Automated echolocation classifiers vary in accuracy for northeastern U.S. bat species

Area of Science:

• Ecology
• Bioacoustics
• Conservation Biology

Background:

• Acoustic surveys are vital for monitoring threatened and endangered bat populations.
• Automated bat call classification software (BCID, KPro, SonoBat) is increasingly used in the northeastern United States.
• Independent assessments of these software programs' accuracy are lacking.

Purpose of the Study:

• To independently assess the accuracy of BCID, KPro, and SonoBat in classifying bat echolocation calls to species.
• To evaluate performance using metrics like Positive Predictive Value (PPV), Negative Predictive Value (NPV), Sensitivity (SN), and Specificity (SP).

Main Methods:

• Utilized 1,500 full-spectrum reference calls from nine northeastern United States bat species.
• Tested the accuracy of BCID, KPro, and SonoBat using established statistical measures.
• Compared the performance of the three automated classification programs.

Main Results:

• BCID demonstrated lower accuracy, potentially due to its reliance on zero-crossing data.
• SonoBat and KPro exhibited high NPV and SP, effectively minimizing false positives.
• Both SonoBat and KPro showed lower PPV and SN, indicating a tendency for false negatives, particularly for Myotis species.
• SonoBat and KPro performed better at differentiating Myotis from non-Myotis species.

Conclusions:

• While SonoBat and KPro excel at excluding species, their accuracy in confirming species presence is limited, especially for Myotis bats.
• Automated classification accuracy may be lower under typical field conditions.
• Expert review of automatically classified bat calls is essential for reliable species identification in conservation and regulatory contexts.