2.1 Study area and bird data

We tested our approach in the Iberian Peninsula (IP; southern Europe) since it covers a wide range of environmental gradients. The IP (581,200 km²) is administratively divided between Portugal (PT; 89,015 km²) and Spain (SP; 492,175 km²) and is characterized by a combination of natural and human history, geologic and topographic heterogeneity, and strong climatic gradients, offering a wide range of environmental conditions for hosting a broad variety of endemic and rare species (Underwood et al., 2009).

To fit SDMs at the first stage, we used bird occurrence data from two different data sources of biodiversity (Table 1): (i) a standardized dataset based on national Bird Atlases (Atlas) and (ii) a citizen science (i.e. non-standardized) dataset based on the EOD—eBird Observation Dataset from the Global Biodiversity Information Facility—GBIF (eBird). To reduce the potential geographical errors in species records that can strongly influence the results of models (Hijmans, 2012), we filtered the original dataset removing duplicates and positional/spatial errors such as outliers using R and qgis programs, and nomenclature errors and taxa misidentification supported by expert knowledge. In addition, we harmonized the species records in grid cells with a resolution higher than 10 km before the modelling procedures. We adjusted eBird data to the spatial resolution of both Atlases (10-km UTM square) to standardize input data and make both datasets comparable. Predictive variables were aggregated at 10-km UTM square (see below). We also matched the eBird data to the years and months within the Atlas data (1999–2012), from late February to mid-August, the breeding season in the IP (SEO/BirdLife, 2020).

Additionally, to perform subsequent comparisons, we built a full dataset from the combination of both Atlas and eBird datasets. This full dataset represents the best knowledge of bird distribution in Iberia as it includes all occurrences recorded in the two datasets. To assess the effect of species distributional range (from narrow-ranged to widespread species) on modelling performance, we grouped the species records from the full dataset in four sets of size classes based on number of occurrences at 10-km UTM squares: (i) Class I (10–100); Class II (101–500); (ii) Class III (501–1000); Class IV (>1000). We also considered the conservation categories of species in the IUCN Red List to assess the potential effects of model uncertainty on decision-making for bird conservation. Therefore, each species always belongs to the same group (n = 236) because the full dataset represents the actual distributional range of each species, and it is not affected by potential sampling biases of each dataset (e.g. in the eBird dataset).

2.2 Environmental predictors

We selected environmental factors that are known to influence bird physiology, distribution and local habitat preferences, as close as possible to the time period covered by the species occurrence datasets (1999–2012) (Table S1 in Appendix S1).

We derived 19 (bio-) climate predictors from monthly temperature and rainfall data. This climatic dataset was obtained for historical conditions (1979–2013) from the CHELSA 1.2 database at a spatial resolution of 30 arc-sec (~1-km pixel size).

Satellite remote sensing data from the Moderate Resolution Imaging Spectrometer (MODIS) on-board the Terra satellite platform were used to derive remotely sensed ecosystem functional attributes (EFAs) (Regos et al., 2022, and references within), to characterize species habitat dynamics as a counterpoint/complement of climate data. To compute EFAs, we used the MODIS Enhanced Vegetation Index (EVI) (MOD13Q1.v006; 232 m pixel every 16 days) as a proxy of vegetation greenness, biomass and leaf area index—with values ranging from −1 to 1, with healthy vegetation generally holding values between 0.20 and 0.80, and for the 2000–2012 time period. For that, we used Google Earth Engine (GEE) cloud-based platform (Gorelick et al., 2017) to derive originally 11 metrics of the EVI seasonal dynamics (Table S1 in Appendix S1). These statistical measures were calculated for each complete year. To capture the multi-year normal conditions of each EFA variable, thus reducing the effect of stochastic interannual climatic fluctuations, we computed the overall mean. EFAs were exported from GEE at 1-km squares of final spatial resolution.

