MODELING NOCTURNAL CHIROPTERAN ACTIVITY: A REVIEW OF STATISTICAL AND COMPUTATIONAL APPROACHES FOR POINT-PROCESS PREDICTION

Abstract

Passive acoustic monitoring produces time-stamped records of bat detections that underpin ecological surveys and wind-energy impact assessments. This paper is a state-of-the-art review of statistical and machine-learning methods used to describe, explain, and predict bat activity from these data. We organize the literature by data representation, moving from aggregated discrete-time counts to continuous-time event sequences. For count data, we summarize generalized linear models and generalized additive models, emphasizing practical treatments of overdispersion and excess zeros (negative binomial, hurdle, and zero-inflated formulations). We then review hierarchical approaches that separate latent biological activity from imperfect detection, focusing on state-space and latent-state formulations. For temporally dependent detections, we discuss latent-state models alongside continuous-time temporal point process, highlighting self-exciting Hawkes processes for clustered echolocation events. Finally, we cover predictive pipelines based on ensemble learners and deep neural networks, including convolutional neural networks, recurrent architectures, Transformers, and neural point processes. Throughout, we discuss when each family is identifiable and interpretable given typical monitoring designs, and we outline common pitfalls and reporting practices that support robust, mitigation-oriented inference.