Conserving saproxylic flagship species by complementing 150 years of natural history with citizen science data—the case of the stag beetles (Lucanidae, Coleoptera) of Portugal

Lucanus cervus is the largest (ranging from 22 to 85 mm (López-Cólon 2000)), the best-studied and most broadly distributed lucanid in Europe. This species was first described in ancient Greek texts (Sprecher-Uebersax 2008) and features prominently in local stories across its wide European distribution (Harvey and Gange 2011). Its iconographic value and high recognizability, paired with its ecological significance and the fact that, at least in Portugal, the species represents a cherished natural heritage symbol, make it an ideal flagship species that inspires and motivates people to participate in conservation actions (Campanaro et al. 2016; Bardiani et al. 2017). The species is known to have a life cycle of 3–7 years, with its larvae feeding and developing predominantly in deadwood of broadleaf trees (preferably white oaks such as Quercus orocantabrica Rivas Mart. & al. and Q. pyrenaica Willd.) that are in contact with soil, like dead roots, stumps or fallen logs. The adults can be found from May to July, mostly at dusk. Although it has been relatively well studied compared to other invertebrates, key aspects of its biology are still misunderstood, which hinders the identification and quantification of its historical and current threats needed to create effective conservation actions (Méndez and Thomaes 2021). Lucanus cervus is under the protection of the Habitats Directive (Annex II) and the Bern Convention (Annex III) and is classified as Least Concern in the Portuguese Red List (Boieiro et al. 2023) and as Near Threatened in the European Red List (Calix et al. 2019).

L. barbarossa (28–40 mm), like L. cervus, uses roots and stumps from Q. pyrenaica, Q. faginea Lam. or Q. canariensis Willd. The adults can be found at night from July to September and are only found in the Iberian Peninsula and border countries (Morocco and South of France), having little published information on its ecology and conservation status. Dorcus parallelipipedus (15–36 mm) is a more generalist species, occurring in more decayed wood (stumps, logs, branches and roots) of a wide range of tree species. The adults can be found during the day and night almost throughout the whole year, with its phenology peak between April and September. It is distributed across Europe, being the most conspicuous of the four species. Platycerus spinifer (9–15 mm), the smallest of the species, feeds of deadwood of large shrubs such as Cytisus spp. or Genista spp. and is more likely to be found in higher altitudes during the days between April and June (López-Cólon 2000). It is almost endemic to the Iberian Peninsula, having recent records in the southern border of France (Brustel and Van Meer, 2020) and not much is published on its ecology and conservation status.

Citizen scientists were called upon to collaborate with the project by reporting opportunistically collected observations. From 2016 to 2019, each observation involved the submission of a simple form with relevant ecological, phenological and spatial information, including the date of the sighting, the geographic coordinates, habitat, number and sex of the individuals observed and also required the upload of a picture of the observed specimen for validation purposes. In 2020, it was decided to adopt iNaturalist–an international citizen science project aiming to collect biodiversity data from the California Academy of Sciences and the National Geographic Society–as the project's recording platform. Using the iNaturalist platform, only the date of the sighting, species, geographic coordinates, and pictures were submitted since this platform does not allow the remaining information to be submitted for each sighting.

As communication is key to the success of citizen science (Constant and Roberts 2017), the project was anchored on a strong component of environmental education. A strong presence on social networks and meetings with stakeholders and potential collaborators were part of the communication strategy. In 2019, a new engagement method was created: the national network of ambassadors of the project, which involved, at its peak, more than 50 naturalists or countrywide scattered institutions that were responsible for the implementation of engagement activities within their local communities.

Historical information on the distribution of lucanid species in Portugal was compiled and analysed to evaluate the added value of this effort. Data were gathered from both published and unpublished data sources before 2016 and included (i) scientific literature (e.g., books, scientific papers and technical reports), (ii) online biodiversity databases, such as Naturdata (https://​naturdata.​com), (iii) institutional and private entomological collections, and (iv) unpublished data from entomologists and ongoing projects. Records were included in a dataset if the following criteria were met: (1) the original publication was fully known; (2) there were no duplicate records from entomological collections previously reported in a publication or that were consequently cited in historical publications; (3) the date (at minimum, the year of recording) and/or site location (locality or coordinates) were validated. When the locality was identified, we used the county- or parish-level centroid point as geographical coordinates.

