• individual acoustic discrimination
• Strix aluco
• survey
• Tawny owl

An acoustical approach has proven to be the most cost-effective method to survey populations of an elusive but vociferous bird of prey like the Tawny owl (Strix aluco). Indeed, recognizable individual variation of male territorial calls has been widely documented in literature.
However, the most widespread methods used for individual acoustic discrimination in this species still suffer some operator subjectivity: many of them rely on measurements directly defined by the exact selection bounds the operator draws on the spectrogram; others simply conduct a visual comparison between hootings. To some extent, this may produce a scarce repeatability between operators and low reproducibility of results. In addition, individual acoustic discrimination in Tawny owls has been conducted through the use of supervised statistical techniques on groups of known individuals; to date, none of these statistical tools could be useful in discriminating calls when number and identities of individuals are not known.
This work had the first goal of assessing the ability of four unsupervised clustering algorithms (K-means, K-medoids, Affinity propagation and Hierarchical clustering) in discriminating between hootings of known male Tawny owls, by the means of their sole acoustic features, accounting for no information about individual number or identities. We accounted for both traditional “selection bounds-based” measurements, broadly employed in literature, and introduce alternative “robust-signal” measurements, which values are way more stable towards variation on spectrogram quality and human subjectivity. Indeed, the second aim of this thesis was to compare the effectiveness of these different sets of measurements throughout the different clustering algorithms, testing whether robust could be as efficient as or even more efficient than traditional ones during the discrimination process. Finally, we aimed at evaluating whether the best-fitting model, resulting by the more suitable combination of clustering algorithm and set of measurements on known individuals, would have been useful to census an entire population of unknown individuals, by predicting number and distribution of males throughout our study area.
A total of 412 good quality hootings, relative to 47 owl localizations, were collected in a period going from November 2019 to July 2021. Repeated visits (2-3 times a week) along defined transects were scheduled, listening for both spontaneous or elicited territorial calls, recording them, and noting each singing owl location. Spectrograms of recorded audio have been analysed on “Raven 1.6 Pro”, measuring both temporal and frequency variables of male songs. Both traditional and robust measurements were extracted by performing manual selections around each note.
For a first validation of our methods, analysis have been conducted on a subset of 80 hootings belonging to a test group of 8 assumed known individuals. Accounting for this subset of data, coefficient of variation ratios have been estimated for every considered acoustic variable, as a measure of inter-individual variability. Unsupervised clustering algorithms have been computed accounting separately for traditional and robust measurements, to compare the respective ability on grouping calls belonging to each individual. To verify the reliability of each cluster solution, optimal number of clusters, internal and external validation indices have been estimated.
We found that three out of the four tested algorithms (K-medoids, Affinity propagation and Hierarchical clustering) returned successful clustering solutions according to all the adopted validation indices. In particular, best results were obtained when robust measurements were used. Hierarchical clustering computed on robust measurements provided the best solution, with 87.5% of correctly classified hootings.
However, when this best-fitting model were extended to the entire dataset of collected vocalizations, several issues emerged. Our findings confirm that, while representing a powerful instrument and an excellent starting point for acoustic surveys on this species, these statistical tools alone are still not enough to achieve an exhaustive census of an entirely unknown population of Tawny owls. Thus, the integration of additional information from alternative monitoring methods are necessary.
This thesis could provide relevant innovation to bioacoustical studies concerning Tawny owls and other nocturnal or elusive species in general. Indeed, robust-signal measurements turned out to be a solid alternative for individual discrimination in this species, improving both discriminative efficiency in addition to comparability and repeatability of results. Moreover, this is one of the few existing studies on which unsupervised classification techniques have been successfully applied for individual acoustic discrimination.