All environmental, climate and remotely sensed variables were aggregated from its original spatial resolution (1 × 1 km) by computing the mean values within each 10-km UTM square, to match to the spatial resolution of bird datasets, 10 × 10 km. To avoid including highly correlated variables in model fitting, we conducted a multicollinearity analysis by testing Pearson pairwise correlations and variance inflation factors (VIF). Based on these multicollinearity analyses, we retained those predictors with Pearson's correlation coefficients of <0.7 and VIF of <5 to calibrate models in following steps (Figures S1.1–S1.4 and Table S1 in Appendix S1). Based on the pairwise correlations and the collinearity assessment from initial 19 (bio-) climate variables, three temperature-related (bio4—Temperature Seasonality; bio8—Mean Temperature of Wettest Quarter; bio9—Mean Temperature of Driest Quarter) and two precipitation (bio16—Precipitation of Wettest Quarter; bio17—Precipitation of Driest Quarter) variables were selected (Table S1 in Appendix S1) to build the climate dataset. Similarly, we selected five remotely sensed EFAs as descriptors of species habitat dynamics: EVI annual mean (EVI_mean as surrogate of annual total amount of primary production), EVI annual minimum (EVI_min as an indicator of the annual extremes), EVI seasonal standard deviation (EVI_sd as descriptor of variations between seasons), and dates of maximum (EVI_dmax) and minimum (EVI_dmin) EVI (indicators of phenology—growing season) (Table S1 in Appendix S1).

2.3 Model fitting

We used the species occurrences as response variables in modelling processes within each 10 × 10 km UTM square (Figure 1).

We designed three groups of models: (1) climate-based models (CLIM_only)—using variables extracted from the monthly (bio-) climate dataset; (2) models based on habitat/ecosystem functioning-related predictors (EFAs_only); and (3) the combined model, combining the most significant five (and uncorrelated) predictors selected from the previous partial climate and EFA-based models (CLIM_EFAs).

We calibrated SDMs using an ensemble forecasting approach based on the nine modelling techniques implemented in the ‘biomod2’ R package (Thuiller, 2014): (1) artificial neural networks (ANN); (2) classification tree analysis (CTA); (3) flexible discriminant analysis (FDA); (4) generalized additive model (GAM); (5) generalized boosted models (GBM); (6) generalized linear model (GLM); (7) multivariate adaptive regression splines (MARS); (8) maximum entropy using Phillip's Maxent software (MAXENT) and (9) random forests (RF). Default parameters were used for all modelling techniques, with the exception of the smoothing degree term in GAM algorithm which was set to k = 4, and the number of boosting trees in GBM (n.trees = 2000) to prevent over-fitting issues (Guisan et al., 2002). We also generated a total of 10 sets of randomly distributed PAs in the model calibration. PAs were generated by assigning unoccupied grid cells with the following constraints: (1) generating the same number of PAs as of presences to avoid potential bias caused by different levels of prevalence in the presence/absence datasets (as recommended by Barbet-Massin et al., 2012; Manel et al., 2001) and (2) defining a minimum distance between PAs, corresponding with the grain size (10 km), and without overlapping with presences (Wisz & Guisan, 2009), to avoid spatial autocorrelation and to cover the different ecological conditions in the study area. Each model was fitted using 70% of the data and tested using the remaining 30%. On the other hand, considering that there is not consensus about the minimum number of species records to fit models (Breiner et al., 2015; Sillero et al., 2021), we followed the criterion provided by Franklin (2010) and the species with less than 10 occurrences at 10 km grid cells were excluded to avoid model overfitting and ensure the most threatened species can be included in the analysis.

2.4 Model evaluation and performance

We employed hold-out cross-validation to evaluate the models with 10 evaluation rounds for each PAs set. The predictive performance and discrimination ability of individual models were evaluated using two metrics (Baasch et al., 2010): (i) the area under receiver operating characteristic curve (AUC) and (ii) and the true skill statistic (TSS). The AUC ranges between 0 and 1 (models with AUC ≥0.7 were considered good), while TSS ranges from 0 or less to 1 (models with TSS ≥0.4 were considered good). Finally, to deal with uncertainty in our models coming from the single-algorithm techniques, we built ensemble (consensus) models among those satisfying the conditions AUC ≥0.7, and TSS ≥0.4. We used the weighted mean of all the partial projections (Marmion et al., 2009), a consensus method that considers the weights proportional to the selected evaluation scores (i.e. the higher the AUC of the model, the greater the importance in the ensemble modelling; Konowalik & Nosol, 2021).

Once the predictions of the ensemble models by each species were obtained for each dataset, all model comparisons (namely data source, range class, predictor type and evaluation metric) were carried out on the full dataset resulting from the combination of both Atlas and eBird datasets (n = 236). Considering that Atlas could be treated as a presence–absence dataset and the eBird as only-presence dataset, we tested the performance of the resulting ensemble models for the species hosted in the subset of 236 species using two complementary measures of accuracy: the AUC and Boyce's Index (BoyceI). Unlike the TSS, a prevalence-independent and threshold-dependent binary measure of model accuracy but usually highly correlated to AUC (Shabani et al., 2016), the Boyce's Index is an appropriate metric in the case of presence-only models, measuring how much model predictions differ from a random distribution of the observed presences across the prediction gradients (Hirzel et al., 2006; Pearce & Boyce, 2006). Both accuracy measures (AUC and Boyce's Index) were calculated from the model predictions using the R packages ‘PresenceAbsence’ (Freeman & Moisen, 2008) and ‘ecospat’ (Di Cola et al., 2017), respectively.