If none of the criteria were met, then the record was eliminated and was not used further in the analysis. The resultant non-citizen science dataset was therefore standardised, including the data source, the record date and the geographical coordinates (when locality was used for geographical reference, the coordinates of the central point of the locality was used (Supp. Information—SI 1). The data sources were divided into: (a) Published data (which included validated entomological collections), and, (b) Unpublished data from taxonomists, other experts or non-published entomological collections.

All the observations reported by citizens were validated by more than one expert between 2016 and 2019 by the VACALOURA.pt’s team and between 2020 and 2023 by the iNaturalist community. The data were downloaded from iNaturalist on September 19th, 2023, standardised and stored (Supp. Information—SI 1). Whenever essential information was missing or was incorrectly entered in the online form, the participants were contacted to clarify and curate the erroneous records. The iNaturalist validation method follows their Data Quality Assessment, which is a “summary of an observation's accuracy, completeness, and suitability for sharing with data partners”. The building block of iNaturalist is the verifiable observation. A verifiable observation is an observation that: (1) has a date; (2) is georeferenced (i.e. has latitude/longitude); (3) has photos and/or audio recordings, and (4) is not of a captive or cultivated organism. Observations become “Research Grade” when the community agrees on the species-level ID or lower, i.e. “when more than two or three of identifiers agree on a taxon” (https://​www.​inaturalist.​org/​pages/​help#general1, September 19th, 2023). Only data validated as research-grade and that had its location public were considered. Following validation, all the records were checked to remove possible duplicates. We have decided not to include data from other data platforms (e.g. GBIF) because of ongoing integrations and links between them, which could increase the risk of including duplicated or falsely validated records. For example, iNaturalist records are automatically added to GBIF once Research Grade validation is reached (which is not always correct). Since iNaturalist was the platform used in the VACALOURA.pt project, we decided to focus on the project's results, conducting, as mentioned, a second validation on all retrieved records from iNaturalist before use. This decision might leave a minor percentage of data from other platforms outside of this paper. However, it was our priority to include all available records with the highest level of validation possible under the “Published” category”.

Spatial information was managed using the software QGIS v-3.28.2 (QGIS Development Team 2023). A 10 × 10 km grid (PT-TM06/ETRS89) was used to update species distribution information at the national level. The establishment of this grid was grounded in the assumption that by possessing the true coordinates of each record, we could project its distribution across various scales.

To model each target species, we used variables on several environmental dimensions that were considered relevant to explain their distribution and ecological constraints at different scales. These include bioclimatic indices, topography/geomorphology, soil, vegetation, and forest structure (Table 1).

As we lacked prior knowledge of how the initial set of variables related to the target species’ environmental niche space, we used a heuristic and model-driven approach to identify the most explanatory predictors by species (Gonçalves et al. 2021; Hespanhol et al. 2022; Regos et al. 2022).

Initially, we ran a set of models that included all available variables and relied exclusively on the Random Forest algorithm. This algorithm is known for handling large numbers of variables and addressing multicollinearity through ‘feature bagging’ (i.e., to train multiple decision trees on randomly selected subsets of input variables to improve model performance) (Biau 2012; Breiman 2001). This preliminary effort was to identify the predictive variables with the strongest predictive power. Variable importance was based on the R biomod2 package (Thuiller et al. 2023), which calculates Pearson’s correlation between reference predictions and predictions for a ‘randomized’ version of each variable. Variables with higher scores strongly influence model predictions, while zero indicates no influence. We averaged importance scores across all five pseudo-absence sets and ten training rounds to rank each variable. Using multiple sets of pseudo-absences (PAs) is a common practice in ensemble species modelling frameworks like the one implemented in biomod2. Instead of generating a single set of random pseudo-absences, multiple sets are created to insert higher variability in environmental data in pseudo-absence subsets. Models generated by each subset are then trained, evaluated, and ensembled, aiming to improve final predictions by creating a distribution of habitat suitability values for each grid cell used for projections (i.e., merging multiple sets of PAs, algorithms and training rounds).

Next, we conducted an iterative variable selection process based on their importance rank to reduce multicollinearity and increase parsimony. We tested the non-parametric Spearman correlation between each pair of variables, starting with the most important and working down the rankings. If two variables were correlated, we kept the one with a higher importance score. Variables with correlation values below 0.75 were considered for the final set. We ultimately included the top ten variables with the highest importance scores in this reduced set, which were used to develop the final models (Supp. Information—SI 2). After this step, some adjustments were made during model calibration to iteratively enhance variable selection and model predictions with a focus on improving model generalization and preventing overfitting.

