically rely on data from different sources, which increases noise and uncertainty, or on data derived using remote sensing techniques, such as satellite imagery (Geldmann et al. 2014, 2019; Jones et al. 2018; Schulze et al. 2018; Bowler et al. 2020). While remote sensing data are becoming increasingly accurate and useful for documenting human pressure on biodiversity, they can miss important human threats, which currently can be only surveyed in situ, such as the spread of invasive species (Peres et al. 2016; Hulme 2018; Mammides et al. 2022). Consequently, we still have limited knowledge about the relationships between multiple human threats and their co-occurrence patterns within PAs. Indeed, to date, we are not aware of any studies that investigated the associations between human threats co-occurring within PAs on a large scale.

The European Union (EU) is home to the largest coordinated network of PAs worldwide, the Natura 2000 network. The network includes over 24,000 terrestrial sites, which cover more than 18% of the terrestrial area of the EU's 27 Member States and nine biogeographical regions (European Commission 2022). Although PAs in the Natura 2000 network can technically fall into any IUCN protected area category, ranging from strict nature reserves to PAs with sustainable use of natural resources (Dudley 2008), only about 4% of the PAs are currently under strict protection (IUCN categories Ia or Ib), as the network was not designed to be “a system of strict nature reserves from which all human activities would be excluded” but rather “managed in a sustainable manner, both ecologically and economically” (European Commission 2013). As part of the management requirements, each Member State is obliged to monitor and report every six years on all Natura 2000 PAs within their borders. The information is collated by the EU and made available online as a unified Natura 2000 database. Each PA is surveyed in situ by local experts (selected by each Member State) using standardized protocols. Among other information reported is the presence of human threats based on a list of 75 types of threats (Table S1). Importantly, the threats monitored include several key pressures that are rarely examined in large-scale studies, such as those related to hunting, invasive species, and pollution (Table S1). Albeit ubiquitous within PAs (Tsiafouli et al. 2013; Schulze et al. 2018), such threats often need to be surveyed in situ and cannot be detected via remote sensing technology on which most studies are based (Schulze et al. 2018; Mammides et al. 2022).

Here, we analyzed this unique dataset to study human threats in the Natura 2000 network and their co-occurrence patterns. Previous studies have proposed the use of network analysis to study the associations between human threats and their connection to a variable of interest (e.g., a focal species, Geary et al. 2019). However, when non-experimental data are used, this approach does not allow one to distinguish between the threats that co-occur due to common drivers as opposed to those that directly influence one another. To address this issue, we used a Joint Species Distribution Modeling (JSDM) approach—used commonly in ecological studies to study species co-occurrence patterns (Ovaskainen et al. 2017, see methods)—to incorporate external variables that might be important common drivers of the associations among threats, such as land-use cover, topography, population density, site accessibility (González-García et al. 2022), area of each PA (Guan et al. 2021), and presence of a management plan (Bowler et al. 2020). This novel approach allows us to account for the influence of these factors on the probability of threats’ co-occurrence, while at the same time assessing the importance of these variables in driving the presence of individual threats. Finally, we discuss the approach's limitations and the available co-occurrence data and provide some recommendations for understanding and improving the management of threats in PAs for future research.

Materials and methods

Data collection

We downloaded the Natura 2000 database (end2020 version) made available by the European Environment Agency (EEA) at www.eea.europa.eu/data-and-maps/data/natura. The database contains the spatial boundaries of each Natura 2000 site in the EU, as well as a series of Microsoft Excel files containing the information that Member States must report to the EU every six years according to Articles 12 and 17 of the Birds (2009/147/EC) and the Habitats (92/43/ECC) Directives, respectively. This information includes data on the human threats, pressures, and activities within each PA. Specifically, for each recorded activity, Member States must specify whether: (a) it has a positive or negative impact on the Natura 2000 site; (b) whether its impact is considered "low", "medium", or "high", and (c) whether the activity has been observed inside the site, outside, or both. Generally, threats, pressures, and activities are recorded using a hierarchical standardized list (https://cdr.eionet.europa.eu/help/natura2000) following international conventions (Salafsky et al. 2008). The first level consists of 13 broad categories (Table S1), the second level consists of 75 subcategories (Table S1), and the third level consists of even more detailed activities. Following previous studies (Tsiafouli et al. 2013; Mazaris et al. 2019), we focused on the list of threats and activities reported at the second level for which the information was most detailed and complete; threats recorded at the first level are too broad, while those listed at the third level are not consistently surveyed and recorded across all PAs. Given our focus on terrestrial threats, we removed marine threats from the dataset (four threats).

