Classification of Panthera Species Based on Vocalization Using Higuchi Fractal Dimension and K-Nearest Neighbor

Abstract

The genus Panthera, comprising the lion, jaguar, leopard, tiger, and snow leopard, exhibits
complex vocalization characteristics as biological communication mechanisms in its natural
environment. However, manual acoustic identification is limited by environmental variability,
individual sound variations, and recording noise, necessitating an automated approach.
This study applies the Higuchi Fractal Dimension (HFD) for feature extraction and the
K-Nearest Neighbor (KNN) algorithm for classification to recognize vocalizations of five
extant Panthera species. The dataset consists of 100 recordings (20 per species) from the
Tierstimmenarchiv repository, preprocessed by noise reduction, two-second segment selection,
mono conversion, resampling, and amplitude normalization. An empirical fundamental
frequency (F0) verification showed a range of 53 to 367 Hz, confirming the adequacy of the
DWT decomposition depth. The signals were decomposed using DWT (Daubechies 4, 7
levels), and HFD was computed on the resulting coefficients with Kmax of 10, 16, 28, 32, 50,
and 60. Performance was evaluated using repeated nested stratified cross-validation (5-fold
outer loop repeated 10 times; 5-fold inner loop for selecting Kmax and k), together with an
ablation study of DWT approximation (cA) and detail (cD) coefficients and a non-DWT
baseline computed directly from the preprocessed signal. The highest mean accuracy, 70.80%
(SD = 8.77%), was obtained by combining all detail coefficients with the three highest-level
approximation coefficients, and was stable across randomly sampled seeds (SD = 1.06%).
The non-DWT baseline reached only 56.40%. These results indicate that HFD computed
on DWT sub-band provides discriminative features for classifying Panthera vocalizations,
although repeated individuals across folds cannot be ruled out.