# Maximum likelihood estimators are ineffective for acoustic detection of rare bat species

## TL;DR

This study shows that current methods for detecting rare bat species using acoustic monitoring can be unreliable, especially when species are rare or audio files are limited.

## Contribution

The study evaluates the effectiveness of Maximum Likelihood Estimators in acoustic surveys for rare bat species and identifies conditions where they fail.

## Key findings

- Kaleidoscope Pro and SonoBat require different numbers of audio files to detect certain bat species.
- Both programs perform poorly when the species ratio is below 25%.
- The total number of audio files recorded in a night significantly affects detection accuracy for rare species.

## Abstract

Acoustic monitoring is an important tool for determining presence or probable absence of threatened and endangered bats in the United States (US). Federal guidance requires the use of automated identification programs that classify audio files and calculate a Maximum Likelihood Estimator (MLE) for each bat species during each night of a survey. Acoustic presence or absence of species is based on a significant or non-significant MLE, which can have profound regulatory effects, positive or negative. Despite relying on this metric to determine presence of rare species for the past ten years, little is known about the number of files required by available programs to trigger significant MLE or the effect of species ratio on this calculation. We used 1,120 audio files containing echolocation calls from nine northeastern US bat species to simulate survey nights containing variable absolute counts and ratios of species’ audio files. We developed models to estimate the number of audio files that Kaleidoscope Pro (KPro) and SonoBat programs required to establish acoustic presence for each species, and we then applied our best model to a long-term acoustic dataset collected at the Fort Drum Military Installation in New York. Each program required a similar number of files to detect presence for some species, such as Myotis septentrionalis and M. sodalis (8 to 10 files), but differed in file requirements for other species, such as Lasiurus cinereus (KPro =  4; SonoBat =  7) and Perimyotis subflavus (KPro =  10; SonoBat =  6). Both programs performed poorly with determining presence for any species at low species ratio (<25%). Applying our model to the Fort Drum dataset revealed that the total number of audio files recorded within a night had a great effect on whether a rare species was correctly determined to be present. We conclude that MLE should be used with caution during surveys of rare species and could produce misleading results in certain conditions.