Cowans, A. | Search bioRxiv

bioRxiv2024

Improving the integration of AI into existing ecological inference workflows

O_LIArtificial Intelligence (AI) has revolutionised the process of identifying species and individuals in audio recordings and camera trap images. However, despite developments in sensor technology, machine learning, and statistical methods, a general AI-assisted data-to-inference pipeline has yet to emerge. C_LIO_LIWe argue that this is, in part, due to a lack of clarity around several decisions in existing workflows, including: the choice of classifier used (e.g., semi- vs. fully automated); how classifier confidence scores are used and interpreted; and the availability and selection of appropriate statistical methods for drawing ecological inferences. C_LIO_LIHere, we attempt to conceptualise a general workflow associated with automated tools in ecology. We motivate this perspective using our experiences with occupancy modelling using monitoring data collected through passive acoustic monitoring and camera trapping, identifying priority areas for future developments. C_LIO_LIWe offer an accessible guide to support the ecological community in navigating and capitalising on rapid technological and methodological advances. We describe how different error types arise from both sensor-based monitoring and from classifiers themselves; how different error types are handled at each stage of the workflow; and finally, implications and opportunities associated with deciding on methods used at each step of the pipeline. C_LIO_LIWe recommend that "black box" tools like neural network classification algorithms should be embraced in ecology, but widespread uptake requires more formal integration of AI into the existing ecological inference workflows. Like ecological AI more broadly, however, successful development of new data-to-inference pipelines is a multidisciplinary endeavour that requires input from everyone invested in collecting, processing, analysing, and using ecological monitoring data. C_LI

ecology↗

bioRxiv2024

Sample size considerations for species co-occurrence models.

Multispecies occupancy models are widely applied to infer interactions in the occurrence of different species, but convergence and estimation issues under realistic sample sizes are common. We conducted a simulation study to evaluate the ability of a recently developed model to recover co-occurrence estimates under varying sample size and interaction scenarios while increasing model complexity in two dimensions: the number of interacting species and the number of covariates. Using both standard and penalized likelihood, we demonstrate that the ability to quantify interactions in species occupancy using this model is highly sensitive to sample size, detection probability and interaction strength. In the simplest scenario, there is high bias in the interaction parameter (used for co-occurrence inference) with less than 100 sites at high detection, and 400-1000 sites at low detection, depending on interaction strength. Strong co-occurrence is detected consistently above 200 sites with high detection probabilities, but weak co-occurrence is never consistently detected even with 2980 sites. We demonstrate that mean predictive ability of the co-occurrence model is less affected by sample size, with low bias in derived probabilities at 50 sites. Our results highlight that while occupancy patterns are often robust to sample size limitations, reliable inference about co-occurrence demands substantially larger datasets than many studies currently achieve. We caution the interpretation of model output in small datasets or when co-occurrence is expected to be weak, but show methods are suitable to quantify strong co-occurrence in larger datasets and generate predictions of site occupancy states.

ecology↗