Species Distribution Models (SDMs) with a 1 × 1 km resolution were developed in the R v-4.3 analysis software (R Core Team 2022) using the biomod2 package (Araújo and New 2007; Thuiller et al. 2023). Biomod2 is a multi-model ensemble forecasting approach to combine predictions from various statistical and machine-learning-based algorithms and explore species-environment relationships.

Models were fitted using nine modelling techniques currently available in biomod2: GLM (Generalised Linear Models), XGBOOST (Extreme Gradient Boosting Models), GAM (Generalised Additive Models), CTA (Classification Tree Analysis), ANN (Artificial Neural Networks), FDA (Flexible Discriminant Analysis); MARS (Multivariate Adaptive Regression Splines), RF (Random Forests) and MAXENT (Maximum Entropy). Default parameters were used except for the smoothing degree term in GAM, which was set to k = 4 to prevent over-fitting issues (Guisan et al. 2002).

As we only had presence data available for the selected species, we randomly generated ten sets of pseudo-absences (PA), each with 10,000 PA points. Despite the considerable number of PAs, the biomod2 package employed in model development allows balancing the training data by attributing equal weights to presences and pseudo-absences points/subsets. Using this approach, we implement a prevalence of 0.5 in model fitting (as recommended by the package developers), which mitigates unbalancing issues in train datasets by assigning equal weights to the presence and pseudo-absence datapoints. Beyond that, using a high number of PAs allows for capturing a more significant amount of variability and heterogeneity in environmental covariates with benefits for model training.

We used a holdout cross-validation method to evaluate the models. This method used 80% of the input records for model fitting and the remaining 20% for model evaluation at each round. A total of ten rounds were performed for model evaluation, plus an additional one containing all the available data. For assessing model performance, the Area Under the Receiver-Operating Curve (AUC), the True-skill Statistic (TSS), and the Sensitivity and Specificity values were calculated (Thuiller et al. 2009). Response curves, following the method described by Elith et al. (2004), were used to plot how each variable influences habitat suitability values and obtain greater insight into each species’ environmental requirements. Further, density plots comparing the observed (i.e., presence) vs region-wide random (i.e., pseudo-absence) distributions of each variable for each species were used to further investigate species niches and their ecological requirements and assess if citizens' observations can encompass the whole range of suitable conditions.

Next, we computed an ensemble consensus model to combine the partial models (n = 990) generated with the different data PA sets and cross-validation subsets. We selected the top 25% percentile best models based on the TSS value. Based on these top-performing models, we computed an ensemble model using the average as the consensus method. Spatialized model predictions were obtained for the entire area using the ensemble model. To ‘binarize’ predictions into suitable/unsuitable habitat, we used the threshold value that maximized the TSS statistic. We evidenced the areas which the models indicate as “Highly Suitable Areas” (HSA) from the remaining “Suitable Areas” (SA). HSA were defined as 10 × 10 km grid cells from which 50% of the containing 1 × 1 km pixels of the model are above the TSS threshold for suitability. SA’s were defined as the remaining 10 × 10 km grid cells with suitable 1 × 1 km pixels.

We calculated niche hypervolumes using the R package hypervolume (Blonder et al. 2014). This method uses a stochastic geometry approach and is a new tool to quantify high-dimensional ecological hypervolumes. Similar to SDMs, we used three main sets of data: (i) previous records for the species from published sources and reference collections (REFS); (ii) occurrence records from the VACALOURA.pt citizen science project (CS), and (iii) all data combined (ALL).

Because the number of variables must be parsimonious when calculating niche hypervolumes, we decided to use only bioclimatic indices highlighted as important in SDMs. A second criterion for niche hypervolume construction relates to comparability. Hence, we decided to use the same variables for all species. From the previous SDM variable importance ranking/selection stage (Supp. Information—SI 2), we included all bioclimatic variables (9 out of 19) for all species. Then, we analysed pairwise correlation to retain only those variables with a Spearman correlation of |r|≤ 0.75. This resulted in the selection of three temperature-related variables (BIO_03 – isothermality, BIO_04 – temperature seasonality and BIO_10 – mean temperature of the warmest quarter) and one precipitation-related variable (BIO_15 – precipitation seasonality coefficient of variation). All bioclimatic variables were previously centred and scaled to avoid differences in hypervolumes due to the ranges of variables.

