The purpose of this study was to obtain an estimate of naïve occupancy and abundance while accounting for imperfect detection and without relying on individual identification of the animals. For occupancy, we followed the modelling approach of MacKenzie et al. (2002). It assumes that, if a species is detected at a location, it can be considered present. Yet, non-detection of a species does not equal absence as the species might have gone undetected by the observer. The model, therefore, uses repeat sampling and the obtained detections and non-detections at each site on each survey to estimate the expected occupancy probability (\(\psi\)) and the probability of detecting the species given that it is present (\(p\)). In doing so, it allows the estimation of the probability of site occupancy when the detection probabilities are <1.

We fitted a single-season occupancy model for each year separately using a Maximum Likelihood framework and calculated the Maximum Likelihood for \(\psi\) and \(p\) following MacKenzie et al. (2002) by:

\(L\left(\psi,p\right)=\left[\psi^{n.}\prod_{t=1}^{T}p_t^{n_t}\left(1-p_t\right)^{n-n_t}\right]\ast\left[\psi\prod_{t=1}^{T}\left(1-p_t\right)+\left(1-\psi\right)\right]^{N-n.} \) Eq. 1

with time (\(t\)), the total number of surveyed sites (\(N\)), the number of distinct sampling occasions (\(T\)), the number of sites where the species was detected at time \(t\) (\(n_t\)) and the total number of sites at which the species was detected at least once (\(n.\)). The detection matrices of presence/absence records used for analysis consisted of one record per species, camera location and day. We used the R package unmarked (Fiske and Chandler 2011) to estimate occupancy and detection probabilities, as well as standard errors (SE). R code for the occupancy analysis is available in Suppl. material 6.

For the estimation of population densities, we applied the camera-trap distance sampling (CTDS) methodology developed by Howe et al. (2017) to the data from 2019. CTDS applies the distance sampling framework by Buckland et al. (2004) without the need for individual identification and allows us to account for imperfect detection. Here, a camera trap which is deployed at point \(k\) for a period of time \(T_k\) serves as the observer. Density can be estimated as:

\(\hat D = \frac{{2t\sum\limits_{k = 1}^K {{n_k}} }}{{\theta {w^2}\sum\limits_{k = 1}^K {{T_k}{{\hat P}_k}} }}\) Eq. 2

where ϴ is the horizontal angle of view of the camera model, \(w\) is the truncation distance beyond which observations are discarded, \(n_k\) is the number of animals recorded at point \(k\) and \(\hat{P}_k\) is the probability at each snapshot moment that an animal within the survey area is detected between 0 and \(w\) in front of the camera (Howe et al. 2017).

We measured the distance to the camera of each animal present on a picture by visually comparing them to reference distances recorded during the deployment in that year. A detection function was fitted to the observation distances in the software Distance 7.3 (Thomas et al. 2010) using a conventional distance sampling approach with a point transect model. Two different key functions (Hazard-rate and Half-normal) with up to two cosine adjustment terms were applied and non-parametric bootstrapping with 999 iterations generated by sampling with replacement from the obtained data being used to calculate the coefficient of variation (Buckland et al. 2004). We estimated \(\hat{P}\) as the mean detection probability of all models weighted by their AIC (Akaike Information Criterion; Akaike (1973)). The density estimates were corrected for the limited availability of species (Howe et al. 2017) by estimating an availability factor \(a\) which was derived by calculating the proportion of time animals spend active (activity level), following Rowcliffe et al. (2014).

Additionally, we calculated the mean encounter rate \(\bar{\varepsilon}\) (number of observations per camera) across all cameras and the Coefficient of Variation (CV) of the density estimates following the delta method (Dorfman 1938):

\(CV\left(\hat{D}\right)^2=CV\left(\bar{\varepsilon}\ \right)^2+CV\left(\hat{p}\right)^2+CV\left(a\right)^2\) Eq. 3

The CV of the encounter rate was obtained following the approach of Fewster et al. (2008) for estimating the encounter rate variance in distance sampling. The CV of the detection probability was calculated as the CV of all models weighted by their AIC. The CV of the activity level was calculated using bootstrapping with 999 iterations resampling from the original data (Rowcliffe et al. 2014).

