There have been numerous studies assessing the effectiveness of acoustic indices in monitoring biodiversity, with mixed results. Most research has focused on the independent use of seven acoustic indices, which have been readily available in the literature. However, there are now over 60 such indices developed. Moreover, to improve the indices’ utility in monitoring animal communities, recent research recommended that the indices be combined using appropriate tools. This is a valid suggestion, considering that each index measures different aspects of the acoustic environment and thus carries complementary information. Building on past research and in an effort to improve the use of acoustic indices as monitoring tools, BIOMON uses machine learning algorithms, mainly the random forests regressor, to identify a combined set of indices that are most useful in measuring bird diversity. To improve the method's generalizability, BIOMON uses acoustic and bird data collected at (a) 30 sites in seven subtropical forest nature reserves in Guangxi, China, (b) 24 sites in the Big Scrub lowland rainforest region in Australia, and (c) 60 sites in agricultural and seminatural areas in Cyprus.
BIOMON’s research activities produced the following key findings. First, it was found that the set of indices with the greatest predictive power for estimating bird species richness in each region varies depending on the area's specific soundscape characteristics. This suggests that no index is universally effective across all areas and that researchers and stakeholders interested in implementing a monitoring protocol based on acoustic indices must first identify the set of indices that are most relevant to their region. Second, it was found that several of the indices commonly used in the literature are less useful than other indices not frequently examined. Therefore, it is recommended that for a more effective monitoring protocol, all available indices be first examined to ensure maximum accuracy when using acoustic indices to monitor biodiversity.
A second key research activity performed during BIOMON was the incorporation of the Conformal Prediction framework into the protocol used to monitor biodiversity using acoustic monitoring methods. Current machine learning methods, such as the aforementioned random forest regressor, have the limitation of not providing a measure of the uncertainty associated with the predictions made. This can be overcome using the conformal prediction framework, which is a novel framework that can provide guaranteed coverage prediction intervals. In layman's terms, this means that when using passive acoustic monitoring methods to survey biodiversity, stakeholders can now obtain both an estimation of the number of species present and the associated range of error based on their desired degree of confidence interval. This information is critical for making informed management decisions.
