Machines as Observers in Ecological Monitoring

Ecological monitoring has greatly benefited from the integration of automated tracking systems and community science data, though combining these sources introduces unique challenges. In this talk, I discuss approaches to integrating semi-structured data from eBird with automated tracking data from the Motus Wildlife Tracking System to study migration patterns and improve ecological inferences. Using hierarchical models, we examine species-specific arrival times along the Atlantic Flyway while accounting for spatiotemporal variations in sampling effort. Drawing on both past research and ongoing work, I explore how these integrated data sources can enhance our understanding of species dynamics, improve abundance estimates, and inform conservation efforts. 

Natural world data collection has increased significantly in recent years, including images, video, and audio captured to understand ecosystems and biodiversity. These data provide a vast and largely uncurated source of scientific information, only a fraction of the data has been analyzed and that analysis is limited, usually focusing on a few target species or categories. However, these sources contain a wealth of "secondary data" including crucial insights into, e.g., interactions, animal social behavior, morphology, habitat, co-occurrence, that is too costly, time-consuming, or expert-dependent to extract at scale. Recent advances in AI methods enable language-based interaction with large-scale databases, potentially enabling the efficient and automated processing techniques needed to unlock the "hidden treasure" in such datasets– being able to directly search large data collections for any scientifically relevant concept would enable richer analyses that span beyond species identification. We propose interactive, open-ended data retrieval via language as a mechanism to support scientific discovery in these collections, and demonstrate a proof-of-concept in partnership with iNaturalist. 

In dynamic riverine systems, such as dammed rivers, habitat conditions can fluctuate rapidly due to environmental and anthropogenic factors. Piping plovers (Charadrius melodus), a threatened bird species, rely on sandbar habitats in these environments for nesting, brood-rearing, and foraging. Effective habitat management requires understanding their habitat usage, with vegetation reduction practices playing a key role in improving reproductive success by creating suitable habitats and reducing predation risks. This study combines Bayesian hierarchical modeling with machine learning to develop a spatial modeling framework that captures plover nesting dynamics along the Missouri River, addressing complex spatial dependencies such as discontinuities and abrupt transitions. Our findings provide valuable insights into plover nesting and habitat usage informing adaptive management strategies to support piping plover conservation efforts in dynamic riverine environments. 

The increasing volumes of species observation data being collected by participatory (or citizen) science projects around the world have great potential for monitoring populations and helping to identify the drivers of population change. However, to realize this potential requires methods that can 1) estimate heterogenous spatial patterns of population change that arise when multiple drivers (e.g. change in land use and climate) affect species populations simultaneously, and 2) control for confounding sources of inter-annual variation common in large-scale crowdsourced spatiotemporal datasets. In this presentation we investigate the use of multi-arm causal machine learning models designed for Conditional Average Treatment Effect (CATE) estimation to address these challenges. Using a simulation study and data from the citizen-science project eBird, we assess performance estimating spatially varying inter-annual trajectories in population size and identifying drivers of population change in the face of real-world confounding. We discuss results showing how this approach can recover heterogenous trends and discuss outstanding challenges.  

Recent advances in tracking technology have made animal trajectory data increasingly common, offering new opportunities to address scientific questions in ecology. Statistical models for such data often leverage remotely sensed information on individual locations over time. Among these, mechanistic models such as the Langevin model are widely used, as they characterize how species use space and respond to environmental conditions while incorporating underlying movement dynamics. Notably, the advection term in the Langevin model is typically defined via spatial derivatives of continuous covariates to ensure differentiability. However, many ecological questions involve discrete covariates with spatial discontinuities. In this talk, we explore the use of neural networks to address the challenges posed by such discontinuities, which are problematic for standard mechanistic models. We apply our method to improve understanding of animal behavior and better capture the effects of discrete covariates on movement of an important invasive species.