Internal clustering validation method for ecosystem health identification using passive acoustic monitoring

Abstract

ABSTRACT : One of the most challenging tasks in unsupervised algorithms is determining the number of clusters, for which Clustering Internal Validity Indices (CIVIs) have been developed. CIVIs are based on metrics such as compactness and separation to evaluate partitions and assist in the quest for the optimal number of clusters. Nevertheless, specialized CIVIs tailored for specific applications have been devised, and there exists no allencompassing CIVI applicable to all scenarios. One contemporary application where such an approach is employed is Passive Acoustic Monitoring (PAM), which employs soundscape data to comprehend community dynamics and complement landscape information. PAM utilizes acoustic variables, including acoustic indices—mathematical functions designed to elucidate various aspects of the complexity within sound recordings. Furthermore, although a relationship between the soundscape and landscape features has been established, there are currently no methodologies that allow for the interpretable integration of acoustic indices into unsupervised algorithms. This gap, in part, arises from the absence of CIVIs based on crisp uncertainty metrics, which is especially critical in decision-making processes like PAM, which often involve ambiguity, non-convex distributions, outliers, and data overlap. This document presents the proposal of a novel CIVI, Uncertainty Frechet (UF), capable of determining the optimal number of clusters for PAM applications. The UF index has also demonstrated proficiency across a multitude of benchmark databases and synthetic datasets. Additionally, the index was employed in two PAM methodologies: the first displayed remarkable performance in identifying ecosystem transformations in an unsupervised manner, tested within a tropical dry forest in Bolivar, Colombia. The second methodology aids in creating acoustic similarity maps, integrating acoustic index information to represent similarities among diferent acoustic patterns across a region. This methodology was tested in an ecosystem with various types of coverage, demonstrating a relationship between i the results and various ecological indicators. The results, both of the UF index and the methodologies, establish the UF index as a valuable tool for researchers and practitioners working for both PAM applications and highly uncertain data applications .