In the field of species distribution modeling, there is often a trade-off between interpretability and performance. This partly explains why the study of species distributions is often based on relatively simple models (GLM, MaxEnt, etc.). The advantage of these models is that they capture easily interpretable relationships between species distribution and environmental factors. In this presentation we describe deep learning-based species distribution models (Deep-SDMs) and their main features. Deep-SDMs have emerged recently and often lead to superior performance for predictions. As for JSDM (Joint Species Distribution Models), Deep-SDMs account for species co-occurrences, and have the ability to model and predict species communities. Deep-SDMs can bring several advantages to the field of SDM and macroecology more generally. In particular, their ability to efficiently predict species communities on big datasets by learning a common latent space and the ability to easily integrate complex data (aerial or satellite imagery, environmental neighborhood information). We show for exemple that Deep-SDMs can capture information about the spatial structure of the environment or the landscape to enable better performance than other state-of-the-art models.

Big data approaches, like large-scale camera trapping studies, are becoming more relevant for studying human-wildlife interactions. In addition, open-source object detection models are rapidly improving and have great potential to enhance the image processing of camera trap data. The resulting large-scale and long-term database enables to understand and predict spatiotemporal patterns of human activities in natural areas as well as their interactions with wildlife.

Building on own results, this talk first highlights the performance of open-source object detection model in visitor and wildlife monitoring. Since the accuracy of the detection model is very high, this approach is suitable for biogeographic analysis of spatiotemporal patterns of humans and wildlife. Besides the great acceleration in processing speed, the approach is suitable for long-term monitoring and allows reproducibility in scientific studies while complying with privacy regulations.

The talk will further show how such long-term monitoring can be used to predict visitor flows along a trail network using multiple spatiotemporal predictors. Building on this, the data allow the assessment of human-wildlife interactions and are thus able to enhanced the adaptation of management measures to reduce social and ecological conflicts.

The IUCN Red List of threatened species (RL) is the most authoritative global quantification of extinction risk, and widely used in ecological research and applied conservation. Yet, due to the time-consuming assessment process, the RL is taxonomically and geographically biased, in particular towards the global North and charismatic taxa. One promising approach to speed up RL assessments and overcome these biases is the use of AI to predict extinction risk based on the combination of information from digitized collection specimens and citizen science data with remote sensing information on the environment. Here, I present IUCNN, an approach using deep learning models to predict species RL status from publicly available geographic occurrence records (and other data, such as traits if available).