Chipperfield, J. D. | bioRxiv Science

bioRxiv2024

: A tool for modelling ecosystem resilience

AimsA number of modelling frameworks exist to aid in the identification and exploration of stable states and the assessment of resilience from ecological datasets. However, because such models are complex to implement there is a substantial barrier for the application in ecological research. Here we develop a flexible model of ecological resilience based on Bayesian approximation of the "stability landscape". We illustrate its usage on a tropical area where variation in tree cover has been previously interpreted as alternative stable states. MethodsThe stability landscape, from which stable states and resilience parameters are computed, is modelled using a mixture of multiple distributions, each representing a regression between the system state variable and the environmental covariates. Our "mixglm" model allows the mean, precision, and probability parameters of these distributions in the landscape to be dependent on multiple external covariates. "Mixglm" is implemented as a function in R package with the same name, internally using Bayesian inference via NIMBLE. We also conducted a power analysis to provide guidance regarding required sample size. ResultsWe illustrate the use of the "mixglm" on a published case of tree cover in South America which reports a stability landscape with three distinct stable states. Using "mixglm", we were able to replicate the identification of these states. Moreover, we quantified uncertainty of our estimates, and computed resilience of South Americas forests. Conclusions"Mixglm" can be readily used for description of stability landscapes and identification of stable states in most spatial datasets of system state variables, and it is accompanied by tools for calculation of resilience metrics. It can also be further expanded using regression framework to account for more complex data structures such as spatio-temporal data.

ecology↗

bioRxiv2020

A hierarchical point process model for spatial capture-recapture data

O_LISpatial capture-recapture (SCR) is a popular method for estimating the abundance and density of wildlife populations. A standard SCR model consists of two sub-models: one for the activity centers of individuals and the other for the detections of each individual conditional on its activity center. So far, the detection sub-model of most SCR models is designed for sampling situations where fixed trap arrays are used to detect individuals. C_LIO_LINon-invasive genetic sampling (NGS) is widely applied in SCR studies. Using NGS methods, one often searches the study area for potential sources of DNA such as hairs and faeces, and records the locations of these samples. To analyse such data with SCR models, investigators usually impose an artificial detector grid and project detections to the nearest detector. However, there is a trade-off between the computational efficiency (fewer detectors) and the spatial accuracy (more detectors) when using this method. C_LIO_LIHere, we propose a point process model for the detection process of SCR studies using NGS. The model better reflects the spatially continuous detection process and allows all spatial information in the data to be used without approximation error. As in many SCR models, we also use a point process model for the activity centers of individuals. The resulting hierarchical point process model enables estimation of total population size without imputing unobserved individuals via data augmentation, which can be computationally cumbersome. We write custom distributions for those spatial point processes and fit the SCR model in a Bayesian framework using Markov chain Monte Carlo in the R package nimble. C_LIO_LISimulations indicate good performance of the proposed model for parameter estimation. We demonstrate the application of the model in a real-life scenario by fitting it to NGS data of female wolverines (Gulo gulo) collected in three counties of Norway during the winter of 2018/19. Our model estimates that the density of female wolverines is 9.53 (95% CI: 8-11) per 10,000km2 in the study area. C_LI

ecology↗