We collected a total of 934,810 pictures on 41,234 trap days with the majority of the pictures being collected in 2015 and 2016. In 2017 and 2018, we obtained considerably less images and observations probably due to the lower sensitivity setting of the camera traps (Table 1). Grouping the pictures into sequences led to a total of 80,191 distinct occurrence events with a mean of 13.2 pictures per event. Out of those, 14,442 contained observations of an animal species. In total, we observed ten wild mammal species and six domestic species (Table 2). The most frequently observed species were domestic cattle (4,572) and domestic horses (4,554), followed by European roe deer (2,548) and wild boar (1,891).

Detection probabilities obtained from occupancy modelling were generally low across species and approached zero for species with fewest observations (Fig. 2). The highest occupancy estimates throughout the study duration were obtained for European roe deer (0.72 - 1) and wild boar (0.48 - 0.92) (Fig. 2), yet their densities were considerably low (Table 3). In contrast to that, domestic cattle and horses showed very high density estimates, while their occupancy was rather low (0.47 - 0.72 and 0.46 - 0.73, respectively). It seems that the domestic species are slightly decreasing in occupancy over the years, while the wild species remain mostly constant. For all other species, we obtained comparatively low occupancy estimates over the years, but for many of them, the number of observations is too low to make any assumptions about temporal trends at this point (Table 2). In 2017, 2018 and 2020, most of the species showed lower occupancy estimates than in other years. A Spearman rank correlation analysis revealed that, for the more common species (e.g. cattle, horses, roe deer and wild boar), detection probabilities were positively correlated with the sensitivity setting of the cameras in that year (R = 0.5, p = 0.002). Yet, we found no significant correlation between occupancy and sensitivity (R = 0.11, p = 0.55). All occupancy and detection probabilities are presented in Suppl. material 1.

Figure 2.  

A) Detection (p) and B) occupancy (Ψ) probabilities for every species per year (see Suppl. material 1 for all values). Transparent bars mark models with too few observations to obtain reasonable results.

When looking at the density analysis, the obtained distances (Suppl. material 2) showed that many cameras only triggered up to 10 metres, even though the motion sensor range was reported to be up to 30 metres (manufacturer's information). For all species, observation distances decreased with increasing distance. Yet, for foxes and wild boar, we obtained very few observations close to the camera, indicating that some animals might have gone undetected at short distances. For both red fox and gray wolf, the density estimates were low with confidence intervals overlapping zero. Therefore, the results should be observed with care. Information and results of each model contributing to the final estimates are presented in Suppl. material 3.

Comparing the occupancy and the density estimates from 2019, we did not observe a strong relationship between them when including all six species in the analysis (R = 0.54, p = 0.30). Yet when only including the wild species, there is a very strong and marginally significant positive correlation between occupancy and density (R = 1, p = 0.08, Fig. 3). With higher densities, the occupancy probability and, therefore, the amount of space used by the population increases.

Figure 3.  

Comparison between occupancy estimates obtained from occupancy modelling (MacKenzie et al. 2002) and density estimates obtained through camera-trap distance sampling (Howe et al. 2017) for the six most common species in the study area in 2019.

We are presenting an extensive dataset obtained from long-term camera trapping that will be maintained and extended in the future. The dataset will offer the opportunity to study local changes in occupancy and density, as well as responses to changes in land-use and vegetation structure or anthropogenic influences on a diverse mammal community. It is currently published in the Darwin Core Archive (DwC-A) format, but GBIF is in the process of implementing a new data model allowing for improved presentation of camera-trap data, as well as the upload of Camtrap DP data in the Integrated Publishing Toolkit (IPT). As the DwC-A standard is not very well suited for camera trap data, for example, by not allowing the hierarchical structure of deployments, media and observations and the manual conversion from Camtrap DP to DwC-A has proven to be challenging (e.g. it was only possible to include images as associatedMedia in the occurrences and not as a distinct source data), we will adopt the new data model once available. Our dataset can then be used for the training for automated computer classification systems, as all images will be made available online.

