Two-stage approach to automatic detection with machine learning for improved surveillance of the invasive Cuban treefrog
The Cuban treefrog ( Osteopilus septentrionalis ), as an invasive species in the southern United States, presents a need for effective surveillance. Automated detection expedites processing of audio data for large-scale surveillance and monitoring programs. However, current available methods commonly used for anuran species have not been sufficient to detect Cuban treefrogs. Here, we present results from a two-stage method for automated detection that employs both cross-correlation template matching and secondary supervised learning classifiers. In the first stage, audio data are screened for initial detections using template matching, in which the detections contain both true and false positives. In the second stage, the false positives are screened out using classifier algorithms. We used this method to process 139,985 audio recordings, consisting of 596,046 total minutes, collected at 13 locations in Louisiana and Florida from 2014 to 2022. From the stage 1 template matching, we detected 83,191 Cuban treefrog signals across recordings. The stage 2 machine learning model was able to identify stage 1 false positive detections with a testing accuracy of 98.46% and a testing false positive rate of 1.116%. After pruning false positive detections, a total of 20,271 individual Cuban treefrog detections remained, distributed mainly across 3 sites in an area with known presence. Locations with presumed absence had an easily verifiable number of false positive detections ( n = 109 across all other sites). The two-stage methodology utilizing both template matching and machine learning algorithms can be integrated into wildlife surveillance or monitoring programs for species with distinctive, conserved calls as an effective way to achieve sensitive species detection with a low incidence of false positives.