We only used threats recorded inside the Natura 2000 PAs since areas outside the boundaries of the PAs are not surveyed as extensively and consistently as inside. We also analyzed only threats listed as having a negative impact since we were mostly interested in the potential relationships between threats that negatively impact biodiversity. We used all negative threats regardless of their intensity level because of the subjectivity involved in such assessments, driven by the data's continental scale and the fact the assessments are made by numerous experts over multiple years. Using the resulting dataset, we built a presence/absence matrix representing the threats within each PA in our dataset.

Using the shapefiles of the Natura 2000 network and the terrestrial boundaries of the EU (https://ec.europa.eu/eurostat/web/gisco/geodata), we removed all marine PAs (since our focus was on terrestrial human pressures) and the parts of the coastal sites that extended into the sea. The latter was necessary to avoid overestimating the total terrestrial area of those sites. Using the resulting shapefile and the "Zonal Statistics as a Table" tool in ArcGIS Pro 2.9, we estimated the mean human population densities within each PA using the LandScan dataset (Rose et al. 2021) available at a resolution of 1 km². We also estimated the terrain ruggedness of each PA, using the topographical data provided by Amatulli et al. (2018), also at a resolution of 1 km². We calculated the accessibility of each PA using the data by Weiss et al. (2018), which represents the number of minutes needed to reach the PA from the nearest urban center based on the transportation network available in each Member State and on other determining factors (e.g., topography). To identify each PA's biogeographical region, we used the spatial boundaries of the regions available by the EEA at https://www.eea.europa.eu/data-and-maps/data/biogeographical-regions-europe-3. Using the most recent CORINE Land Cover map of 2018 (https://land.copernicus.eu/pan-european/corine-land-cover), we estimated the percentage of each PA covered by each of the following land cover classes: artificial surfaces, agricultural lands, semi-natural vegetation, forests, wetlands, and waterbodies. Lastly, we recorded whether each PA had a management plan (0 = no plan; 1 = plan available) using the corresponding information in the Natura 2000 database.

Data analysis

We used a JSDM approach to model the co-occurrence of human threats within the Natura 2000 network (Ovaskainen et al. 2016). This method is commonly used to model the distribution of species according to a series of environmental factors and based on incidence or abundance community data (e.g., Odriozola et al. 2021). Using a hierarchical generalized mixed model framework, JSDM allows specifying a model structure with fixed and random effects to account for the effects of environmental covariates as well as the spatial structure of the data on the associations between 'species' (i.e., residual co-occurrence associations). In our case, JSDM allows us to detect potential relationships among co-occurring threats while accounting for the effects of external factors, suggesting that residual correlations represent associations in which the threats directly influence each other's probability of occurrence, even if the causal direction of the relationship cannot be determined with the data used in this study (i.e., if threat A causes B, or whether B causes A). At the same time, it also allows the measurement of the influence of external factors on the probability of the occurrence of each threat separately, providing additional information for managing the threats. Since our data structure resembles that of a "species X site" incidence matrix, we used Hierarchical Modeling of Species Communities (HMSC, Ovaskainen et al. 2017) and the HMSC 3.0 package in R to conduct the analysis (Tikhonov et al. 2020; R Core Team 2023).

Before running the model, we removed all PAs with an area < 0.1 km² (n = 2819), which given the resolution of other data used (e.g., topographical factors), are likely to be too small to be relied upon for the goals of our analysis. We also removed all sites in which threats were surveyed prior to 2012 (n = 1268) to reduce temporal mismatches with the rest of data used in the analysis and because in 2011 there was a change in the EU methodology used to report threats within the Natura 2000 network.

