Cicada species recognition based on acoustic signals using dynamic time warping-graph based GraphMix, graph convolution network

• Published in: Procedia computer science Vol. 245; pp. 508 - 517
• Main Authors: Yohanes, Gabriel, Prabowo, Abram Setyo, Kurniadi, Felix Indra
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 2024
• Subjects: Acoustic Signal; Audio; GAT; GCN; Graph Representation; GraphMix; Insect Classification; RNN; Speech Recognition; Speech Recognition; Acoustic Signal; GCN; RNN; GraphMix; Insect Classification; GAT; Audio; Graph Representation
• ISSN: 1877-0509; 1877-0509
• DOI: 10.1016/j.procs.2024.10.277

Cicadas, known for their distinctive acoustic signals, have been subjects of classification research for years. Recent researches elaborated the species composition as effect of climate change, further raising the need of effective classification system. Tra- ditional methods rely on manual classification by domain experts, while recent trends favor Artificial Intelligence (AI)-assisted approaches due to their efficiency. However, image-based recognition faces challenges due to cicadas’ varied appearances and environmental factors. Deep learning approaches, particularly utilizing Mel-frequency cepstral coefficients (MFCC) spectrograms, have been effective but are limited by dataset size. Graph Neural Networks (GNN) have surfaced as a promising alternative, lever- aging graph represen- tations to provide additional information like data relationships. In this study, we address the challenge of efficient classification with a small dataset while maximizing feature representation. We explore the effectiveness of MFCC and Chromagram features in a noisy environment, constructing unique graphs for each. Dynamic Time Warping (DTW) is employed to establish connec- tions between nodes. Our experiments on the cicada audio dataset demonstrate the superiority of Chroma- gram over MFCC, with graph-based approaches outperforming graph-less methods such as Recurrent Neural Networks (RNN). Our findings suggest the potential of graph neural networks in audio classification tasks and contribute to advancing the field's methodologies.