Assessing spatial and temporal trends over time in potential species richness using satellite time-series and ecological niche models

The special conservation area of Montesinho/Nogueira (SAC-MN) is located in the northeast of Portugal mainland, and is a European Union’s Natura 2000 sites (https://​natura2000.​eea.​europa.​eu/​Natura2000/​SDF.​aspx?​site=​PTCON0002) (Fig. 1). It aggregates two main regions within its boundaries: i) Montesinho Natural Park (MNP) and ii) the Nogueira Mountains (NM) (Figure SM1). The MNP has a spatial extend of approximately 74 thousand hectares, hosting a high diversity of species. It stands out as an important PA in Portugal due to its rich array of endemic and highly endangered species (e.g., Galemys pyrenaicus (É. Geoffroy, 1811); Herniaria lusitanica (Chaudhri); or Monticola saxatilis (Linnaeus, 1766)), along with priority habitats (i.e., “lameiros”, which are mountain seminatural irrigated meadows). The region predominantly features a traditional mountainous agricultural landscape, with chestnut crops (Castanea sativa (Mill. 1768)) being a significant source of income for the local community (Castro et al. 2010). Conversely, the NM region is primarily characterized by dense and homogenous oak forests (e.g., the black oak, Quercus pyrenaica (Willd 1805) that have remained unaffected by land-use changes over the years (Alves et al. 2022) due to topography. The SAC-MN is a crucial Mediterranean crossroad zone for an extensive array of species, featuring a supra and oro-Mediterranean bioclimate, along with an Atlantic climate. In this sense, the region experiences high-temperature fluctuations, ranging from − 12 to 40 °C, and has an elevation range of over 1000 m, from its lowest point along the western border (380 m) to its highest peak at 1472 m (Figure SM2). The SAC-MN aggregates different habitats and ecosystems, making it an ideal model area for evaluating species richness patterns over time.

As described in Garcia et al. (2024), species occurrences we compiled from several data sources such as online databases, distribution atlases, inventory datasets and field-collected data for a timeframe of two decades (2001–2021). We used the dataset from Garcia et al. (2024), which ensured data quality by addressing common errors identified by Zizka et al. (2019), such as nomenclature mistakes, coordinate absence and/or inaccuracy, and data entry errors. For this study, we selected data with a finer resolution (less than 1 km and aggregated data at 1 km; see Table SM1 and Figure SM3), focusing on four major taxonomic groups with highly threatened species (IUCN 2022a): amphibians (41%), reptiles (21%), birds (13%) and mammals (27%), following the last accepted taxonomy for each species (Bencatel et al. 2019; IUCN 2022b). Aggregate data at a 1 km resolution is commonly found in datasets like distribution atlases, where presence records are typically compiled. Also, we gathered field-collected data to supplement the existing data for groups (i.e., amphibians and reptiles) with limited available information Table SM2). The compilation of occurrence records and the creation of distribution maps were conducted under R software 4.2.2 (Script SM1 and Maps SM1; https://​www.​r-project.​org/​). Given the lack of biodiversity data in the NM region, we specifically focused on selecting biodiversity data within MNP boundaries (consult Figure SM1). We removed record duplicates using QGIS software Version 3.28.1 (https://​www.​qgis.​org/​), and included species with 15 or more presence-only records to ensure the inclusion of ecologically significant species, even though fewer records can influence model outcomes (Sillero et al. 2021). In total, we used 10,190 occurrences records to model 130 species across the groups (Table 1).

We collected environmental predictors (EPs) from the Moderate Resolution Imaging Spectroradiometer (MODIS) datasets, available in Google Earth Engine (GEE; Gorelick et al. 2017; https://​earthengine.​google.​com/​). We selected nineteen MODIS products based on ecosystem functioning and dynamics (Table SM3) (Arenas-Castro & Sillero 2021). We used MODIS monthly burned area product (MCD64A1.061) to generate three fire disturbance-derived variables: Area Annually Burned (AAB), Fire Recurrence (FR), and Time-Since Fire (TSF). Additionally, the study area is highly affected by annual fire disturbances, making these fire-related variables crucial for modelling. We computed the annual means (e.g., ee.Reducer.mean(); https://​developers.​google.​com/​earth-engine/​apidocs/​ee-reducer-mean) of each predictor for the 2001–2021 interval using the original spatial resolutions (e.g., 250 m, 500 m and 1 km) and aggregated to 1 km grid cells. We selected this two-decade period (21 years) because it is the longest continuous period between all variables, except for Leaf Area Index (LAI), and Fraction of Absorbed Photosynthetically Active Radiation (FPAR). To measure the effects of multicollinearity and variance influence, we conducted an analysis involving variance influence factors (VIF) and Pearson and Spearman’s correlation coefficients for all variables. We performed the analyses in R using the packages “raster”, “usdm”, “corrplot” and “GGally” (Hijmans et al. 2022a; Naimi 2017; Schloerke et al. 2021; Wei et al. 2021). The evapotranspiration variable was removed from analyses due to algorithmic product issues (see Mu et al. 2007; Laipelt et al. 2021; Guerschman et al. 2022). Finally, we selected six final predictors (Fig. 2) with VIF < 4 and correlation coefficients (both Pearson and Spearman’s) < 0.75 (Figure SM4). Additionally, we conducted a Principal Component Analysis (PCA) using the “factoextra” R package (Kassambara et al. 2020) to examine the composition of variances and covariances among the primary variables (Figure SM5). The selected EPs (i.e., AAB—Annual area burned; EVI—Enhanced vegetation index; LST-Day—Day land surface temperature; LST-Night—Night land surface temperature; SR-Band1—surface reflectance (620-670 nm); TSF—Time since fire) capture key factors driving habitat suitability for the taxonomic groups. While their impact may vary across taxa, they effectively represent the ecological gradients essential for modelling diversity changes in these groups.