We used the incidence matrix of the threats as our response variable. Given the presence-absence nature of the data, we modeled threats with a probit-link function. As environmental and anthropogenic covariates, we included each site's: area (km²), human population density (individuals km⁻²), terrain ruggedness index (meters), accessibility from the closest urban center (minutes), whether there was a management plan in place (0/1), and proportion classified as 'artificial', 'agriculture', 'semi-natural', 'waterbodies', 'wetlands'. We initially also included the proportion of 'forest' area, but we removed it because it was strongly correlated with the proportion of land classified as 'agriculture' (r = − 0.59), causing collinearity in the model. We retained 'agriculture' instead of 'forest' because the former has a generally negative impact on biodiversity (Henle et al. 2008), while forest cover typically relates positively to biodiversity (Estavillo et al. 2013), and our focus was on the negative impacts. We decided to use these variables despite a potential concern about circularity between the response variable and some predictors (e.g., agricultural land and the presence of threats related to agriculture) because not all land cover classes (e.g., agriculture as captured by the CORINE Land Cover map) necessarily represent a threat to biodiversity (e.g., agriculture threats as captured by in the Natura 2000 database). We further included the longitude and latitude of the centroid of each site, obtained using the “Calculate Geometry” tool in ArcGIS Pro, as fixed effects to account for possible spatial relationships. We used the biogeographical region (hereafter, bioregion) and the country in which each site was located as random effects to control for both within-bioregion and within-country variation due to common regional characteristics not captured by the model.

We built our model using three Markov chain Monte Carlo (MCMC) chains, each chain consisting of 150 000 iterations, with the first 50 000 discarded as burn-in, and the other 100 000 thinned every 100th sample to give a total of 1000 samples per chain, and 3000 posterior samples in total. These values are in line with other studies employing the same analytical method (e.g., Odriozola et al. 2021). We evaluated the performance of the model by Tjur-R² (Tjur 2009) and area under the curve (AUC), and we measured the amount of variance in the residual co-occurrences association that was explained by fixed and random effects of the model. Due to the continental scale of the data and the fact that they were collected by many different observers over many years, we used a conservative 99% posterior probability to consider the relationships between pairs of threats as well supported. We applied the same support level for the relationships between threats and the modeled environmental and anthropogenic variables.

Results

Our final dataset included 18 013 PAs and 71 threats distributed across nine bioregions and 25 EU countries (Malta and Estonia were excluded from the analysis following the data preparation steps described in the methods; Fig. 1). The Continental bioregion hosted the highest number of PAs (n = 6941), while the Black Sea bioregion the lowest (n = 52). The Black Sea bioregion reported the most threats per PA (696 total threats, mean = 13.38, median = 13 per PA), while the lowest number of threats per PA were reported in the Boreal bioregion (n = 11 601, mean = 3.05, median = 2 per PA; Figure S1). Belgium reported the most threats per PA (n = 4059, mean = 12.72, median = 12), and Finland the least (n = 3054, mean = 2.54, median = 2; Figure S2). The models had mean values of Tjur-R² = 0.08 and AUC = 0.82 (Figure S3). The random effects explained most of the variance of the model (country = 70.2, range 32–93; bioregion = 8.3, range 0.5–34), while the fixed environmental and anthropogenic variables explained only a small portion of the variance (Fig. 2). However, the proportion of variance explained by the random and fixed effects varied considerably among threats (Fig. 2).

Fig. 1.

Map of the 18 013 terrestrial Natura 2000 sites (PAs) used in the study, shown in dark green color. Also highlighted the distribution of the biogeographical regions within the European Union (EU) Member States. United Kingdom (UK) is not included because the data used are from the version 2020 of the Natura 2000 dataset, which was submitted after the UK left the EU

Fig. 2.

