Two-stage models improve machine learning classifiers in wildlife research: A case study in identifying false positive detections of Ruffed Grouse

Autonomous recording units are increasingly being used to monitor wildlife on large geographic and temporal scales, paired with machine learning (ML) to automate detection of wildlife. However, false positive detections from ML classifiers can result in erroneous ecological models that can lead to misguided management and conservation actions. We used a two-stage general approach to understand and reduce false positive detections, a technique in which outputs of the primary classification model are passed to a secondary classification model to yield the probability that a detection from the primary model is a true positive detection. This approach is demonstrated on two open-source models that detect Ruffed Grouse ( Bonasa umbellus ). We analyzed over 9500 h of acoustic data collected in 2022–2023 from the Green Mountain National Forest in Vermont, USA, and found the two models detected different types of acoustic signals associated with differing life history traits. The first model yielded 4106 detections (71.5 % true positives) while the second model yielded 524 detections (17.0 % true positives). Secondary logistic regression models separated true positives and false positives with high accuracy (84.5 % and 89.8 % respectively). Our findings go beyond improving Ruffed Grouse monitoring and conservation efforts to, more broadly, illustrate how two-stage ML approaches can improve the use of model-derived detections in wildlife research. • A plethora of machine learning models are being used to automate detection of wildlife. • False positive wildlife detections can impact downstream ecological modeling. • Two-stage models can help weed out false positive detections, improving performance. • The method is demonstrated for detection of Ruffed Grouse in field recordings. • Using established and novel examples, we show significant model improvements. • Two-stage models separated true and false positives with high accuracy (>80 %).