We generated the SRS-ENMs using the Maximum Entropy (MaxEnt) algorithm integrated into GEE (see https://​developers.​google.​com/​earth-engine/​apidocs/​ee-classifier-amnhmaxent) to determine the suitability of habitats of species in SAC-MN (Campos et al. 2023). MaxEnt is a machine-learning approach that uses presence and background data (Sillero & Barbosa 2021) to determine habitat suitability (Phillips et al. 2006, 2017). We wrote a JavaScript script in GEE to ensure a standardised and reproducible modelling process (Script SM2). We randomly selected 70% of the occurrences as training data and 30% as testing data. We defined the background data according to the number of 1 × 1 km grid cells (e.g., 857) within the spatial extent of the MNP. By using background points sampled from across the study area, we aimed to ensure a balanced representation of both occupied and unoccupied pixels, thereby maintaining consistency in model evaluation and providing a robust basis for predicting habitat suitability. We conducted 10 model replications to capture a representative sample of the variability associated with the random selection of training and test data. We executed the models with default settings for all parameters (see https://​developers.​google.​cn/​earth-engine/​apidocs/​ee-classifier-amnhmaxent). We selected the “extrapolate” argument to extend the modelling process to regions with the environmental space beyond training areas (i.e., the NM region). Additionally, each model underwent a maximum of 500 iterations. We employed the GEE “.explain()” command to obtain additional information for each model, such as the percentage contribution of predictors, and evaluation metrics of the Area Under the Curve (AUC). The AUC is a commonly used metric for evaluating ENMs to discriminate presences from absences and ranges from 0.5 (random predictions) to 1 (perfect predictions; Lawson et al. 2014). Additionally, we calculated a set of null models following the methodology by Raes and Ter Steege (2007). To do this, we generated different datasets with the same number of random points as each dataset following a Poisson distribution. We calculated MaxEnt models for each of these random datasets and obtained the AUC values. Then, we compared the training AUC values of the species models with the ones calculated for the null models using the non-parametric Wilcoxon test for repeated measures. We calculated the null models in GEE, in the same way as the empirical models.

To facilitate the export process, we stored the results in new empty images, as we had modelled 130 species with 10 replications over 21 years. These images were used to store the average probability occurrence values from the model replications as distinct bands (per species and year). In the end, we generated a total of 13 images which were then exported from GEE. Additionally, we exported the supplementary model information (e.g., predictors contribution and evaluation metrics) in table format (.csv).

We calculated PSR (from 2001 to 2021) using the R software (see Script SM3). We converted the ENMs into binary presence-absence maps, employing the maximum true skill statistics as species-specific thresholds (maxTSS, Allouche et al. 2006; Biber et al. 2023). Then, we used these categorical maps to compute the PSR across all species, as well as for each taxonomic group separately throughout the 21 years. To analyse the PSR changes over time (Chase et al. 2019), we applied a Mann–Kendall trend test using the R packages “raster”, “terra”, “spatialEco”, and “Kendall” (Evans 2022; Hijmans et al. 2022a, 2022b; McLeod 2022), which detects positive, neutral, or negative trends. Additionally, we performed a Spearman correlation analysis between the trend maps of the different taxonomic groups to detect potential positive or negative correlations among trends.

MaxEnt algorithm was employed for 130 species (conducting 10 model replications per species) over 21 years, resulting in the creation of 27,300 MaxEnt models in GEE.