Variance partition derived from JSDM (Joint Species Distribution Modelling) of the threats co-occurrence matrix. The values of each bar indicate the amount of variance explained by each group of fixed and random effects for each threat (code in the x-axis; A = Agriculture, B = Sylviculture, forestry, C = Mining, extraction of materials and energy production, D = Transportation and service corridors, E = Urbanization, residential and commercial development, F = Biological resource use other than agriculture & forestry. G = Human intrusions and disturbances, H = Pollution. I = Invasive, other problematic species and genes, J = Natural system modifications, K = Natural biotic and abiotic processes, L = Geological events, natural catastrophes, M = Climate change. Refer to Table S1 for meaning of the individual codes). On top of the figure are the mean values among the 71 threats assessed here, for each fixed or random effect. 'Coordinates' shows the variance explained by longitude and latitude as a combined value (dark green color), while 'Land Use' incorporates the combined effects of the five land uses assessed here: artificial, agriculture, semi-natural, waterbodies, wetlands (red color)

Of the 5,041 possible pairwise correlations, 30.5% were positive and supported with at least 99% posterior probability, while 0.8% were negative and supported, and 68.7% were not supported (Fig. 3, Figure S4, Supplementary file 1). To facilitate the interpretation of the results, we summarized the occurrences of positive correlations for each first-level threat with all other first-level threats (Table S1), and we further removed all correlations with a value < 0.7 to only account for the strongest relationships (Fig. 4). Agricultural threats were the most widespread, reported in 53% of all PAs and were also more often positively related with other threats. Conversely, threats related to pollution and invasive species were rarely related with other threats (Fig. 4). The number of supported and strong (i.e., r > 0.7) positive correlations did not appear to be related to how often a threat (or threat category) was reported (Fig. 4). For instance, threats related to pollution and urbanization were recorded in 24% of the sites, but urbanization was positively related with many other threats compared to pollution (109 and 12, respectively; Fig. 4).

Fig. 3.

Residual co-occurrences among all measured threats. Only co-occurrences with at least 99% posterior probability of are shown in red (positive) or blue (negative) color, while those non-supported are in white. The shade of the color indicates the strength of the correlation between each pair of threats, with darker colors indicating stronger correlations. Included on the outer part of the figure are the 13 first-level threats to which the 71 second-level threats belong. See Table S1 for the meaning of individual second-level threats, and Supplementary file 1 for the value of each correlation

Fig. 4.

Count of positive co-occurrences between each first-level category of human threats. The y-axis shows the total number of positively supported co-occurrences at 99% posterior probability and with a correlation value > 0.7. The percentage on top of each bar indicates the percentage of sites (PAs) where at least one second-level threat of that category was reported (out of the 18 013 Natura 2000 sites analyzed in this study). The lack of correspondence between these numbers and the bar below them demonstrates that the most frequent threats were not always those that were most associated with other threats

With regards to the environmental and anthropogenic drivers, the total area of the PA was positively related to the occurrence of 75% of the threats at 99% posterior probability, meaning that most threats were more likely to be present in larger PAs (Fig. 5, Table S2). The proportion of area classified as agricultural land was positively related to all the agricultural threats but negatively related to forestry and sylvicultural threats (Fig. 5). The rest of the variables showed both positive and negative effects on the occurrence of human threats (Fig. 5).

Fig. 5.

Threats responses to environmental covariates with at least 99% posterior probability of being positive (red) or negative (blue) in our model. The name of the threats is in the y-axis (see Table S1 for acronyms' meaning), while the name of the covariates is in the x-axis

Discussion

We present the first large-scale analysis of co-occurrence patterns of human threats within terrestrial PAs, and of their responses to environmental and anthropogenic variables, stemming from data collected through field surveys. Several key findings emerge from our study. First, many of the threats recorded within the Natura 2000 network are positively rather than negatively related with each other (Fig. 3). Second, agricultural and urbanization threats are more frequently related with other threats (Fig. 4). Third, most of the variation explained by our model was due to the random effect ‘country’ (Fig. 2), pointing to considerable differences among the EU Member States. Lastly, although we used the most comprehensive and homogeneous dataset on human threats available worldwide, we still found it challenging to reach robust conclusions in some instances. More efforts are needed to improve the quality and distribution of threat data in the future, which would allow the researchers to better understand their co-occurrence patterns.

