Abstract
There is a growing interest in using passive acoustic monitoring methods to survey biodiversity. Many studies have investigated the efficacy of acoustic indices in monitoring animal communities, particularly bird species richness, with mixed results. It has been suggested that combining multiple acoustic indices could improve accuracy. To accomplish this, researchers have employed machine learning methods, such as the Random Forest Regression, which are considered more robust in this context. However, most machine learning methods have a limitation in that they do not provide well-calibrated uncertainty quantification measures for their predictions. Quantifying uncertainty with the use of appropriate prediction intervals is of paramount importance for making informed management decisions. In this study, we propose addressing this issue using a Machine Learning framework, called Conformal Prediction, which has been developed to provide guaranteed coverage prediction intervals. Specifically, we examine the application of a recently proposed combination of Conformal Prediction with Gaussian Process Regression using data collected through bird and acoustic surveys at biodiverse sites in Cyprus and Australia. Our goal is to demonstrate how the Conformal Prediction framework can be used to assess the models’ prediction accuracy and associated uncertainty when monitoring biodiversity using acoustic indices and machine learning methods. Moreover, we discuss how the framework can be integrated into a wider range of ecological applications to help make more informed conservation management decisions.