In addition, to test the effect of data source, type of predictor (quantitative), size class and accuracy metric (namely AUC and Boyce's Index) on models performance, we developed generalized linear mixed-effects models (GLMMs; Bolker et al., 2009) with R package ‘lmer4’ (Bates et al., 2015). We fitted GLMMs with gamma probability distribution and inverse link (after checking for overdispersion effects in Poisson models and robust estimation via weighted likelihood), as the dependent variable (the values of AUC and the Boyce's Index) showed continuous probability distributions. We fitted ‘data source’, ‘predictor’, ‘size class’ and ‘accuracy metric’ as fixed effects, while ‘species’ was considered a random effect. We only fitted interactions between ‘data source’ and ‘size class’. Fixed effects were considered significant at p-values <0.01.

2.5 The most contributing predictors and spatial projections across set of models

To examine the contribution of satellite-derived EFAs as integrative predictors of habitat dynamics in SDMs for narrow-ranged and widely ranged bird species, we compared the relative importance of three sets of predictors for a total of six final ensemble models (CLIM_only, EFAs_only and CLIM_EFAs), ranging between 0 (no importance) and 1 (high importance) (Araújo & New, 2007). Only predictors with importance >0.1 were considered.

To assess the spatial confidence associated with model results, we compared the spatial projections of the six sets of models based on two data sources (Atlas vs. eBird). For that, we performed pairwise comparisons between spatial projections of each species distribution and obtained at each combination of data source, set of predictors and finally grouped by size class. To do so, we used the ‘raster. overlap’ function in the ‘ENMTools’ R 1.0.2 package (Warren et al., 2008, 2021). This function measures similarity in the geographical distribution of suitability scores from pairwise SDMs. Among other metrics (Warren et al., 2008), the niche overlap is calculated using Schoener's D Index which varies from 0 (complete divergence/no overlap) to 1 (high similarity/complete overlap). In addition, to map uncertainty from the different data sources, we examined the agreement of spatial projections between models obtained by different sets of predictors. For that, we first performed binary transformations of the habitat suitability predicted by the ensemble models into presence–absence maps based on the AUC optimized thresholds available on the ‘biomod2’ R package (Thuiller, 2014), since AUC it is a robust threshold-independent measure of a models ability to discriminate presence from absence (Lawson et al., 2014). Then, we combined the predictions by simply overlapping the binary maps, resulting in similarity maps that ranged from 0 (no prediction) to 6 (prediction based on the six models).

The entire process of obtaining, editing and analysing data, as well as model calibration, mapping and evaluations, was carried out in R software 4.0.1 available at CRAN (http://cran.r-project.org/), the GEE platform and in qgis 3.14 software.

3.1 Data structure

The number of species records available in the eBird dataset was, approximately, 30.7% higher than in the Atlas dataset (Figure S2.1 in Appendix S2). However, the number of 10-km UTM squares with data in the eBird dataset was 71.75% lower than in the Atlas dataset (Figure S2.1 in Appendix S2), which did not prevent to cover a similar environmental gradient than the Atlas (see maps of the multivariate environmental similarity surface—MESS analysis in Figure S2.2 in Appendix S2). The number of species recorded in the Atlas was 15 species higher than in the eBird (Figure S2.3 in Appendix S2). Considering the size classes based on species ranges, while the Atlas dataset hosted the highest number of narrow- and wide-ranged species (I: 10–100 and IV: >1000 size classes), eBird hosted the highest scores for the II (101–500) and III (501–1000) classes in the IP (Figures S2.3 and S2.4 [datasets combined] in Appendix S2). Overall, the most represented size classes were those that include more widely distributed or common species (IV, II and III, in this order). In terms of IUCN conservation categories, the Least Concern level was overwhelmingly more frequent in both datasets followed by Vulnerable level and considering the selected period (1999–2012) (Figures S2.5 and S2.6 [datasets combined] in Appendix S2).

