Biomonitoring from incomplete surverys

Ecological systems in nature harbor rich biodiversity. However, this inevitably leads to ecologists’ incomplete characterization of entire ecological systems. We may understand the interactions between two or three species in detail, other interactions qualitatively, and yet most are poorly characterized. Conventional methods for inferring the dynamics of diverse systems must then ignore or guess the effects of the unknown interactions. However, the conclusions derived from the characterized part of an ecological system may be completely misleading due to the influence of the uncharacterized remainder. For example, we can imagine a system consisting of three species with rock-paper-scissors competition. If we have only characterized the interaction between two species, we would predict competitive exclusion, yet they actually coexist because of the presence of the third species. Therefore, we need to establish links between the small snapshots of data that ecologists and conservationists feasibly collect, and the big picture conclusions about ecosystem health demanded by policymakers, scientists, and society.

Achieving this goal may seem impossible, and indeed it would be if the aim was to obtain a deterministic link between the dynamics of the whole ecological system and the dynamics of a subsystem. However, if the subsystem contains interactions that are a representative sample of the entire system (i.e. information redundancy), we might be able to infer the dynamics of the system as a whole with incomplete information. By exploiting this information redundancy, my research centers on establishing generic probabilistic links to scale up information about small amounts of an incompletely characterized system to rapidly assess the persistence of entire ecological systems. For example, I found that the persistence of small interaction subnetworks (i.e. the coexistence of all species in the subnetwork) isolated from their larger network is a reliable probabilistic indicator of the persistence of the whole ecological network (i.e. the coexistence of all species in the full network) (BiorXiv, 2022). Specifically, when the whole ecological network is persistent, more than half of all the subnetworks persist in isolation. In contrast, when the whole ecological network is not persistent, almost all the subnetworks fail to persist in isolation. By leveraging the large differences in coexistence in these two cases, I developed a Bayesian statistical framework to effectively and rigorously update our beliefs about the persistence of the whole network from whether subnetworks are persistent in isolation.

Beyond incomplete characterization problems in local ecological systems, I have extended these approaches to spatial metapopulations. A metapopulation consists of spatially separated local populations coupled to one another via dispersal. The metapopulation concept is central to conservation science, where it is used to assess and predict the consequences of habitat loss for species persistence in highly fragmented landscapes. However, as all patches in the metapopulation have some non-zero influence on metapopulation persistence, quantifying metapopulation persistence would seem to require characterizing all the habitat patches in a landscape. However, a complete survey is logistically challenging to conduct and repeat. To address whether we can learn enough about metapopulation persistence from an incomplete survey, I have built a robust statistical framework that can infer metapopulation persistence with just few samples of patches (PRSB, 2022). What makes the inference possible is that metapopulation dynamics pose strong ecological constraints on the effect of dispersal among habitat patches.