Our findings advance the current understanding of the co-occurrences of threats within terrestrial PAs. In many cases, the results confirmed expected and previously known relationships, adding confidence to the validity of our analysis. For example, we found that the presence of 'roads, paths and railroads' is strongly related with 'hunting and collection of wild animals', and we can reasonably assume that the first is likely driving the second (Espinosa et al. 2014). It follows that reducing the occurrence of roads is likely to reduce the impact of hunting. However, the direction of the relationship between other threats is not always immediately obvious. For example, 'forest and plantation management & use' is positively related with 'urbanized areas, human habitation', but which drives the other is likely to depend on local conditions. Yet, knowing that two threats are related can highlight areas where interactions between threats may produce greater consequences than what each would do individually (i.e., cumulative environmental effects) (Côté et al. 2016). This information can help improve PA management, forecast potential combined impacts, and direct future research (Driscoll et al. 2021).

Previous studies have mentioned that threats related to agriculture are the most widespread within the Natura 2000 sites (European Environment Agency 2020). Here, we also show that agriculture-related threats are often related to other threats. For instance, 'cultivation' is strongly related with 'hunting and collection of wild animals', suggesting that open areas of agricultural land are related with increased hunting pressure. Interestingly, we found no clear relationship between the number of sites in which a threat was recorded and the frequency with which the specific threat is related with others (Fig. 4). This suggests that the co-occurrence patterns we report are not determined by the frequency of the threats, lending credence to the hypothesis that some pressures are more likely than others to exacerbate other threats (Baldwin 2010). For example, threats linked to 'pollution' and 'urbanization' are recorded in approximately the same number of sites, but urbanization is positively related with many more threats than pollution. Therefore, urbanization might be expected to drive the presence of other threats (Elmqvist et al. 2013), while pollution is less likely to. Based on these findings, we suggest that the management of agricultural and urbanization threats should be prioritized. Yet, we note that the source of some threats might occur outside the boundaries of the PA, and thus be out of the control of the PA managers, with limited possibilities to eliminate or reduce their impacts.

The proportion of the PAs under agricultural use was positively related to the occurrence of all agricultural threats, which represent a major threat to biodiversity (Tilman et al. 2017) and are likely to worsen due to continuous expansion of cropland within PAs (Meng et al. 2023). About 38% of threats were positively related to the accessibility of the PAs (i.e., travel time to major cities), supporting previous findings that more remote areas suffer less human pressure (Gallardo et al. 2017; Schulze et al. 2018; Guan et al. 2021) and suggesting that greater efforts are needed to protect and manage PAs closer to large cities from where many threats may originate. Globally, the reasons for variation in the distribution of human threats between countries or regions are frequently linked to socioeconomic factors such as levels of human development (Schulze et al. 2018; Geldmann et al. 2019). Here, the random effect 'country' explained 70% of the variation in our model, implying that it is critical to understand the effects of the missing covariates in our models on the occurrence of threats within each EU Member State in order to develop effective management plans. Variables such as population density, area of the PA, accessibility, and terrain ruggedness only explained a small part of the variance, even though they were previously reported to be relevant drivers of human pressure in PAs and to negatively affect biodiversity (Luck 2007; Mcdonald et al. 2008; Schulze et al. 2018; Mammides 2020). This may be due to other uncontrolled factors that can be important at the country level, such as the number of hunters and other socioeconomic factors (Massei et al. 2015), as well as possible differences in data collection among Member States. The presence of a management plan also showed limited explanatory power, possibly because it is a simple binary variable (0/1) and, perhaps more importantly, because its presence does not necessarily inform on whether and to what extent the plan is implemented. Overall, our findings support previous studies noting significant country-level variation in human threats and associated processes (Tsiafouli et al. 2013).