Niches’ hypervolumes were calculated using the Support Vector Machine (SVM) method. This technique is considered the most appropriate method in high dimensionality analyses and for most modelling applications of realised niches where the limits of the observed data are of most interest (Blonder et al. 2018). We then quantified niche breadth from the calculated hypervolume log size. We compared it for each species and by each of the three datasets (REFS, CS and ALL) to assess differences and potential gains in niche quantification due to CS data increments.

We compiled 7782 records of the four lucanids using a combination of the different data collection strategies (88.8% from citizen science sources). In total, 93.9% of the total records were validated. From the validated ones, 93.7% are from citizen science sources. Furthermore, most of the non-validated records are from non-citizen science sources (86.1%) (Table 2).

The 7306 validated records were unevenly distributed by the four species (Table 3). L. cervus accounted for 59.8% of the validated records, followed by D. parallelipipedus (27.8%), and L. barbarossa (11.8%), while P. spinifer was the least recorded species (0.7%). All species had a higher percentage of validated records from citizen science when compared to non-citizen science sources, the exception being P. spinifer where 51.0% of its records were from non-citizen science sources (for comparison: L. cervus—4.7%; L. barbarossa—7.1%; D. parallelipipedus—8.2%).

Validated records have increased over the years, especially when citizen science data started to be collected. Between 1870 and 2015, 493 records were collected and validated for the four species, with an average of three validated records per year. On the other hand, between 2016 and 2023, 6813 records were validated, with an average of 851 validated records per year and a maximum of 1610 in 2020 and a minimum of 409 in 2016. This represents a 284 times increase in average annually collected and validated records.

We found that adding citizen science records to “traditionally” gathered data substantially increased the known distribution for all species (Table 4). L. barbarossa showed the greatest increase (7.2⨉), followed by D. parallelipipedus (3.9⨉), L. cervus (3.6⨉) and lastly, P. spinifer (1.6⨉). Citizen science records allowed the validation of most of the current 10 × 10 km known grid cells for all species. The species with the highest validation rate was L. cervus (95.5%). The second was L. barbarossa (93.5%), followed by D. parallelipipedus (90.0%) and lastly, P. spinifer (50.0%) (Fig. 2).

The distribution models resulting from biomod2 ensembles identified the potentially suitable areas for each species (Fig. 3.d). These were created based on the top variables with the highest importance scores (Supp. Information—SI 2). Importance scores for each variable strongly varied across species and without having a direct correlation with differences between observed (i.e., presence) and random (i.e., pseudo-absence) distributions (with larger differences in the mean/median usually signalling greater ability to discriminate between presence vs. absence conditions; see Supp. Information—SI 3). The observational interval for each variable did not assess the whole range of suitable conditions (but most of it), although in most cases, the median value for each species showed a significant difference compared to the median of the overall distribution of the variable in the country, identifying particular ranges where the species are mostly present/observed (Fig. 4).

Model performances (Fig. 6a) varied across species and data sources. Overall, model performances for AUC when combining all data sources ranged from excellent (i.e., AUC ≥ 0.9) for P. spinifer, L. cervus, L. barbarossa (in decreasing order), and very-good (0.8 ≤ AUC < 0.9) for D. parallelipipedus. For TSS, slightly more conservative performances were found with models for L. cervus, L. barbarossa, and D. parallelipipedus displaying very-good performances (0.6 ≤ TSS < 0.8) for the combined dataset, whereas P. spinifer attained excellent performance (TSS ≥ 0.8). Overall, D. parallelipipedus, followed by L. cervus, recorded the highest gains in predictive performance when combining all data sources, whereas L. barbarossa and P. spinifer had no performance gain. However, it should be noted that comparisons are limited by the different sizes of each dataset (Fig. 6a).

By combining citizen science with pre-existing data, it was possible to increase the known species' suitable habitat (1 × 1 km grid cells). The biggest increase was found in P. spinifer (2.1 times, 3638 to 7696 cells), followed by L. barbarossa (1.9 times, 8857 to 16,759 cells). The third species with the biggest increase was D. parallelipipedus (1.3 times, 14,837 to 19,122 cells) and finally L. cervus (1.1 times, 14,608 to 16,368 cells)—Fig. 3 d). The number of 10 × 10 km grid cells that comprise suitable areas (SA) and Highly suitable areas (HSA) for all species is present in Fig. 3c.

