Awarded CESU Grant
Matt Koslovsky (PI) received CESU grant “CESU: Advancing Bayesian statistical methods for bat acoustic data” funded by the USGS.
The North American Bat monitoring program collects acoustic recordings to monitor bat populations using echolocation calls. Prior to analysis, recordings are typically processed with automated species identification software (auto-classifier) that assigns a species label to a recording deemed to contain a bat. Regardless of survey effort, these data are subject to imperfect detection as a species may go undetected and a recording may be misclassified as the wrong species by the auto-classifier. One approach for handling potential species misclassification is to collect a calibration data set that relies on a single human observer to view, or vet, a subset of the recordings. When performing validation, the expert may agree with the assigned label, assign a different species label, downgrade to a species couplet, or downgrade to a frequency group. Recordings may also be assigned a species couplet by the auto-classifier if the file contains information that two species have in common or two distinct species were recorded simultaneously. Due to limitations of existing statistical models designed to accommodate imperfect detection, only the recordings that have one species assigned by the auto-classifier and the human expert are included in statistical applications.
The goal of this project is to develop a multispecies count detection model that 1) accommodates auxiliary information regarding the reliability of the auto-classifier using calibration data; 2) informs occupancy and misclassification probabilities using knowledge of species characteristics and behaviors (e.g., species ranges) and other available covariate information (e.g., time of day/weather patterns during collection); and 3) incorporates multiple observer assignments for the species label of a given recording. These developments will allow researchers to fully leverage acoustic bat recording data to provide more accurate estimates of occupancy, misclassification probabilities, and encounter rates. As a result, ecologists will have a better understanding of bat species characteristics, abundances, distributions, and interactions, which can be used to guide future conservation efforts and policy. Further, the proposed methodology has the potential to transform how ecologists perform multispecies occupancy modeling in the presence of imperfect detection. The computational scalability of the proposed approach will enable it to be embedded into integrated population models and will allow for range-wide surveillance efforts.