Contributions of EPs for all models showed diverse patterns across groups (Fig. 3). AAB showed the least contribution to the models (5.77%) compared to the other fire-related variables (e.g., TSF contributed 13.5%). However, surface-proxy (SR-Band1) and day period LST (LST-Day) had significant contributions to the models (17.9 and 14.8%, respectively). However, night-period LST (LST-Night) and vegetation index (EVI) were the variables that most contributed to the models (21.2 and 26.8%, respectively). Examining the EPs contributions for each group in more detail, surface reflectance had a major impact on mammals (Fig. 3). Both day- and night- LSTs revealed a notable impact on each group, with a particular emphasis on LST-day for amphibians and reptiles. Interestingly, the contributions of EVI for amphibians and mammals were found to have lower maximum values compared to birds and reptiles.

Evaluation metrics of the empirical models revealed an overall moderate discrimination of the models (Fig. 4 and Table SM4). The training AUC was higher for mammal species models (0.80), indicating a moderate level of discrimination. The other groups showed lower training AUC values, still suggesting moderate discrimination for amphibians (0.70), birds (0.71), and reptiles (0.71). However, the test AUC was lower for all models, with mammal models showing a test AUC of 0.669, while the rest of the groups had test AUC values below 0.6, specifically amphibians (0.6), birds (0.58), and reptiles (0.57). Nonetheless, empirical models presented higher AUC values than null models, either for training data (Wilcoxon test for paired data: V = 2,739,800,038, p-value < 0.0001; mean AUC empirical: 0.74 ± 0.07; mean AUC null: 0.68 ± 0.06) or test data (Wilcoxon test for paired data: V = 2,633,132,453, p-value < 0.0001; mean AUC empirical: 0.62 ± 0.12; mean AUC null: 0.5 ± 0.11).

After calculating PSR for each year between 2001 and 2021, variations were observed temporally and spatially on multiply occasions (consult Maps SM2). In general, the yearly maps generated indicated that the highest levels of species diversity were predominantly concentrated in the central areas of the SAC-MN, and in regions near the Spanish border. Regarding each group: amphibians demonstrated a pattern indicating high species diversity throughout the southeastern region over time (Maps SM3); birds showed a stable concentration of species diversity in the central SAC-MN, while mammals displayed high species diversity in the northern and northeastern regions, with the highest richness observed in the northern regions adjacent to the Spanish border (Maps SM4 and SM5, respectively); reptiles, on the other hand, displayed the most significant (spatial and temporal) changes, yet distinct areas with high species diversity persisted, particularly in the western and central regions of the NM area (Maps SM6).

Additionally, the Mann–Kendall trend analysis results indicated some significant trends in PSR considering all species (Fig. 5). Notably, the western and eastern regions of SAC-MN showed prominent positive trends, indicating an increasing diversity of species. In contrast, multiple regions within the MNP, central part of SAC-MN, showed clear negative trends. PSR in the NM region remained largely stable over time, with positive trends occurring across much of the area.

Each taxonomic group revealed different trend results (Fig. 6): mammals displayed increasing hotspots, primarily in the MNP, but also revealed a significant area with a negative trend (see Fig. 6); reptiles exhibited increasing hotspots primarily in the western and eastern sides of the SAC-MN; birds had higher (positive and negative) trend values overall despite displaying multiple locations with negative trends; amphibians, on the other hand, demonstrated a unique pattern with not so strong positive trend across the entire SAC-MN, except certain negative trends observed near the centre of the study area. P-values and Z-values of the Mann–Kendall trend test can be consulted in the supplementary material (Figures SM6 and SM7).

The Spearman’s correlation analysis of PSR trends among the different groups revealed a moderate correlation between reptiles and birds (0.38; 0.4), while the correlation between mammals and amphibians was notably low (< 0.1; 0) (Figure SM8). In all other cases, the correlations among the various taxonomic groups were consistently classified as low.

Our methodology involved generating PSR maps for each year and analysing trends over time using SRS-ENMs, which offer a more comprehensive understanding of potential species distributions (Biber et al. 2023). High-spatial resolution SRS-ENMs generated in GEE with moderate but acceptable testing discrimination (0.5 < AUC’s < 0.7) proved to be effective for evaluating species habitat suitability across all groups (Campos et al. 2023). AUC is a threshold-independent metric, assessing a model’s ability to distinguish between presences and random points, not true absences. When applied to presence-only data, AUC values becomes sensitive to species prevalence (Lobo et al. 2008; VanDerWal et al. 2009), challenging the conventional AUC threshold of 0.7 considered indicative of acceptable accuracy in species modelling (Raes & Ter Steege 2007). For instance, specialist species, with narrower distributions, often show higher AUC values, while generalist species, with broader distributions, may have AUC values closer to random due to their widespread presence (Sillero et al. 2021). While the p and z values from the Mann–Kendall test indicate non-significant trends in some areas, the robustness of SRS-ENMs in estimating PSR over time is evident. Plus, the results from the null models confirm that the empirical models perform well (Raes & Ter Steege 2007).