Citizen science data is responsible for a different growth of the known niche hypervolume information for each species. For L. cervus about 21.5% of the niche hypervolume comes from citizen-science data. For L. barbarossa it is 75.8%, for D. parallelipipedus it is 41.4% and for P. spinifer it is 10.1%, the smallest share for all species. There is a percentage of the hypervolume niche that is shared between data sources: about 47.8% for L. cervus, 21.0% for L. barbarossa, 54.5% for D. parallelipipedus, and 3.0% for P. spinifer (Fig. 3a).

The two-dimensional representation of niche hypervolumes allows us to understand how the data sources bring a different perception of the species niches, by adding novel or previously unknown parts of the suitable bioclimatic niche. The biplot in Fig. 5, showing the mean temperature of the warmest quarter (BIO-10, x-axis) vs temperature seasonality (BIO-04, y-axis), shows not only increases, expansions and new niche ‘regions’ (L. barbarossa example), but also how previously isolated “parts” of the niches are actually continuous (D. parallelipipedus example). It is also interesting to notice that the centroid of the hypervolume changes based on the data source. For P. spinifer, (despite being few) records brought novel information regarding the species niche, significantly increasing it and showing potential new subpopulations (identified in the blue region of the low right corner of Fig. 5). The current niche hypervolume size, combining citizen science and non-citizen science data, is larger for all species compared to only non-citizen science data. Total niche hypervolume has increased 1.1 × for L. cervus, 4.2 × for L. barbarossa, 1.8 × for D. parallelipipedus and 1.6 × for P. spinifer (Fig. 6b).

Our results allow us to identify and describe the environmental conditions that best predict the presence/absence of the species:

In general, L. cervus and D. parallelipipedus are both related to high levels of precipitation, low annual temperature variation (although D. parallelipipedus is more suited to habitats with higher annual mean temperature), low solar radiation (indirectly linked to less productive areas that commonly were not converted to other land uses throughout the years) and a high percentage of native forests with tall trees and low density. Overall, this latter result clearly indicates a need for mature/old-growth native forests where large/tall native trees dominate the habitat. Both species can also occur in landscapes more dominated by non-native forests, probably due to the presence of natural features that are retained throughout the landscape (with different levels of conservation, abundance and structural conditions), like large trees and deadwood. L. cervus prefers habitats where the soil has a significant percentage of coarse fragments. Although D. parallelipipedus does not prefer particular soil conditions, it seems to be more commonly associated with soils similar to those preferred by L.cervus.

The low amount of data for Platycerus spinifer does not allow us to extract as much ecological information on its niche as the other species. Nonetheless, the species seems to be more suited to areas with lower temperatures during the warmest season and related to high levels of annual precipitation, even in the driest season, with a low mean annual temperature and a high temperature and low precipitation seasonality. Altogether, these conditions clearly relate to the mountain ranges of the north of Portugal. In these regions, P. spinifer is, in coherence with the former species, linked to a high percentage cover of native forests and is more frequently observed in areas with a high tree density with low maximum and mean forest canopy height, potentially indicating a preference for habitats dominated by very tall and dense shrubs (almost arboreal).

Our historical perspective (before 2016) on lucanid species distribution in Portugal showed that the available information did not fit conservation and management needs. Other than the latest Portuguese Habitat Directive report, the last major publications on lucanid species distributions were published more than twenty years ago. Grosso-Silva (1999) synthesised and updated the available information on the distribution of two lucanids in Portugal, and López-Cólon (2000) provided an overview of the diversity, ecology and distribution of Iberian lucanids. Collectively, the results of our effort showed that the number of georeferenced historical records available was low, spatially biased and highly fragmented, even when considering some additional records that were published after 1999 (Grosso-Silva 2005, 2009; Whitehead 2007; Ferreira 2012; Costa 2012; Cox et al. 2013; Ferreira et al. 2021; Gil et al. 2021). The low number of records could be a symptom of the ongoing European “insect taxonomist extinction” (Hochkirch et al. 2022), where the pool of taxonomic experts has been declining for decades. This reduction in insect taxonomists leads to a threatened or eroded taxonomic capacity for 41.4% of the insect orders (Hochkirch et al. 2022), with Portugal having a “poor taxonomic capacity”, ranking 20 out of 46 EU countries, with only 33 identified taxonomists who have expertise in ~ 24% of the insect orders (Hochkirch et al. 2022). Additionally, Coleoptera is one of the insect orders with reduced numbers of taxonomic experts, resulting in a rating of “poor taxonomic capacity”, a status equivalent to “vulnerable” in Red List assessments (Hochkirchet al. 2022).

