Open Access

Gabriella Krivek

• Amid the current global biodiversity crisis, being able to accurately monitor the changing state of biodiversity is essential for successful conservation actions and policy. Despite the pressing need for reliable and cost-effective monitoring methods, collecting such data remains extremely difficult for elusive species, such as temperate zone bats. Although bats are important indicators of environmental changes, monitoring bat populations is challenging because they are nocturnal, volant, small, and highly sensitive to human activities and disturbance. Thus far, population trends of temperate zone bats have been mainly based on visual surveys, including winter hibernation counts at underground sites. However, as bats may not always be roosting in visible locations within the hibernacula, it is currently unknown how these estimates relate to actual population sizes. Infrared light barriers combined with camera traps are a novel method to monitor bats at underground sites. When installed at the entrance of hibernacula, infrared light barriers have the potential to estimate site-level population sizes more accurately than visual surveys, by counting all bats flying in and out of the site. Moreover, camera traps, consisting of a digital camera and white flash, can be used for species-level identification. However, for this new method to be applicable as a large-scale bat monitoring technique, it is important to characterize it with regard to three main criteria: is the method minimally invasive, is it accurate, and is it scalable in terms of spatial and temporal resolution? Therefore, the purpose of this thesis was to investigate the invasiveness and accuracy of this novel bat monitoring method, and to develop standardized and automated data analysis pipelines, both for the light barrier and camera trap data, to support the deployment of this method at scale. In Publication I, we used light barrier data, infrared video recordings and acoustic data from an experimental field study to investigate whether the white flash of the camera trap has any measurable short- or long-term effect on bat activity and behavior. The flash of the camera trap was turned on and off every week at each site, which allowed us to compare the activity and behavior of bats between flash-on and flash-off nights. We found that despite the high sensitivity of bats to disturbance, they did not change their nightly activity patterns, flight direction, echolocation behavior, or long-term site use in response to the white flash of the camera trap. Based on these results, we concluded that camera traps using a white flash are a minimally invasive method for monitoring bat populations at hibernacula, providing high quality images that allows species-level identification. In Publication II, we used infrared video surveillance to quantify the accuracy of infrared light barriers, and we described a standardized methodology to estimate population sizes and trends of hibernating bat assemblages using light barrier data. We showed that light barrier accuracy varies based on the model and location of the installation relative to the entrance, with the best combination achieving nearly perfect accuracy over the spring emergence phase. When compared to light barrier-based estimates, we found that visual counts markedly underestimated population sizes, recovering less than 10% of the bats at the most complex hibernacula. Moreover, light barrier-based population trends showed regional patterns of growth and decline that were not detectable using the visual count data. Overall, we established that the light barrier data can be used to estimate the population size and trends of hibernating bat assemblages with unprecedented accuracy and in a standardized way. In Publication III, we described a deep learning-based tool, BatNet, that can accurately and efficiently identify bat species from camera trap images. The baseline model was trained to identify 13 European bat species or species complexes using camera trap images collected at 32 hibernation sites (i.e., trained sites). We showed that the baseline model performance was very high across all 13 bat species on trained sites, as well as on untrained sites when the camera angle and distance from the entrance were comparable to the training images. At untrained sites with more atypical camera placements, we demonstrated the ability to retrain the baseline model and achieve an accuracy comparable to the trained sites. Additionally, we showed that the model can learn to identify a new species, while maintaining high classification accuracy for all original species. Finally, we established that BatNet can be used to accurately describe ecological metrics from camera trap images (i.e., species diversity, relative abundance, and species-specific phenology) that are relevant for bat conservation. We conclude that infrared light barriers and camera traps offer a minimally invasive and accurate method to monitor site-level bat population trends and species-specific phenological estimates at underground sites. Such remote data collection approaches are particularly relevant for monitoring large, complex hibernation sites, where traditional visual surveys are not feasible or account only for a small fraction of the actual population. Combining this automated monitoring method with a deep learning-based species identification tool, BatNet, allows us quickly and accurately analyze millions of camera trap images resulting from large-scale, long-term camera trap studies. As a result, we can gain unprecedented insights into the behavior and population dynamics of these enigmatic species, drastically improving our ability to support data-driven bat conservation.