The EPs contributions in our models underscored the substantial importance of integrating SRS variables such as vegetation indexes, surface properties, and fire characteristics. Curiously, SR-Band1 had an extremely high importance for certain mammals species models. This may be due to various factors, such as the availability and precision of field-collected data on mammal distributions as well as the suitability of the environmental predictors or the MaxEnt algorithm for modelling mammal distributions (Raman et al. 2020). The method by Arenas-Castro & Sillero (2021) improves cross-scale habitat suitability monitoring over time using ENMs with SRS data, offering a more dynamic alternative to traditional methods. Incorporating additional SRS products from alternative satellites (e.g., Sentinel or Landsat) in future models could boost overall model performance and augment predictions of species’ habitat suitability (José-Silva et al. 2018; Arenas-Castro & Sillero 2021).

Integrating species-specific thresholds (e.g., maxTSS) into these SRS-ENMs can enhance the accuracy of PSR models, identifying areas with high conservation significance, and forecast the potential impacts of climate change in future scenarios (Biber et al. 2023). However, it is important to acknowledge the ENMs methodology limitations (e.g., overfitting, niche truncation or spatial autocorrelation issues), as it is restricted to SRS-ENMs due to insufficient data on species presence (consult José-Silva et al. 2018; Sillero et al. 2021; Campos et al. 2023), potentially resulting in an inadequate representation of biodiversity. Alternative methods must be explored to evaluate PSR and monitor changes in biodiversity, such as the integration of deep learning (DL) (see Choe et al. 2021). By applying SRS (MODIS) products, this DL framework stacked species distribution models to estimate PSR across the entire Korean Peninsula due to limited survey data availability (Choe et al. 2021). So, integrating remote sensing data/methods into ENMs for detecting high species diversity areas and monitoring changes in PSR over time presents a cost-effective and efficient approach to biodiversity monitoring in challenging environments (e.g., mountain regions) as well as remote and/or inaccessible regions.

Analysing PSR trends is vital for successful conservation planning and management (Condro et al. 2021). Our analysis of PSR over time (Fig. 5 and 6) has identified specific locations within our study area characterized by high or low levels of PSR.

Positive trends were particularly evident in the western and eastern regions of the SAC-MN, likely due to species-specific adaptations to fire-prone environments. For example, some vascular plants in these regions may quickly regenerate after fire events, creating conditions that support their survival and attract other species relying on the new ecological niches. In contrast, the central areas of the SAC-MN, including parts of the MNP, show negative trends, potentially linked to increased human-related influences (e.g., urban expansion, low-intensity rural settlements, agricultural intensification of arable and permanent crops) and habitat fragmentation. The lower levels of PSR in MNP, compared to the NM region, despite the sampling bias towards MNP, align with prior findings (Castro et al. 2010), likely influenced by changes in diverse habitat types (i.e., crop fields). Moreover, taxa-specific responses were evident: mammals showed an increase in richness in the northern areas, whereas reptiles showed significant changes in PSR in regions affected by fire. The stable and high levels of PSR throughout the years in the NM region can be attributed to the remaining areas relatively unaffected by human-related disturbances, which indicates a state of moderate to good ecological preservation. Various direct and indirect factors contribute to changes in biodiversity patterns (Talukder et al. 2022) and influence species’ ecological contributions (i.e., seed dispersal, pollination, soil and water quality). In that sense, anthropogenic disturbances that affect the natural environment and create habitat fragmentation can disturb species richness and habitat quality (Tripp et al. 2019). Hence, the variations in species and taxa responses emphasize the need to tailor conservation strategies to their ecological needs, treating biodiversity as a public good and integrating conservation responsibilities into societal and governmental sectors (Rands et al. 2010).

Having a comprehensive understanding of the spatial and temporal patterns of species richness across biologically different taxonomic groups, including those containing highly threatened species, is crucial to gaining valuable insights into the ecological drivers and/or processes that shape biodiversity in SAC-MN. Given that many species in the region are under constant threats from illegal human activities (e.g., slash-and-burn agriculture), the protection of high and low-species-rich spots is of utmost importance. For instance, forthcoming survey methodologies, strategies or initiatives (i.e. IUCN key species surveys and the monitoring of vital habitats such as the “lameiros”) can be designed using our results, proving crucial for mammal and herpetofauna conservation, especially in areas impacted by water-related agriculture activities and local man-made ponds (Wei et al. 2021). This evaluation of local biodiversity highlights the importance of effective management practices in protected areas for both biodiversity preservation and regions impacted by anthropogenic pressure (Noroozi et al. 2018).

Nonetheless, in the future, raising awareness and promoting sustainable practices among local communities can help to mitigate the effects of human disturbances and ensure the long-term persistence of species and habitat diversity.