In contrast to taxonomists' decline, parataxonomist and citizen science projects have been growing throughout Europe (Vohland. et al. 2021; Hochkirch et al. 2022). Insects have been one of the focal topics, leading to an increase in monitoring actions and improved efficiency of research on insect conservation. Citizens interacting with taxonomists benefit by increasing insect observation, detection and identification capacities, which, in some cases, has led some to become parataxonomists (Hochkirch et al. 2022). Scientific projects rooted in citizen participation have the advantage of covering large areas over long periods with limited funding, which is needed for long-term monitoring of species such as L. cervus, which has a maximum imago life-span of three months in Portugal and hence requires “temporally-concentrated” efforts from volunteers. Our citizen science collected data contributed to a significant increase in the number of records and resulted in an increase in the known distribution of lucanid species in Portugal. Since none of the species’ distributions was 100% validated by citizen science records, some taxonomic expertise or more focused prospection is still necessary to complete and fully validate model outputs on species distributions and habitat preferences.

The success of VACALOURA.pt is consistent with other citizen science initiatives aimed at increasing the knowledge about L. cervus national distributions, as evidenced in Italy (Zapponi et al. 2017), Croatia (Katušić et al. 2017) and more recently in Spain (Méndez and Cortés-Fossati 2021). These outcomes demonstrate the necessity of collaborative efforts to improve species dynamics and to efficiently gather extensive information that cannot be obtained through traditional methods or solely by experts.

Nevertheless, citizen science does not work the same for every species (Van Eupen et al. 2022). In our specific case, P. spinifer, a smaller, rarer and less conspicuous species, showed a comparatively modest increase in the known distribution compared to the other lucanid species.

In our case, there were generally more reports for larger species that commonly occur closer to urban areas, which are readily detected and identified by citizens and/or have larger natural ranges. We posit that the species’ natural range, ecology (habitat and phenology), body size and cultural recognition represent crucial traits that can help us understand the variation in our results and discern the success of citizen science initiatives between species. However, further work should be done to quantitatively assess this topic with benefits for improving citizen science projects similar to ours.

Data quality and quantity are paramount for modelling species distributions and obtaining robust inferences on species-environmental relationships (Elith and Leathwick 2009). For instance, with low sample sizes, the environmental space covered by species' occurrences may be incomplete and/or biased, often degrading model performance (Wisz et al. 2008). This limited coverage can result in an incomplete spatial representation of the species’ niche, not only because of the small amount of data but also because of the methodological limitations that these few records bring (Radomski et al. 2022), leading to truncation in particular environmental dimensions (Chevalier et al. 2021). The absence of data from specific environmental conditions may also result in the exclusion of suitable habitats from predicted distributions (Thomaes et. al 2008). In turn, these insufficiencies leak into species conservation, management, monitoring and vulnerability assessment (Kühl et al. 2020).

By modelling each species distribution and niche hypervolume with each data source (i.e. past references, citizen-science and combining both) (Redolfi De Zan et al. 2023), it is possible to assess how the project’s continuous efforts in recording the species are helping to preserve them.

Our analysis of niche hypervolume for key climatic variables, along with identifying and understanding critical ecological structural variables from SDMs, brings new insights to survey and preserve lucanid species. Further than validating the current known distribution, we were able to identify new potentially suitable locations for each species (Rhoden et al. 2017; Marsh et al. 2023). Moreover, niche hypervolume analyses with citizen science data (and its combination with pre-existing sources) allowed us to identify particular regions with previously unknown combinations of environmental conditions. For the species with more records (L. cervus and D. parallelipipedus), it is possible to confirm distribution ranges larger than the modelled highly suitable areas (HSA), showing that the increased amount of information provided by the project can help to map the overall suitable distribution of the species. The effort was marked by the surprising discovery of new subpopulations for all species, particularly for L. barbarossa and P. spinifer, highlighted by the citizen science data through the niche hypervolume analysis (Fig. 6b). Both these species (which have fewer records) show a continuous and steady expansion of the recorded distribution throughout the years. Still, increased survey work is needed to validate the overall modelled distribution areas.