3.2 Model evaluation and performance

Overall, the ensemble SDMs yielded higher predictive ability for both bird datasets and all groups of predictors when compared to single-algorithm models as measured by median, and interquartile range to be compared across models (AUC_median-Atlas: 0.89 ± 0.09 and BoyceI_median-Atlas: 0.83 ± 0.33; AUC_median-eBird: 0.72 ± 0.19 and BoyceI_median-eBird: 0.82 ± 0.36) (Figure 2; Text S3.1 and Figures S3.1 and S3.2 in Appendix S3). Models showed contrasting results depending on the evaluation metric. For equal species number (n = 236), and for ensemble models, while AUC values were higher for models calibrated with Atlas data, the presence-only Boyce's Index indicates better performance of models fitted with eBird data (Table S3.1 in Appendix S3). Differences in model accuracy were found between both data sources and size classes (t_eBird-IV: 7.869; p-value <0.05) and evaluation metrics (t_BoyceI: 4.838; p-value <0.05) (Table S3.1 in Appendix S3 and see accuracy metrics per each species in Table S3.2).

Based on GLMM results, and considering the species range (size classes) and type of predictors, climate-based models (individually or in combination with EFAs) fitted with Atlas data showed the highest AUC values for narrow-ranged species (AUC_{CLIM_only-I}: 0.99 ± 0.02; AUC_{CLIM_EFAs-I}: 0.99 ± 0.03) (Figure 2a; Table S3.1 in Appendix S3), while climate- and EFA-based models alone fitted with eBird data showed the highest performances for the same size class when the BoyceI was used (BoyceI_{CLIM_only-I}: 0.93 ± 0.22; BoyceI_{EFAs_only-I}: 0.93 ± 0.10) (Figure 2b; Table S3.1 in Appendix S3). In addition, although the BoyceI showed higher performance than AUC for the less widespread species (size class I, II, III), Atlas-fitted models showed higher values than eBird-fitted models. Predictor types also affected modelling performance (t_{EFAs_only}: 3.434; p-value <0.05) (Table 2).

Regarding the IUCN conservation categories, while the Atlas showed the highest values of the AUC for all classes (Figure S3.3a in Appendix S3), the highest Boyce's Index values were found for the most threatened species when modelled with the eBird dataset (Figure S3.3b in Appendix S3).

3.3 The variable importance ranking

The contribution of each independent variable to models by type of predictors was also affected by the combination of data source, type of predictor and size class. Overall, bio17 (Precipitation of Driest Quarter), bio4 (Temperature Seasonality) and bio9 (Mean Temperature of Driest Quarter) were the most contributing predictors of climate group to model performance across size classes, while EVI_mean, EVI_min and EVI_sd, descriptors related to productivity and seasonality in primary productivity, were the most important ones within the group of the habitat attributes (EFAs_only) (Figure 3). However, when we combined both types of predictors (CLIM_EFAs) in the models, bioclimate predictors (bio17 and bio4) held the highest importance across size classes, followed by EVI_mean (Figure 3).

3.4 Overlap and uncertainty in spatial projections

Overall, the similarity test based on Schoener's D Index indicated that pairwise niche overlaps between models based on eBird and Atlas datasets for each species were high for the tested models (D_overall-mean = 0.82 in all cases). However, the similarity between spatial predictions increased with the distributional range, from narrowly distributed species (I_Dmean = 0.78 and II_Dmean = 0.81) to more widespread species (III_Dmean = 0.82, IV_Dmean = 0.83) (Figure 4). The similarity test also showed that, in general, the overlap between species niches, from both Atlas and eBird datasets, predicted by the EFAs was higher than the predicted by the climate and the combination of climate and EFAs, being highly significant in the case of widely distributed species (based on Kruskal–Wallis test [K–W]: class II: K–W, p = 3.3e-5; class III: K–W, p = 0.0057; and class IV: K–W, p = 0.00012; Figure 4).

The species-specific spatial projections of the similarity maps derived from the Atlas and eBird datasets and the three groups of predictors (CLIM_only, EFAs_only and CLIM_EFAs) allowed the identification of UTM squares where the predictions of the six models overlapped (Class 6), and therefore were the most consistent across models, for widespread (Figure 5b) and narrow-ranged (Figure 5c) species. Maps for all species in Figure S3.4 are found in Appendix S3.

4.1 The added value of citizen science occurrence data for more accurate predictions

Albeit eBird has evolved from a citizen-science project into cooperative partnerships among experts in a wide range of fields, it still hosts several challenges that can inhibit robust ecological inferences (e.g. species and spatial biases, variation in effort and observer skills, among other). A priori, observers (citizens) may be more likely to misidentify or miss, especially rare or not very abundant species, but overall identify target species as accurately as trained researchers (Aceves-Bueno et al., 2017). This, therefore, suggests that citizen science (such as eBird) data may complement more traditional standardized occurrence datasets (such as Atlases), or even substitute for other presence-only datasets (Cox et al., 2012).