The very simple index of occupancy presented in this paper already shows that there might be certain trends, such as a decrease in the space used by domestic species and a slightly increasing occupancy of species, such as the red deer or the Iberian ibex that are currently in the process of repopulating the area (Moço et al. 2006; Santos et al. 2011). The data from following years will hopefully render further insights into those processes, especially if also environmental variables or multiple species interactions are included in the analysis. However, we observed an effect of the sensitivity setting of the camera traps on the detection probabilities which needs to be explored in more depth in future analysis by incorporating it directly in modelling. Additionally, it needs to be considered that, because of the design of our camera trapping grid (cameras being only 500 m apart), the assumption of independence amongst sampling locations (MacKenzie et al. 2002) is very likely to be violated, so \(\psi\) should be interpreted with care.

Regarding the density estimates obtained from CTDS, our results were consistent with literature from the same or similar study areas. Unfortunately, for many species, there is no recent literature on population densities. For example, for the gray wolf the last national assessment of populations in Portugal was performed in 2002 - 2003 by Pimenta et al. (2005) using a combination of different indirect sampling methods and resulting in a much lower density estimate of 0.03 – 0.06 individuals/km² (compared to 0.15 individuals/km² in our study) for the study area. Yet, several studies have recorded an increase in wolf numbers and positive population growth rates in Portugal since 2007 (Campos 2018). For roe deer, a relatively recent study by Valente et al. (2014) estimated densities using the distance-sampling framework with data from pellet-group counts in the Montesinho Natural Park (approximately 100 km east of the survey area) and calculated a density of 3.5 (2.26 – 5.45) individuals per km² (5.88 individuals/km² in our study). Many roe deer populations have increased greatly in Portugal within the last 20 years (Torres et al. 2015), so an increase in the Peneda-Gerês National Park due to changes in land-use practices and re-naturalisation of habitats is likely. To date, there are no studies estimating wild boar abundances or densities in northern Portugal. Bosch et al. (2012) and Pittiglio et al. (2018) have conducted surveys using habitat suitability models to predict wild boar densities in northern Portugal. Yet, their estimates were close to zero. This shows that modelled densities from remote sensing data, based on assumptions about habitat suitability, should always be interpreted with care. Data from the National Statistics Institute (Instituto Nacional de Estatística (INE) 2019) reported cattle densities in the study area to be 18.76 individuals/km², ranging from 9 individuals/km² in the Lamas de Mouro and Castro Laboreiro area to 37 individuals/km² in the Gaveira Region (Instituto Nacional de Estatística (INE) 2019). Our estimates (approx. 23 individuals/km²) lie within this interval. For horses, we also obtained data from the INE. Yet, these data seem to be underestimated as it only reports a density of 0.76 individuals/km² and about 115 horses for the region, whereas we obtained densities of almost 11 individuals/km².

For future analysis, we aim at analysing temporal trends, not only in occupancy, but also in abundance and investigate the occupancy-abundance-relationship in more depth. With the data from just one year, it was not possible to observe a relationship between those two metrics. Potential reasons are that we only have one estimate per species for the entire study area and the relationship might depend on species specific variables, such as home range or site specific variables, such as habitat (Efford and Dawson 2012). Another reason is that there can be time lags in the response of occupancy to increases and decreases in abundance that only become apparent from long-term monitoring initiatives (Gaston et al. 2000). Yet, if only considering the wild species, we do observe that higher populations densities could be correlated with higher occupancy estimates. While the domestic species seem to appear in larger groups only in certain parts of the study area, leading to high density estimates with lower occupancy estimates, wild species seem to have higher densities with increasing occupancy. We expect that new technological advancements will soon make the application of CTDS to previously-collected data possible, so that we are planning to also obtain density estimates for the years of 2015 - 2018 and from 2020 onwards. This should hopefully give us further insights into the relationship of occupancy and abundance. Additionally, to including data from several years, we plan to also include environmental variables, as well as potential interactions between wild and domestic species in the analysis, to gain a better understanding of the relationship on a finer spatial scale.