These results underscore the significance of complementarity provided by data from new sources, even when only a limited number of records are available (see also SI 04). Overall, the amount of pre-existing data, species detectability and its adoption in citizen science programs may justify the different levels of improvement in data quantity, model accuracy and niche hypervolume increase across species. This suggests that species with more significant knowledge gaps but better detectability will benefit the most from citizen science initiatives, although we suggest increasing surveying efforts for all taxa. In order to reach it, better sampling schemes and monitoring protocols supported by the best data and model inferences should be put forward and that is only possible if we also improve the ecological knowledge of the species.

Identifying previously unknown sites and potentially new subpopulations through citizen science efforts is critical for fully understanding species niches, their distributions and assessing the impacts of climate and land use, which currently influence species conservation and maintenance. By modelling species suitable areas, we can not only validate the known distribution but also identify and recommend new locations to survey and monitor with the support and collaboration of citizens, parataxonomists, taxonomists or even with the help of trained conservation dogs and the use of new tools like eDNA (Bennett et al. 2020; Riaz et al. 2020; McKeague et al. 2024). This is extremely important for species conservation since validating occurrences in new areas might help increase the known species range and improve future conservation actions.

We now know that the species range is larger than the citizen-led observations can record (Fig. 3b), highlighting the need for targeted surveying and monitoring in new regions. Identifying new areas of occurrence can trigger the discovery of other novel subpopulations, the possibility of expanding knowledge on species' environmental niches and allow us to improve our understanding of how to integrate species conservation in humanised landscapes. Large trees, deadwood, and the structure and abundance of deciduous forests are small natural features (SNF) (Lindenmayer 2017) that, if protected or enhanced, can support our target species’ preservation (Parmain and Bouget 2018; Méndez and Thomaes 2021) since they can hold a large part of the species niche requisites throughout large periods of time (Hunter et al. 2017; Poschlod and Braun-Reichert 2017). SNFs are remnants of natural ecosystems that were once more abundant than today and that are considered keystone structures with ecological importance disproportionate to their scale (Lindenmayer 2017) while also having a significant role in engaging citizens and local stakeholders in conservation actions (Blicharska and Mikusiński 2014).

Identifying areas rich in these small natural features or habitats that support the presence of the species can assist in delineating and promoting citizen science networks and detailed protocols for in-field assessment. The current effort of “passive” citizen science collecting data might be improved by an “active” identification of areas to survey and monitor. Remote sensing data and spatially explicit model predictions can aid in this effort by identifying fine-scale areas for priority field surveying and monitoring.

In the face of climate change, sustainable and integrative land management can significantly impact landscape and territorial resilience. By identifying which features to conserve and how, it becomes feasible to sustain species, their habitats and, thereby, enhance the ecosystem services they provide.

The increased knowledge on our species is vital to understand where the natural habitats and SNF that they need are located, leading to localised actions towards their conservation, which will consequently enhance citizen science initiatives (e.g. improved national and local governance, integration of evidence-based management and implementing payments for ecosystem services that support private owners who integrate nature conservation in their land (Farley and Costanza 2010; Lindenmayer et al. 2014; Blicharska and Mikusiński 2014; Hunter et al. 2017; Hansen et al. 2023)).

Citizen science initiatives can then include wider and often more informative aspects for conservation, such as assessing species abundance, habitat attributes, quality and quantity. In a changing world, where policies targeted at human settlements and activities are sometimes antagonistic with nature conservation, it is crucial to find a middle ground to integrate conservation instruments (Krumm et al. 2020) that also help regenerate the economic activities that local communities have. It could even be possible to start integrating these conservation instruments in the EU Natura 2000 network, which has not been able to fulfil its goals of reducing biodiversity decline (especially for insects) despite the increase in nature conservation efforts and expenditure (Engelhardt et al. 2023).