Despite differences in sampling designs and protocols between data sources (e.g. a lower number of 10-km UTM squares with data in eBird than Atlas), our results confirmed the usefulness of both citizen science project eBird and standardized and well-structured Atlas datasets to predict bird species distributions with different distributional ranges in the Iberian Peninsula. Overall, Atlas-based ensemble models showed better performance through the AUC, while eBird-based ensemble models showed better performance when the specific presence-only metric BoyceI was used (Figure 2; Table 2; Table S3.1 in Appendix S3). These results can be related to the type of evaluation metric and the nature of the response variable. The AUC is a metric for models fitted with presence/absence data, which are better represented by standardized surveys such as those performed in Atlas projects than opportunistic records from citizen-science programs. Considering species grouped by distributional range through sample size classes, models fitted with Atlas data showed higher BoyceI values for widespread species for the same group of predictors, while the eBird-based ensemble models showed higher BoyceI values across all range sizes except the most widespread species (Figure 2; Table S3.1 in Appendix S3). Our results suggest that SDMs fitted with eBird data can predict species distributions as accurately as those based on Atlas data.

4.2 Contribution of predictors to the performance and predictions of SDMs

Unlike the quality of species observations and their effect on the performance of SDMs, which have been extensively documented (Fei & Yu, 2016), the uncertainty associated with environmental variables as predictors should have been more deeply addressed considering its key role in the process of calibrating and evaluating SDMs (Petitpierre et al., 2017; Scherrer & Guisan, 2019). Overall, predictors showed different effects depending on data source and size class. In terms of performance, our results showed that models fitted with Atlas-based data showed higher AUC values than models fitted by eBird-based data for the same group of predictors (Figure 2a). However, results were different when the BoyceI metric was used. Climate-based models showed the highest performance for the Atlas dataset based on the AUC, while EFAs-based models, and the combined model (CLIM_EFAs), showed the highest performances for the presence-only eBird dataset through the BoyceI (Figure 2b). Considering the contribution of predictors to model predictions, the variable importance overall seems to be consistent across bird datasets used to fit the models (Figure 3), being mostly driven by the species range (size classes) than the data source (eBird and Atlas) (Table 2).

Although SDMs have traditionally relied on climatic data as abiotic factors, among other, more integrative descriptors of high ecological relevance for species in terms of habitat dynamics and ecosystem functions have been neglected (Gonzalez et al., 2020). Satellite-derived EFAs stand up as integrative predictors of SDMs, thank its quicker spatiotemporal response of ecosystem dynamics to environmental drivers and changes than interpolated climate data and structural/compositional attributes. EFAs also timely capture shifts in habitat conditions that may trigger changes in populations and communities (Arenas-Castro et al., 2018; Regos et al., 2021; Vila-Viçosa et al., 2020; and references within). Our results showed that EFA-based models alone or in combination with climate data, performed similarly, or even better, than models only based on climate. Thus, our findings support the need for mainstreaming other key environmental factors for improving model performance, and hence, improving ecological inferences for decision-making or conservation planning.

4.3 Similarity and stability across spatial projections in SDMs

Equally important to the environmental and statistical modelling space is the confidence or scepticism related to the spatial projections or maps derived from SDMs, as they are usually inferred into geographical space from the previously fitted environmental space (Goberville et al., 2015; Thuiller et al., 2019). Overall, the similarity between our spatial predictions of species from the Atlas and eBird datasets was high and consistent across species range and compared models (Figure 4). As expected, the similarity between overlapping species niches was progressively increased as the species ranges were larger, but it was significantly higher between widely distributed species in EFAs-based models. The low similarity between overlapped spatial predictions for species with narrow distributions (Class I) are in line with the BoyceI-based predictive capacity (Figure 2) that was also higher for eBird-based data than for Atlas-based data for this same size class. The high similarity measurements for niche overlaps (D statistic) between eBird-based and Atlas-based models suggest that occurrence data derived from the eBird dataset can be an effective presence-only sample data source in ensemble SDMs for the Iberian Peninsula. In addition, the species compatibility/uncertainty maps built on the base of both Atlas and eBird datasets and calibrated with the three group of predictors (CLIM_only, EFAs_only and CLIM_EFAs) clearly represent a valuable resource to visualize the level of uncertainty and confidence hosted when modelling species distributions with different data sources (Figure 5; Figure S3.4 in Appendix S3).