The data that we used, and our results, present some limitations. Notably, knowing that two threats are strongly related does not always allow us to determine whether the presence of one is driving the presence of the other (Blanchet et al. 2020). This is a limitation associated with correlative approaches, such as JSDM, which are based on non-experimental co-occurrence data. Moreover, occurrence data are valuable, but they do not offer insights into the intensity of the impact of the threat. As a result, two PAs where the same threats are present but have significantly different impacts will be scored equally. However, as field data at the scale of this study are inevitably collected by many different observers, the use of occurrence data could also reduce the uncertainty caused by differences among observers compared to the use of quantitative estimates or levels of intensity (Brown and Williams 2016). Still, inter-observer biases may also be one of the reasons for the moderate explanatory power of our models (mean Tjur-R² = 0.08), even though we included many of the variables known to influence human threats (Schulze et al. 2018; Guan et al. 2021). Similarly, the predominant effect of the random effect 'country' may be partly due to possible differences in the reporting practices among Member States. However, the standardization in the protocols and guidelines for reporting the threats minimizes this concern to the best extent possible. Missing confounding variables could also be a reason why some of the supported relationships between threats that we found are difficult to explain. For instance, the strongest supported relationship (r = 0.994) was between "cultivation" and "sport and leisure structures" (Supplementary file 1). Including other socioeconomic factors, such as indices of economic growth or, as previously mentioned, of human development, might help to remove those unlikely relationships. However, no such data is available at the level of each PA, as required by our analysis. Moreover, we note that data are not updated as often as they are expected, and some Member States did not update data on specific PAs for over ten years. This causes temporal mismatches that further increase the noise in the data. Therefore, we support the call for growing coordinated efforts within the Natura 2000 network (Hermoso et al. 2022), including the timely reporting of data to improve the management of PAs.

Conclusions and recommendations

Our study supports the idea that some human threats co-occur more often than others (Geary et al. 2019). Further, our JSDM approach helps identify potential links between threats, indicating that one threat's presence may increase the chances of another, instead of them merely co-occurring due to shared drivers (which were included as covariates in the model). We found that some threats, such as agriculture and urbanization, are disproportionately related with other threats. We also found that many patterns are country-specific, suggesting that management actions need to be tailored to the countries' specific conditions.

To our knowledge, this is the first time that the co-occurrence patterns of threats within PAs have been analyzed on a continental scale and based on data collected in the field under a standardized protocol rather than remotely sensed data. This is a valuable step in better understanding how human threats are interconnected. Therefore, minimizing threats that are strongly related with others should be prioritized. For instance, reducing the extent or the expansion of 'Urbanised areas, human habitation' is likely to also reduce related threats such as 'Interspecific faunal relations', 'Fire and fire suppression', 'Introduced genetic material, GMO', and 'Taking/removal of terrestrial plants', among others (Supplementary file 1).

There are still some limitations that, if addressed, will improve the utility of these results and help make PAs more effective (Geldmann et al. 2019). A quantitative reporting of human threats, instead of presence-absence or broad levels of intensity, would allow a more accurate analysis of the associations between them. For example, estimating the proportion of the PAs affected by each threat could provide valuable information. However, we acknowledge that increased funding and resources would be required to implement such measures and that funding is reportedly limited even today (Hermoso et al. 2022). Future studies should attempt to adopt more quantitative modeling approaches, such as the JSDM used here, which allows modeling threat data with relevant covariates in a comprehensive manner. We hope that our study will inspire future researchers to move beyond documenting individual threats, and towards understanding associations, which are necessary for effective management and therefore protect biodiversity.

Supplementary Information

Below is the link to the electronic supplementary material.

Biographies

Francesco Martini

is a Postdoctoral Researcher at Trinity College Dublin, Ireland. His research interests include community ecology, ecosystem services, and biodiversity conservation.

Constantinos Kounnamas

is the Head of the Nature Conservation Unit of Frederick University, Cyprus. His research interests relate to biodiversity and natural areas conservation and management.

Eben Goodale

is Professor at Xi’an Jiaotong-Liverpool University, China. His interests include behavioral and community ecology and conservation biology.

Christos Mammides

is a Senior Researcher at the Nature Conservation Unit of Frederick University, Cyprus. He researches topics related to biodiversity conservation.

Author contributions

FM: Conceptualization, Data Curation, Formal analysis, Investigation, Methodology, Visualization, Writing—original draft, Writing—review and editing. CK: Investigation, Visualization, Writing—review and editing. EG: Investigation, Visualization, Writing—review and editing. CM: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing—original draft, Writing—review and editing.

Data availability

All data used in this study are publicly available. The aggregated dataset used to run the Joint Species Distribution Model is available in Figshare: 10.6084/m9.figshare.21594387.