Portugal’s Natura 2000 network could serve as an ideal “pilot territory” for implementing a comprehensive conservation strategy that integrates SNF conservation with Citizen Science initiatives. The major land-use changes that have occurred on mainland Portugal over recent decades that are also evident in Natura 2000 areas were driven primarily by socio-economic growth, leading to a generalised urban sprawl (in coastal and metropolitan areas), increased landscape fragmentation and an inland abandonment of rural areas and traditional land uses (Jones et al. 2011; Meneses et al. 2017). The result is a country with an overall increase in urban and forested areas and a decline in agricultural areas, with more than 95% of mainland rural areas privately owned (ICNF 2019). Moreover, the forest increment brought a marked shift in forest quality, with monoculture plantations of Eucalyptus globulus expanding more than 400% between 1970 and 2010 (Deus et al. 2018) and currently accounting for 26% of the forested areas in Portugal (ICNF 2019). With these transformations came a generalised decrease in biodiversity and wildlife species abundance (Matos and Dos, 2011; Cruz 2013; Deus et al. 2018) and, consequently, an increase in high-intensity fires (Fernandes et al. 2016; Silva de Oliveira et al. 2021). In 2017, a forest “cleaning” law–aiming to reduce forest biomass–was updated and implemented at a large scale with the purpose of preventing new events like the catastrophic ones that occurred that year (Turco et al. 2019). Extensive fines were implemented for those who would not comply with the removal of any vegetation or debris on a buffer area of 10–50 m surrounding buildings and roads. This led to a national-scale generalised removal of finer debris and deadwood from forests (each with different impacts on forest fires (Larjavaara et al. 2023)), which was already culturally ongoing and, to the felling of large trees in the wildland-urban interface leading to a reduction of SNF abundance in an already degraded territory.

Based on our results, where all of our focal species were observed in low productivity areas (e.g., low solar radiation and vegetation cover), a focus on applying integrative conservation instruments, such as the preservation of SNF (Krumm et al., 2020) in these sites, could start a new nature conservation era in highly fragmented and humanised landscapes.

Our results show that an increased knowledge of the species distribution and a deeper understanding of their ecological niches can be used to address species conservation needs. Nonetheless, the data collected in the VACALOURA.pt Project cannot be used to assess species’ risk of extinction since population or habitat trends data were not collected. Furthermore, we only used four saproxylic species, which are not considered umbrella species for the presence of old-growth forests or the “naturalness” of forest ecosystems (Eckelt et al. 2018), although they can be considered flagship species of these ecosystems (Lachat et al. 2012; Tini et al. 2018). Efforts towards conserving these species may then help promote the conservation of SNFs in fragmented habitats, which are remnants of natural or old-growth ecosystems (rare in Portugal (Barredo et al. 2021)). Improved information on SNFs distribution and conservation value can enhance monitoring efforts and help us to set up policies to regulate their protection (Lindenmayer and Laurance 2016). For instance, integrative conservation measures should be improved to reduce detrimental pressures to the target species (and biodiversity overall). Examples of these pressures are land use changes (e.g. increased urbanisation and monoculture plantations) and the direct exploitation of forest and natural resources (e.g. large tree logging and deadwood removal) that have been reducing and fragmenting these species’ habitats and metapopulations for decades (IPBES, 2019). Conservation measures such as the integration of ecological indicators on land use planning, the implementation of deadwood retention or restoration strategies (based on national and regional criteria) or the protection and valuation/valorisation of large trees (e.g. based on the ecosystem services that they provide) are just some examples that can be implemented to benefit biodiversity (Farley and Costanza 2010; Kraus and Krumm, 2013). These measures have recently been approved for implementation and monitoring across Europe as part of the new European Nature Restoration Law. Building on model results, we have confirmed our species' preference towards landscapes with a high percentage of native forests with tall trees and low density, conditions normally met in mature/old-growth native forests. As such, their abundance in certain areas might help us to identify SNFs that could be targeted for conservation with clear benefits for forest biodiversity from local to regional scales.

By increasing and improving monitoring efforts, Citizen Science projects create regional or national scale networks, engage different stakeholders, and refine conservation needs, which can be better tackled at different scales while increasing citizens' awareness regarding local pressures on biodiversity, leading to more active and participatory communities.

This work is a first step towards using citizen science data to improve integrative conservation for Portuguese forest biodiversity. It brings vital insights to tackle the conservation of saproxylic species in different habitats and similarly fragmented regions or countries. Increasing the detail level regarding species niches and species ecological requirements at a fine/local scale makes it possible to keep improving monitoring efforts and conservation actions in a fast-paced changing world.